Doven
commited on
Commit
·
f7009b3
1
Parent(s):
6abdfdc
update code.
Browse filesThis view is limited to 50 files because it contains too many changes.
See raw diff
- .gitignore +10 -0
- README.md +79 -3
- checkpoint/generalization.pth +0 -0
- dataset/__init__.py +1 -0
- dataset/cifar100_resnet18bn/model.py +27 -0
- dataset/cifar100_resnet18bn/prepare.py +192 -0
- dataset/cifar100_resnet18bn/test.py +28 -0
- dataset/cifar100_resnet18bn/train.py +195 -0
- dataset/cifar10_cnnmedium/model.py +48 -0
- dataset/cifar10_cnnmedium/test.py +28 -0
- dataset/cifar10_cnnmedium/train.py +192 -0
- dataset/cifar10_cnnsmall/model.py +48 -0
- dataset/cifar10_cnnsmall/test.py +28 -0
- dataset/cifar10_cnnsmall/train.py +192 -0
- dataset/cifar10_mobilenetv3/model.py +21 -0
- dataset/cifar10_mobilenetv3/test.py +28 -0
- dataset/cifar10_mobilenetv3/train.py +199 -0
- dataset/cifar10_resnet18/model.py +17 -0
- dataset/cifar10_resnet18/test.py +28 -0
- dataset/cifar10_resnet18/train.py +191 -0
- dataset/cifar10_vitbase/model.py +17 -0
- dataset/cifar10_vitbase/test.py +28 -0
- dataset/cifar10_vitbase/train.py +199 -0
- dataset/condition_classinput_inference/dataset.py +41 -0
- dataset/condition_classinput_inference/model.py +25 -0
- dataset/condition_classinput_inference/test.py +30 -0
- dataset/condition_classinput_inference/train.py +209 -0
- dataset/condition_classinput_vittiny/dataset.py +41 -0
- dataset/condition_classinput_vittiny/detail.py +58 -0
- dataset/condition_classinput_vittiny/finetune.py +215 -0
- dataset/condition_classinput_vittiny/model.py +25 -0
- dataset/condition_classinput_vittiny/split.sh +28 -0
- dataset/condition_classinput_vittiny/test.py +30 -0
- dataset/condition_classinput_vittiny/train.py +212 -0
- dataset/condition_classinput_vittiny/train.sh +10 -0
- dataset/condition_imageinput_vittiny/README.md +1 -0
- dataset/condition_imageinput_vittiny/dataset.py +46 -0
- dataset/condition_imageinput_vittiny/model.py +18 -0
- dataset/condition_imageinput_vittiny/test.py +30 -0
- dataset/condition_imageinput_vittiny/train.py +208 -0
- dataset/condition_imageinput_vittiny/train.sh +11 -0
- dataset/condition_permutation_vittiny/model.py +18 -0
- dataset/condition_permutation_vittiny/test.py +31 -0
- dataset/condition_permutation_vittiny/train.py +210 -0
- dataset/condition_permutation_vittiny/train.sh +10 -0
- dataset/config.json +1 -0
- dataset/dataset.py +327 -0
- dataset/downtask_detection/README.md +1 -0
- dataset/downtask_detection/test.sh +11 -0
- dataset/downtask_dora_r16/adapter_config.json +23 -0
.gitignore
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
/.idea
|
2 |
+
/.vscode
|
3 |
+
**/checkpoint*/
|
4 |
+
**/__pycache__/
|
5 |
+
**/generated*/
|
6 |
+
**/wandb/
|
7 |
+
**/full_model.pth
|
8 |
+
/rubbish
|
9 |
+
**/*cache*
|
10 |
+
/workspace/classinput/Qwen25llm/
|
README.md
CHANGED
@@ -1,3 +1,79 @@
|
|
1 |
-
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Recurrent Parameter Generation
|
2 |
+
The official repository of paper [Recurrent Diffusion for Large-Scale Parameter Generation]().
|
3 |
+
|
4 |
+
|
5 |
+
## Introduction
|
6 |
+
Parameter generation has long struggled to scale, significantly limiting its applications.
|
7 |
+
In this study, we introduce Recurrent diffusion for large-scale Parameter Generation, or RPG,
|
8 |
+
which models large-scale parameter generation through a recurrent diffusion process.
|
9 |
+
We divide the trained parameters into non-overlapping parts and propose a recurrent model to learn their relationships.
|
10 |
+
The outputs of this recurrent model, serving as conditions, are then input into a diffusion model to generate neural network parameters.
|
11 |
+
Utilizing only a single GPU, our method can generate parameters for popular vision and language models, such as ConvNeXt-L and LoRA parameters for LLaMA-7B.
|
12 |
+
Across various architectures and tasks, the generated parameters consistently achieve comparable performance to those of trained networks.
|
13 |
+
Additionally, our approach demonstrates potential in generating models capable of handling unseen tasks,
|
14 |
+
indicating that recurrent diffusion greatly enhances the practicality of parameter generation.
|
15 |
+
|
16 |
+
|
17 |
+
|
18 |
+
|
19 |
+
|
20 |
+
|
21 |
+
|
22 |
+
|
23 |
+
## Environment
|
24 |
+
Before you get started, you need to set up a conda environment first.
|
25 |
+
1. Create your conda environment.
|
26 |
+
```shell
|
27 |
+
conda create -n rpg python=3.11
|
28 |
+
conda activate rpg
|
29 |
+
conda install pytorch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 pytorch-cuda=12.1 -c pytorch -c nvidia
|
30 |
+
```
|
31 |
+
2. Install mamba-ssm. (You may run into compilation issues, refer to the [official mamba-ssm repository](https://github.com/state-spaces/mamba) for details.)
|
32 |
+
```shell
|
33 |
+
pip install mamba-ssm[causal-conv1d]
|
34 |
+
pip install causal-conv1d
|
35 |
+
```
|
36 |
+
3. Install other dependencies for this repository.
|
37 |
+
```shell
|
38 |
+
git lfs install
|
39 |
+
git clone https://huggingface.co/MTDoven/Recurrent-Parameter-Generation
|
40 |
+
cd Recurrent-Parameter-Generation
|
41 |
+
pip install -r requirements.txt
|
42 |
+
```
|
43 |
+
|
44 |
+
|
45 |
+
|
46 |
+
|
47 |
+
## Quick Start
|
48 |
+
1. Modify your config file.
|
49 |
+
```shell
|
50 |
+
# Set up your configs interactively.
|
51 |
+
python ./workspace/set_configs.py
|
52 |
+
```
|
53 |
+
|
54 |
+
2. Download checkpoint datasets.
|
55 |
+
```shell
|
56 |
+
|
57 |
+
# Download the ViTTiny1022 dataset to /path/to/your/download/ViTTiny1022
|
58 |
+
mv /path/to/your/download/ViTTiny1022/* ./dataset/condition_classinput_vittiny/
|
59 |
+
```
|
60 |
+
|
61 |
+
3. Try to generate with RPG model.
|
62 |
+
```shell
|
63 |
+
cd ./workspace
|
64 |
+
CUDA_VISIBLE_DEVICES=0 python ./classinput/launch.py
|
65 |
+
# CUDA_VISIBLE_DEVICES=<GPU_index> python ./classinput/launch.py
|
66 |
+
```
|
67 |
+
|
68 |
+
You can get more information from [Github](https://github.com/NUS-HPC-AI-Lab/Recurrent-Parameter-Generation).
|
69 |
+
|
70 |
+
|
71 |
+
|
72 |
+
|
73 |
+
## Acknowledgment
|
74 |
+
coming soon...
|
75 |
+
|
76 |
+
|
77 |
+
## Citation
|
78 |
+
coming soon...
|
79 |
+
|
checkpoint/generalization.pth
ADDED
File without changes
|
dataset/__init__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
from .register import *
|
dataset/cifar100_resnet18bn/model.py
ADDED
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import torch
|
3 |
+
import timm
|
4 |
+
import os
|
5 |
+
|
6 |
+
|
7 |
+
def Model():
|
8 |
+
model = timm.create_model("resnet18", pretrained=True)
|
9 |
+
model.fc = nn.Linear(512, 100)
|
10 |
+
if os.path.exists(os.path.join(os.path.dirname(__file__), "full_model.pth")):
|
11 |
+
model.load_state_dict(torch.load(os.path.join(os.path.dirname(__file__), "full_model.pth"), map_location="cpu"))
|
12 |
+
for k, v in model.named_parameters():
|
13 |
+
if k in ["layer4.1.bn1.weight", "layer4.1.bn1.bias", "layer4.1.bn2.weight", "layer4.1.bn2.bias"]:
|
14 |
+
v.requires_grad = True
|
15 |
+
else: # requires_grad = False
|
16 |
+
v.requires_grad = False
|
17 |
+
return model, model.fc
|
18 |
+
|
19 |
+
|
20 |
+
if __name__ == "__main__":
|
21 |
+
model, _ = Model()
|
22 |
+
print(model)
|
23 |
+
num_param = 0
|
24 |
+
for k, v in model.named_parameters():
|
25 |
+
num_param += v.numel()
|
26 |
+
print(k)
|
27 |
+
print("num_param:", num_param)
|
dataset/cifar100_resnet18bn/prepare.py
ADDED
@@ -0,0 +1,192 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
seed = SEED = 20
|
6 |
+
torch.manual_seed(seed)
|
7 |
+
torch.cuda.manual_seed(seed)
|
8 |
+
torch.cuda.manual_seed_all(seed)
|
9 |
+
torch.backends.cudnn.deterministic = True
|
10 |
+
torch.backends.cudnn.benchmark = True
|
11 |
+
np.random.seed(seed)
|
12 |
+
random.seed(seed)
|
13 |
+
|
14 |
+
|
15 |
+
try: # relative import
|
16 |
+
from model import Model
|
17 |
+
except ImportError:
|
18 |
+
from .model import Model
|
19 |
+
|
20 |
+
# import
|
21 |
+
import torch.nn as nn
|
22 |
+
from torch import optim
|
23 |
+
from torch.optim import lr_scheduler
|
24 |
+
from torch.utils.data import DataLoader
|
25 |
+
import torchvision.transforms as transforms
|
26 |
+
from torchvision.datasets import CIFAR100 as Dataset
|
27 |
+
from tqdm.auto import tqdm
|
28 |
+
import os
|
29 |
+
import warnings
|
30 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
31 |
+
|
32 |
+
# load additional config
|
33 |
+
import json
|
34 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
35 |
+
with open(config_file, "r") as f:
|
36 |
+
additional_config = json.load(f)
|
37 |
+
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
# config
|
42 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
43 |
+
config = {
|
44 |
+
"dataset_root": "from_additional_config",
|
45 |
+
"batch_size": 500 if __name__ == "__main__" else 200,
|
46 |
+
"num_workers": 32,
|
47 |
+
"learning_rate": 0.0005,
|
48 |
+
"weight_decay": 0.000005,
|
49 |
+
"epochs": 200,
|
50 |
+
"save_learning_rate": 0.0,
|
51 |
+
"total_save_number": 1,
|
52 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
53 |
+
}
|
54 |
+
config.update(additional_config)
|
55 |
+
|
56 |
+
|
57 |
+
|
58 |
+
|
59 |
+
# Data
|
60 |
+
dataset = Dataset(
|
61 |
+
root=config["dataset_root"],
|
62 |
+
download=True,
|
63 |
+
train=True,
|
64 |
+
transform=transforms.Compose([
|
65 |
+
transforms.Resize(80),
|
66 |
+
transforms.RandomHorizontalFlip(),
|
67 |
+
transforms.AutoAugment(policy=transforms.AutoAugmentPolicy("cifar10")),
|
68 |
+
transforms.ToTensor(),
|
69 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
70 |
+
])
|
71 |
+
)
|
72 |
+
train_loader = DataLoader(
|
73 |
+
dataset=dataset,
|
74 |
+
batch_size=config["batch_size"],
|
75 |
+
num_workers=config["num_workers"],
|
76 |
+
shuffle=True,
|
77 |
+
drop_last=True,
|
78 |
+
pin_memory=True,
|
79 |
+
)
|
80 |
+
test_loader = DataLoader(
|
81 |
+
dataset=Dataset(
|
82 |
+
root=config["dataset_root"],
|
83 |
+
download=True,
|
84 |
+
train=False,
|
85 |
+
transform=transforms.Compose([
|
86 |
+
transforms.Resize(80),
|
87 |
+
transforms.ToTensor(),
|
88 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
89 |
+
])),
|
90 |
+
batch_size=config["batch_size"],
|
91 |
+
num_workers=config["num_workers"],
|
92 |
+
shuffle=False,
|
93 |
+
pin_memory=True,
|
94 |
+
)
|
95 |
+
|
96 |
+
# Model
|
97 |
+
model, head = Model()
|
98 |
+
model = model.to(device)
|
99 |
+
criterion = nn.CrossEntropyLoss()
|
100 |
+
pre_optimizer = optim.AdamW(
|
101 |
+
head.parameters(),
|
102 |
+
lr=0.001,
|
103 |
+
weight_decay=config["weight_decay"],
|
104 |
+
)
|
105 |
+
optimizer = optim.AdamW(
|
106 |
+
model.parameters(),
|
107 |
+
lr=config["learning_rate"],
|
108 |
+
weight_decay=config["weight_decay"],
|
109 |
+
)
|
110 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
111 |
+
optimizer,
|
112 |
+
T_max=config["epochs"],
|
113 |
+
eta_min=config["save_learning_rate"],
|
114 |
+
)
|
115 |
+
|
116 |
+
|
117 |
+
|
118 |
+
|
119 |
+
# Training
|
120 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
121 |
+
model.train()
|
122 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(train_loader),
|
123 |
+
total=len(dataset) // config["batch_size"]):
|
124 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
125 |
+
optimizer.zero_grad()
|
126 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
127 |
+
outputs = model(inputs)
|
128 |
+
loss = criterion(outputs, targets)
|
129 |
+
loss.backward()
|
130 |
+
optimizer.step()
|
131 |
+
if scheduler is not None:
|
132 |
+
scheduler.step()
|
133 |
+
|
134 |
+
# test
|
135 |
+
@torch.no_grad()
|
136 |
+
def test(model=model):
|
137 |
+
model.eval()
|
138 |
+
all_targets = []
|
139 |
+
all_predicts = []
|
140 |
+
test_loss = 0
|
141 |
+
correct = 0
|
142 |
+
total = 0
|
143 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(test_loader),
|
144 |
+
total=len(test_loader.dataset) // config["batch_size"]):
|
145 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
146 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
147 |
+
outputs = model(inputs)
|
148 |
+
loss = criterion(outputs, targets)
|
149 |
+
# to logging losses
|
150 |
+
all_targets.extend(targets.flatten().tolist())
|
151 |
+
test_loss += loss.item()
|
152 |
+
_, predicts = outputs.max(1)
|
153 |
+
all_predicts.extend(predicts.flatten().tolist())
|
154 |
+
total += targets.size(0)
|
155 |
+
correct += predicts.eq(targets).sum().item()
|
156 |
+
loss = test_loss / (batch_idx + 1)
|
157 |
+
acc = correct / total
|
158 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}")
|
159 |
+
model.train()
|
160 |
+
return loss, acc, all_targets, all_predicts
|
161 |
+
|
162 |
+
# save train
|
163 |
+
def save_train(model=model, optimizer=optimizer):
|
164 |
+
model.train()
|
165 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
166 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
167 |
+
optimizer.zero_grad()
|
168 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
169 |
+
outputs = model(inputs)
|
170 |
+
loss = criterion(outputs, targets)
|
171 |
+
loss.backward()
|
172 |
+
optimizer.step()
|
173 |
+
# Save checkpoint
|
174 |
+
_, acc, _, _ = test(model=model)
|
175 |
+
if not os.path.isdir('checkpoint'):
|
176 |
+
os.mkdir('checkpoint')
|
177 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items()}
|
178 |
+
torch.save(save_state, f"full_model.pth")
|
179 |
+
print("save:", f"full_model.pth")
|
180 |
+
|
181 |
+
|
182 |
+
|
183 |
+
|
184 |
+
# main
|
185 |
+
if __name__ == '__main__':
|
186 |
+
test(model=model)
|
187 |
+
train(model=model, optimizer=pre_optimizer, scheduler=scheduler)
|
188 |
+
train(model=model, optimizer=pre_optimizer, scheduler=scheduler)
|
189 |
+
for epoch in range(config["epochs"]):
|
190 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
191 |
+
test(model=model)
|
192 |
+
save_train(model=model, optimizer=optimizer)
|
dataset/cifar100_resnet18bn/test.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
if __name__ == "__main__":
|
4 |
+
from train import *
|
5 |
+
else: # relative import
|
6 |
+
from .train import *
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
|
11 |
+
try:
|
12 |
+
test_item = sys.argv[1]
|
13 |
+
except IndexError:
|
14 |
+
assert __name__ == "__main__"
|
15 |
+
test_item = "./checkpoint"
|
16 |
+
test_items = []
|
17 |
+
if os.path.isdir(test_item):
|
18 |
+
for item in os.listdir(test_item):
|
19 |
+
item = os.path.join(test_item, item)
|
20 |
+
test_items.append(item)
|
21 |
+
elif os.path.isfile(test_item):
|
22 |
+
test_items.append(test_item)
|
23 |
+
|
24 |
+
|
25 |
+
for item in test_items:
|
26 |
+
state = torch.load(item, map_location="cpu")
|
27 |
+
model.load_state_dict({key: value.to(torch.float32).to(device) for key, value in state.items()}, strict=False)
|
28 |
+
loss, acc, all_targets, all_predicts = test(model=model)
|
dataset/cifar100_resnet18bn/train.py
ADDED
@@ -0,0 +1,195 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
seed = SEED = 20
|
6 |
+
torch.manual_seed(seed)
|
7 |
+
torch.cuda.manual_seed(seed)
|
8 |
+
torch.cuda.manual_seed_all(seed)
|
9 |
+
torch.backends.cudnn.deterministic = True
|
10 |
+
torch.backends.cudnn.benchmark = True
|
11 |
+
np.random.seed(seed)
|
12 |
+
random.seed(seed)
|
13 |
+
|
14 |
+
|
15 |
+
try: # relative import
|
16 |
+
from model import Model
|
17 |
+
except ImportError:
|
18 |
+
from .model import Model
|
19 |
+
|
20 |
+
# import
|
21 |
+
import torch.nn as nn
|
22 |
+
from torch import optim
|
23 |
+
from torch.optim import lr_scheduler
|
24 |
+
from torch.utils.data import DataLoader
|
25 |
+
import torchvision.transforms as transforms
|
26 |
+
from torchvision.datasets import CIFAR100 as Dataset
|
27 |
+
from tqdm.auto import tqdm
|
28 |
+
import os
|
29 |
+
import warnings
|
30 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
31 |
+
|
32 |
+
# load additional config
|
33 |
+
import json
|
34 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
35 |
+
with open(config_file, "r") as f:
|
36 |
+
additional_config = json.load(f)
|
37 |
+
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
# config
|
42 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
43 |
+
config = {
|
44 |
+
"dataset_root": "from_additional_config",
|
45 |
+
"batch_size": 100 if __name__ == "__main__" else 200,
|
46 |
+
"num_workers": 4,
|
47 |
+
"learning_rate": 0.01,
|
48 |
+
"weight_decay": 5e-6,
|
49 |
+
"epochs": 1,
|
50 |
+
"save_learning_rate": 0.01,
|
51 |
+
"total_save_number": 200,
|
52 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
53 |
+
}
|
54 |
+
config.update(additional_config)
|
55 |
+
|
56 |
+
|
57 |
+
|
58 |
+
|
59 |
+
# Data
|
60 |
+
dataset = Dataset(
|
61 |
+
root=config["dataset_root"],
|
62 |
+
download=True,
|
63 |
+
train=True,
|
64 |
+
transform=transforms.Compose([
|
65 |
+
transforms.Resize(80),
|
66 |
+
transforms.RandomHorizontalFlip(),
|
67 |
+
transforms.RandAugment(),
|
68 |
+
transforms.ToTensor(),
|
69 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
70 |
+
])
|
71 |
+
)
|
72 |
+
train_loader = DataLoader(
|
73 |
+
dataset=dataset,
|
74 |
+
batch_size=config["batch_size"],
|
75 |
+
num_workers=config["num_workers"],
|
76 |
+
shuffle=True,
|
77 |
+
drop_last=True,
|
78 |
+
pin_memory=True,
|
79 |
+
persistent_workers=False,
|
80 |
+
)
|
81 |
+
test_loader = DataLoader(
|
82 |
+
dataset=Dataset(
|
83 |
+
root=config["dataset_root"],
|
84 |
+
download=True,
|
85 |
+
train=False,
|
86 |
+
transform=transforms.Compose([
|
87 |
+
transforms.Resize(80),
|
88 |
+
transforms.CenterCrop(80),
|
89 |
+
transforms.ToTensor(),
|
90 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
91 |
+
])),
|
92 |
+
batch_size=config["batch_size"],
|
93 |
+
num_workers=config["num_workers"],
|
94 |
+
shuffle=False,
|
95 |
+
pin_memory=True,
|
96 |
+
persistent_workers=False,
|
97 |
+
pin_memory_device="cuda",
|
98 |
+
)
|
99 |
+
|
100 |
+
# Model
|
101 |
+
model, head = Model()
|
102 |
+
model = model.to(device)
|
103 |
+
criterion = nn.CrossEntropyLoss()
|
104 |
+
optimizer = optim.AdamW(
|
105 |
+
model.parameters(),
|
106 |
+
lr=config["learning_rate"],
|
107 |
+
weight_decay=config["weight_decay"],
|
108 |
+
)
|
109 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
110 |
+
optimizer,
|
111 |
+
T_max=config["epochs"],
|
112 |
+
eta_min=config["save_learning_rate"],
|
113 |
+
)
|
114 |
+
|
115 |
+
|
116 |
+
|
117 |
+
|
118 |
+
# Training
|
119 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
120 |
+
model.train()
|
121 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(train_loader),
|
122 |
+
total=len(dataset) // config["batch_size"]):
|
123 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
124 |
+
optimizer.zero_grad()
|
125 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
126 |
+
outputs = model(inputs)
|
127 |
+
loss = criterion(outputs, targets)
|
128 |
+
loss.backward()
|
129 |
+
optimizer.step()
|
130 |
+
if scheduler is not None:
|
131 |
+
scheduler.step()
|
132 |
+
|
133 |
+
# test
|
134 |
+
@torch.no_grad()
|
135 |
+
def test(model=model):
|
136 |
+
model.eval()
|
137 |
+
all_targets = []
|
138 |
+
all_predicts = []
|
139 |
+
test_loss = 0
|
140 |
+
correct = 0
|
141 |
+
total = 0
|
142 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(test_loader),
|
143 |
+
total=len(test_loader.dataset) // config["batch_size"]):
|
144 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
145 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
146 |
+
outputs = model(inputs)
|
147 |
+
loss = criterion(outputs, targets)
|
148 |
+
# to logging losses
|
149 |
+
all_targets.extend(targets.flatten().tolist())
|
150 |
+
test_loss += loss.item()
|
151 |
+
_, predicts = outputs.max(1)
|
152 |
+
all_predicts.extend(predicts.flatten().tolist())
|
153 |
+
total += targets.size(0)
|
154 |
+
correct += predicts.eq(targets).sum().item()
|
155 |
+
loss = test_loss / (batch_idx + 1)
|
156 |
+
acc = correct / total
|
157 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}\n")
|
158 |
+
model.train()
|
159 |
+
return loss, acc, all_targets, all_predicts
|
160 |
+
|
161 |
+
# save train
|
162 |
+
def save_train(model=model, optimizer=optimizer):
|
163 |
+
model.train()
|
164 |
+
saved_number = 0
|
165 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
166 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
167 |
+
optimizer.zero_grad()
|
168 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
169 |
+
outputs = model(inputs)
|
170 |
+
loss = criterion(outputs, targets)
|
171 |
+
loss.backward()
|
172 |
+
optimizer.step()
|
173 |
+
# Save checkpoint
|
174 |
+
if batch_idx % (len(dataset) // train_loader.batch_size // config["total_save_number"]) == 0:
|
175 |
+
_, acc, _, _ = test(model=model)
|
176 |
+
if not os.path.isdir('checkpoint'):
|
177 |
+
os.mkdir('checkpoint')
|
178 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items() \
|
179 |
+
if key in ["layer4.1.bn1.weight", "layer4.1.bn1.bias", "layer4.1.bn2.weight", "layer4.1.bn2.bias"]}
|
180 |
+
torch.save(save_state, f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
181 |
+
print("save:", f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
182 |
+
saved_number += 1
|
183 |
+
if saved_number >= config["total_save_number"]:
|
184 |
+
break
|
185 |
+
|
186 |
+
|
187 |
+
|
188 |
+
|
189 |
+
# main
|
190 |
+
if __name__ == '__main__':
|
191 |
+
test(model=model)
|
192 |
+
for epoch in range(config["epochs"]):
|
193 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
194 |
+
test(model=model)
|
195 |
+
save_train(model=model, optimizer=optimizer)
|
dataset/cifar10_cnnmedium/model.py
ADDED
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from torch.nn import functional as F
|
4 |
+
import timm
|
5 |
+
|
6 |
+
|
7 |
+
class CNNMedium(nn.Module):
|
8 |
+
def __init__(self):
|
9 |
+
super().__init__()
|
10 |
+
self.module = nn.Sequential(
|
11 |
+
nn.Conv2d(3, 16, 3),
|
12 |
+
nn.MaxPool2d(2, 2),
|
13 |
+
nn.LeakyReLU(),
|
14 |
+
nn.Conv2d(16, 32, 3),
|
15 |
+
nn.MaxPool2d(2, 2),
|
16 |
+
nn.LeakyReLU(),
|
17 |
+
nn.Conv2d(32, 15, 3),
|
18 |
+
nn.MaxPool2d(2, 2),
|
19 |
+
nn.LeakyReLU(),
|
20 |
+
nn.Flatten(start_dim=1),
|
21 |
+
)
|
22 |
+
self.head = nn.Sequential(
|
23 |
+
nn.Linear(60, 20),
|
24 |
+
nn.LeakyReLU(),
|
25 |
+
nn.Linear(20, 10),
|
26 |
+
)
|
27 |
+
|
28 |
+
def forward(self, x):
|
29 |
+
x = self.module(x)
|
30 |
+
x = self.head(x)
|
31 |
+
return x
|
32 |
+
|
33 |
+
|
34 |
+
def Model():
|
35 |
+
model = CNNMedium()
|
36 |
+
return model, model.head
|
37 |
+
|
38 |
+
|
39 |
+
if __name__ == "__main__":
|
40 |
+
model, _ = Model()
|
41 |
+
x = torch.ones([4, 3, 32, 32])
|
42 |
+
y = model(x)
|
43 |
+
print(y.shape)
|
44 |
+
print(model)
|
45 |
+
num_param = 0
|
46 |
+
for v in model.parameters():
|
47 |
+
num_param += v.numel()
|
48 |
+
print("num_param:", num_param)
|
dataset/cifar10_cnnmedium/test.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
if __name__ == "__main__":
|
4 |
+
from train import *
|
5 |
+
else: # relative import
|
6 |
+
from .train import *
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
|
11 |
+
try:
|
12 |
+
test_item = sys.argv[1]
|
13 |
+
except IndexError:
|
14 |
+
assert __name__ == "__main__"
|
15 |
+
test_item = "./checkpoint"
|
16 |
+
test_items = []
|
17 |
+
if os.path.isdir(test_item):
|
18 |
+
for item in os.listdir(test_item):
|
19 |
+
item = os.path.join(test_item, item)
|
20 |
+
test_items.append(item)
|
21 |
+
elif os.path.isfile(test_item):
|
22 |
+
test_items.append(test_item)
|
23 |
+
|
24 |
+
|
25 |
+
for item in test_items:
|
26 |
+
state = torch.load(item, map_location="cpu")
|
27 |
+
model.load_state_dict({key: value.to(torch.float32).to(device) for key, value in state.items()})
|
28 |
+
loss, acc, all_targets, all_predicts = test(model=model)
|
dataset/cifar10_cnnmedium/train.py
ADDED
@@ -0,0 +1,192 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
seed = SEED = 20
|
6 |
+
torch.manual_seed(seed)
|
7 |
+
torch.cuda.manual_seed(seed)
|
8 |
+
torch.cuda.manual_seed_all(seed)
|
9 |
+
torch.backends.cudnn.deterministic = True
|
10 |
+
torch.backends.cudnn.benchmark = True
|
11 |
+
np.random.seed(seed)
|
12 |
+
random.seed(seed)
|
13 |
+
|
14 |
+
|
15 |
+
try: # relative import
|
16 |
+
from model import Model
|
17 |
+
except ImportError:
|
18 |
+
from .model import Model
|
19 |
+
|
20 |
+
# import
|
21 |
+
import torch.nn as nn
|
22 |
+
from torch import optim
|
23 |
+
from torch.optim import lr_scheduler
|
24 |
+
from torch.utils.data import DataLoader
|
25 |
+
import torchvision.transforms as transforms
|
26 |
+
from torchvision.datasets import CIFAR10 as Dataset
|
27 |
+
from tqdm.auto import tqdm
|
28 |
+
import os
|
29 |
+
import warnings
|
30 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
31 |
+
|
32 |
+
# load additional config
|
33 |
+
import json
|
34 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
35 |
+
with open(config_file, "r") as f:
|
36 |
+
additional_config = json.load(f)
|
37 |
+
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
# config
|
42 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
43 |
+
config = {
|
44 |
+
"dataset_root": "from_additional_config",
|
45 |
+
"batch_size": 500 if __name__ == "__main__" else 200,
|
46 |
+
"num_workers": 32,
|
47 |
+
"learning_rate": 1e-2,
|
48 |
+
"weight_decay": 0.00666,
|
49 |
+
"epochs": 50,
|
50 |
+
"save_learning_rate": 1e-5,
|
51 |
+
"total_save_number": 50,
|
52 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
53 |
+
}
|
54 |
+
config.update(additional_config)
|
55 |
+
|
56 |
+
|
57 |
+
|
58 |
+
|
59 |
+
# Data
|
60 |
+
dataset = Dataset(
|
61 |
+
root=config["dataset_root"],
|
62 |
+
download=True,
|
63 |
+
train=True,
|
64 |
+
transform=transforms.Compose([
|
65 |
+
transforms.Resize(32),
|
66 |
+
transforms.RandomCrop(32),
|
67 |
+
transforms.RandomHorizontalFlip(),
|
68 |
+
transforms.AutoAugment(policy=transforms.AutoAugmentPolicy("cifar10")),
|
69 |
+
transforms.ToTensor(),
|
70 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
71 |
+
])
|
72 |
+
)
|
73 |
+
train_loader = DataLoader(
|
74 |
+
dataset=dataset,
|
75 |
+
batch_size=config["batch_size"],
|
76 |
+
num_workers=config["num_workers"],
|
77 |
+
shuffle=True,
|
78 |
+
drop_last=True,
|
79 |
+
pin_memory=True,
|
80 |
+
persistent_workers=True,
|
81 |
+
)
|
82 |
+
test_loader = DataLoader(
|
83 |
+
dataset=Dataset(
|
84 |
+
root=config["dataset_root"],
|
85 |
+
download=True,
|
86 |
+
train=False,
|
87 |
+
transform=transforms.Compose([
|
88 |
+
transforms.Resize(32),
|
89 |
+
transforms.CenterCrop(32),
|
90 |
+
transforms.ToTensor(),
|
91 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
92 |
+
])),
|
93 |
+
batch_size=config["batch_size"],
|
94 |
+
num_workers=config["num_workers"],
|
95 |
+
shuffle=False,
|
96 |
+
pin_memory=True,
|
97 |
+
persistent_workers=True,
|
98 |
+
pin_memory_device="cuda",
|
99 |
+
)
|
100 |
+
|
101 |
+
# Model
|
102 |
+
model, head = Model()
|
103 |
+
model = model.to(device)
|
104 |
+
criterion = nn.CrossEntropyLoss()
|
105 |
+
optimizer = optim.SGD(
|
106 |
+
model.parameters(),
|
107 |
+
lr=config["learning_rate"],
|
108 |
+
weight_decay=config["weight_decay"],
|
109 |
+
momentum=0.9,
|
110 |
+
)
|
111 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
112 |
+
optimizer,
|
113 |
+
T_max=config["epochs"],
|
114 |
+
eta_min=config["save_learning_rate"],
|
115 |
+
)
|
116 |
+
|
117 |
+
|
118 |
+
|
119 |
+
|
120 |
+
# Training
|
121 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
122 |
+
model.train()
|
123 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(train_loader),
|
124 |
+
total=len(dataset) // config["batch_size"]):
|
125 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
126 |
+
optimizer.zero_grad()
|
127 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
128 |
+
outputs = model(inputs)
|
129 |
+
loss = criterion(outputs, targets)
|
130 |
+
loss.backward()
|
131 |
+
optimizer.step()
|
132 |
+
if scheduler is not None:
|
133 |
+
scheduler.step()
|
134 |
+
|
135 |
+
# test
|
136 |
+
@torch.no_grad()
|
137 |
+
def test(model=model):
|
138 |
+
model.eval()
|
139 |
+
all_targets = []
|
140 |
+
all_predicts = []
|
141 |
+
test_loss = 0
|
142 |
+
correct = 0
|
143 |
+
total = 0
|
144 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(test_loader),
|
145 |
+
total=len(test_loader.dataset) // config["batch_size"]):
|
146 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
147 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
148 |
+
outputs = model(inputs)
|
149 |
+
loss = criterion(outputs, targets)
|
150 |
+
# to logging losses
|
151 |
+
all_targets.extend(targets.flatten().tolist())
|
152 |
+
test_loss += loss.item()
|
153 |
+
_, predicts = outputs.max(1)
|
154 |
+
all_predicts.extend(predicts.flatten().tolist())
|
155 |
+
total += targets.size(0)
|
156 |
+
correct += predicts.eq(targets).sum().item()
|
157 |
+
loss = test_loss / (batch_idx + 1)
|
158 |
+
acc = correct / total
|
159 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}\n")
|
160 |
+
model.train()
|
161 |
+
return loss, acc, all_targets, all_predicts
|
162 |
+
|
163 |
+
# save train
|
164 |
+
def save_train(model=model, optimizer=optimizer):
|
165 |
+
model.train()
|
166 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
167 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
168 |
+
optimizer.zero_grad()
|
169 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
170 |
+
outputs = model(inputs)
|
171 |
+
loss = criterion(outputs, targets)
|
172 |
+
loss.backward()
|
173 |
+
optimizer.step()
|
174 |
+
# Save checkpoint
|
175 |
+
if batch_idx % (len(dataset) // train_loader.batch_size // config["total_save_number"]) == 0:
|
176 |
+
_, acc, _, _ = test(model=model)
|
177 |
+
if not os.path.isdir('checkpoint'):
|
178 |
+
os.mkdir('checkpoint')
|
179 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items()}
|
180 |
+
torch.save(save_state, f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
181 |
+
print("save:", f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
182 |
+
|
183 |
+
|
184 |
+
|
185 |
+
|
186 |
+
# main
|
187 |
+
if __name__ == '__main__':
|
188 |
+
test(model=model)
|
189 |
+
for epoch in range(config["epochs"]):
|
190 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
191 |
+
test(model=model)
|
192 |
+
save_train(model=model, optimizer=optimizer)
|
dataset/cifar10_cnnsmall/model.py
ADDED
@@ -0,0 +1,48 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
from torch.nn import functional as F
|
4 |
+
import timm
|
5 |
+
|
6 |
+
|
7 |
+
class CNNSmall(nn.Module):
|
8 |
+
def __init__(self):
|
9 |
+
super().__init__()
|
10 |
+
self.module = nn.Sequential(
|
11 |
+
nn.Conv2d(3, 8, 5),
|
12 |
+
nn.MaxPool2d(2, 2),
|
13 |
+
nn.LeakyReLU(),
|
14 |
+
nn.Conv2d(8, 6, 5),
|
15 |
+
nn.MaxPool2d(2, 2),
|
16 |
+
nn.LeakyReLU(),
|
17 |
+
nn.Conv2d(6, 4, 2),
|
18 |
+
nn.LeakyReLU(),
|
19 |
+
nn.Flatten(start_dim=1),
|
20 |
+
)
|
21 |
+
self.head = nn.Sequential(
|
22 |
+
nn.Linear(36, 20),
|
23 |
+
nn.LeakyReLU(),
|
24 |
+
nn.Linear(20, 10),
|
25 |
+
)
|
26 |
+
|
27 |
+
def forward(self, x):
|
28 |
+
x = F.interpolate(x, (28, 28), mode='bilinear')
|
29 |
+
x = self.module(x)
|
30 |
+
x = self.head(x)
|
31 |
+
return x
|
32 |
+
|
33 |
+
|
34 |
+
def Model():
|
35 |
+
model = CNNSmall()
|
36 |
+
return model, model.head
|
37 |
+
|
38 |
+
|
39 |
+
if __name__ == "__main__":
|
40 |
+
model, _ = Model()
|
41 |
+
x = torch.ones([4, 3, 28, 28])
|
42 |
+
y = model(x)
|
43 |
+
print(y.shape)
|
44 |
+
print(model)
|
45 |
+
num_param = 0
|
46 |
+
for v in model.parameters():
|
47 |
+
num_param += v.numel()
|
48 |
+
print("num_param:", num_param)
|
dataset/cifar10_cnnsmall/test.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
if __name__ == "__main__":
|
4 |
+
from train import *
|
5 |
+
else: # relative import
|
6 |
+
from .train import *
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
|
11 |
+
try:
|
12 |
+
test_item = sys.argv[1]
|
13 |
+
except IndexError:
|
14 |
+
assert __name__ == "__main__"
|
15 |
+
test_item = "./checkpoint"
|
16 |
+
test_items = []
|
17 |
+
if os.path.isdir(test_item):
|
18 |
+
for item in os.listdir(test_item):
|
19 |
+
item = os.path.join(test_item, item)
|
20 |
+
test_items.append(item)
|
21 |
+
elif os.path.isfile(test_item):
|
22 |
+
test_items.append(test_item)
|
23 |
+
|
24 |
+
|
25 |
+
for item in test_items:
|
26 |
+
state = torch.load(item, map_location="cpu")
|
27 |
+
model.load_state_dict({key: value.to(torch.float32).to(device) for key, value in state.items()})
|
28 |
+
loss, acc, all_targets, all_predicts = test(model=model)
|
dataset/cifar10_cnnsmall/train.py
ADDED
@@ -0,0 +1,192 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
seed = SEED = 20
|
6 |
+
torch.manual_seed(seed)
|
7 |
+
torch.cuda.manual_seed(seed)
|
8 |
+
torch.cuda.manual_seed_all(seed)
|
9 |
+
torch.backends.cudnn.deterministic = True
|
10 |
+
torch.backends.cudnn.benchmark = True
|
11 |
+
np.random.seed(seed)
|
12 |
+
random.seed(seed)
|
13 |
+
|
14 |
+
|
15 |
+
try: # relative import
|
16 |
+
from model import Model
|
17 |
+
except ImportError:
|
18 |
+
from .model import Model
|
19 |
+
|
20 |
+
# import
|
21 |
+
import torch.nn as nn
|
22 |
+
from torch import optim
|
23 |
+
from torch.optim import lr_scheduler
|
24 |
+
from torch.utils.data import DataLoader
|
25 |
+
import torchvision.transforms as transforms
|
26 |
+
from torchvision.datasets import CIFAR10 as Dataset
|
27 |
+
from tqdm.auto import tqdm
|
28 |
+
import os
|
29 |
+
import warnings
|
30 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
31 |
+
|
32 |
+
# load additional config
|
33 |
+
import json
|
34 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
35 |
+
with open(config_file, "r") as f:
|
36 |
+
additional_config = json.load(f)
|
37 |
+
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
# config
|
42 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
43 |
+
config = {
|
44 |
+
"dataset_root": "from_additional_config",
|
45 |
+
"batch_size": 500 if __name__ == "__main__" else 200,
|
46 |
+
"num_workers": 32,
|
47 |
+
"learning_rate": 1e-2,
|
48 |
+
"weight_decay": 0.001,
|
49 |
+
"epochs": 50,
|
50 |
+
"save_learning_rate": 1e-5,
|
51 |
+
"total_save_number": 50,
|
52 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
53 |
+
}
|
54 |
+
config.update(additional_config)
|
55 |
+
|
56 |
+
|
57 |
+
|
58 |
+
|
59 |
+
# Data
|
60 |
+
dataset = Dataset(
|
61 |
+
root=config["dataset_root"],
|
62 |
+
download=True,
|
63 |
+
train=True,
|
64 |
+
transform=transforms.Compose([
|
65 |
+
transforms.Resize(32),
|
66 |
+
transforms.RandomCrop(32),
|
67 |
+
transforms.RandomHorizontalFlip(),
|
68 |
+
transforms.AutoAugment(policy=transforms.AutoAugmentPolicy("cifar10")),
|
69 |
+
transforms.ToTensor(),
|
70 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
71 |
+
])
|
72 |
+
)
|
73 |
+
train_loader = DataLoader(
|
74 |
+
dataset=dataset,
|
75 |
+
batch_size=config["batch_size"],
|
76 |
+
num_workers=config["num_workers"],
|
77 |
+
shuffle=True,
|
78 |
+
drop_last=True,
|
79 |
+
pin_memory=True,
|
80 |
+
persistent_workers=True,
|
81 |
+
)
|
82 |
+
test_loader = DataLoader(
|
83 |
+
dataset=Dataset(
|
84 |
+
root=config["dataset_root"],
|
85 |
+
download=True,
|
86 |
+
train=False,
|
87 |
+
transform=transforms.Compose([
|
88 |
+
transforms.Resize(32),
|
89 |
+
transforms.CenterCrop(32),
|
90 |
+
transforms.ToTensor(),
|
91 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
92 |
+
])),
|
93 |
+
batch_size=config["batch_size"],
|
94 |
+
num_workers=config["num_workers"],
|
95 |
+
shuffle=False,
|
96 |
+
pin_memory=True,
|
97 |
+
persistent_workers=True,
|
98 |
+
pin_memory_device="cuda",
|
99 |
+
)
|
100 |
+
|
101 |
+
# Model
|
102 |
+
model, head = Model()
|
103 |
+
model = model.to(device)
|
104 |
+
criterion = nn.CrossEntropyLoss()
|
105 |
+
optimizer = optim.SGD(
|
106 |
+
model.parameters(),
|
107 |
+
lr=config["learning_rate"],
|
108 |
+
weight_decay=config["weight_decay"],
|
109 |
+
momentum=0.9,
|
110 |
+
)
|
111 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
112 |
+
optimizer,
|
113 |
+
T_max=config["epochs"],
|
114 |
+
eta_min=config["save_learning_rate"],
|
115 |
+
)
|
116 |
+
|
117 |
+
|
118 |
+
|
119 |
+
|
120 |
+
# Training
|
121 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
122 |
+
model.train()
|
123 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(train_loader),
|
124 |
+
total=len(dataset) // config["batch_size"]):
|
125 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
126 |
+
optimizer.zero_grad()
|
127 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
128 |
+
outputs = model(inputs)
|
129 |
+
loss = criterion(outputs, targets)
|
130 |
+
loss.backward()
|
131 |
+
optimizer.step()
|
132 |
+
if scheduler is not None:
|
133 |
+
scheduler.step()
|
134 |
+
|
135 |
+
# test
|
136 |
+
@torch.no_grad()
|
137 |
+
def test(model=model):
|
138 |
+
model.eval()
|
139 |
+
all_targets = []
|
140 |
+
all_predicts = []
|
141 |
+
test_loss = 0
|
142 |
+
correct = 0
|
143 |
+
total = 0
|
144 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(test_loader),
|
145 |
+
total=len(test_loader.dataset) // config["batch_size"]):
|
146 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
147 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
148 |
+
outputs = model(inputs)
|
149 |
+
loss = criterion(outputs, targets)
|
150 |
+
# to logging losses
|
151 |
+
all_targets.extend(targets.flatten().tolist())
|
152 |
+
test_loss += loss.item()
|
153 |
+
_, predicts = outputs.max(1)
|
154 |
+
all_predicts.extend(predicts.flatten().tolist())
|
155 |
+
total += targets.size(0)
|
156 |
+
correct += predicts.eq(targets).sum().item()
|
157 |
+
loss = test_loss / (batch_idx + 1)
|
158 |
+
acc = correct / total
|
159 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}\n")
|
160 |
+
model.train()
|
161 |
+
return loss, acc, all_targets, all_predicts
|
162 |
+
|
163 |
+
# save train
|
164 |
+
def save_train(model=model, optimizer=optimizer):
|
165 |
+
model.train()
|
166 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
167 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
168 |
+
optimizer.zero_grad()
|
169 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
170 |
+
outputs = model(inputs)
|
171 |
+
loss = criterion(outputs, targets)
|
172 |
+
loss.backward()
|
173 |
+
optimizer.step()
|
174 |
+
# Save checkpoint
|
175 |
+
if batch_idx % (len(dataset) // train_loader.batch_size // config["total_save_number"]) == 0:
|
176 |
+
_, acc, _, _ = test(model=model)
|
177 |
+
if not os.path.isdir('checkpoint'):
|
178 |
+
os.mkdir('checkpoint')
|
179 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items()}
|
180 |
+
torch.save(save_state, f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
181 |
+
print("save:", f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
182 |
+
|
183 |
+
|
184 |
+
|
185 |
+
|
186 |
+
# main
|
187 |
+
if __name__ == '__main__':
|
188 |
+
test(model=model)
|
189 |
+
for epoch in range(config["epochs"]):
|
190 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
191 |
+
test(model=model)
|
192 |
+
save_train(model=model, optimizer=optimizer)
|
dataset/cifar10_mobilenetv3/model.py
ADDED
@@ -0,0 +1,21 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import timm
|
3 |
+
|
4 |
+
|
5 |
+
def Model():
|
6 |
+
model = timm.create_model("mobilenetv3_large_100", pretrained=True)
|
7 |
+
model.classifier = nn.Linear(1280, 10)
|
8 |
+
for name, param in model.named_parameters():
|
9 |
+
if "bn" in name:
|
10 |
+
# print(f"freeze {name}")
|
11 |
+
param.requires_grad = False
|
12 |
+
return model, model.classifier
|
13 |
+
|
14 |
+
|
15 |
+
if __name__ == "__main__":
|
16 |
+
model, _ = Model()
|
17 |
+
print(model)
|
18 |
+
num_param = 0
|
19 |
+
for v in model.parameters():
|
20 |
+
num_param += v.numel()
|
21 |
+
print("num_param:", num_param)
|
dataset/cifar10_mobilenetv3/test.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
if __name__ == "__main__":
|
4 |
+
from train import *
|
5 |
+
else: # relative import
|
6 |
+
from .train import *
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
|
11 |
+
try:
|
12 |
+
test_item = sys.argv[1]
|
13 |
+
except IndexError:
|
14 |
+
assert __name__ == "__main__"
|
15 |
+
test_item = "./checkpoint"
|
16 |
+
test_items = []
|
17 |
+
if os.path.isdir(test_item):
|
18 |
+
for item in os.listdir(test_item):
|
19 |
+
item = os.path.join(test_item, item)
|
20 |
+
test_items.append(item)
|
21 |
+
elif os.path.isfile(test_item):
|
22 |
+
test_items.append(test_item)
|
23 |
+
|
24 |
+
|
25 |
+
for item in test_items:
|
26 |
+
state = torch.load(item, map_location="cpu")
|
27 |
+
model.load_state_dict({key: value.to(torch.float32).to(device) for key, value in state.items()})
|
28 |
+
loss, acc, all_targets, all_predicts = test(model=model)
|
dataset/cifar10_mobilenetv3/train.py
ADDED
@@ -0,0 +1,199 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
seed = SEED = 20
|
6 |
+
torch.manual_seed(seed)
|
7 |
+
torch.cuda.manual_seed(seed)
|
8 |
+
torch.cuda.manual_seed_all(seed)
|
9 |
+
torch.backends.cudnn.deterministic = True
|
10 |
+
torch.backends.cudnn.benchmark = True
|
11 |
+
np.random.seed(seed)
|
12 |
+
random.seed(seed)
|
13 |
+
|
14 |
+
|
15 |
+
try: # relative import
|
16 |
+
from model import Model
|
17 |
+
except ImportError:
|
18 |
+
from .model import Model
|
19 |
+
|
20 |
+
# import
|
21 |
+
import torch.nn as nn
|
22 |
+
from torch import optim
|
23 |
+
from torch.optim import lr_scheduler
|
24 |
+
from torch.utils.data import DataLoader
|
25 |
+
import torchvision.transforms as transforms
|
26 |
+
from torchvision.datasets import CIFAR10 as Dataset
|
27 |
+
from tqdm.auto import tqdm
|
28 |
+
import os
|
29 |
+
import warnings
|
30 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
31 |
+
|
32 |
+
# load additional config
|
33 |
+
import json
|
34 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
35 |
+
with open(config_file, "r") as f:
|
36 |
+
additional_config = json.load(f)
|
37 |
+
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
# config
|
42 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
43 |
+
config = {
|
44 |
+
"dataset_root": "from_additional_config",
|
45 |
+
"batch_size": 500 if __name__ == "__main__" else 200,
|
46 |
+
"num_workers": 4,
|
47 |
+
"learning_rate": 3e-3,
|
48 |
+
"weight_decay": 0.1,
|
49 |
+
"epochs": 5,
|
50 |
+
"save_learning_rate": 1e-6,
|
51 |
+
"total_save_number": 50,
|
52 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
53 |
+
}
|
54 |
+
config.update(additional_config)
|
55 |
+
|
56 |
+
|
57 |
+
|
58 |
+
|
59 |
+
# Data
|
60 |
+
dataset = Dataset(
|
61 |
+
root=config["dataset_root"],
|
62 |
+
download=True,
|
63 |
+
train=True,
|
64 |
+
transform=transforms.Compose([
|
65 |
+
transforms.Resize(224),
|
66 |
+
transforms.RandomCrop(224),
|
67 |
+
transforms.RandomHorizontalFlip(),
|
68 |
+
transforms.AutoAugment(policy=transforms.AutoAugmentPolicy("cifar10")),
|
69 |
+
transforms.ToTensor(),
|
70 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
71 |
+
])
|
72 |
+
)
|
73 |
+
train_loader = DataLoader(
|
74 |
+
dataset=dataset,
|
75 |
+
batch_size=config["batch_size"],
|
76 |
+
num_workers=config["num_workers"],
|
77 |
+
shuffle=True,
|
78 |
+
drop_last=True,
|
79 |
+
pin_memory=True,
|
80 |
+
persistent_workers=True,
|
81 |
+
)
|
82 |
+
test_loader = DataLoader(
|
83 |
+
dataset=Dataset(
|
84 |
+
root=config["dataset_root"],
|
85 |
+
download=True,
|
86 |
+
train=False,
|
87 |
+
transform=transforms.Compose([
|
88 |
+
transforms.Resize(224),
|
89 |
+
transforms.CenterCrop(224),
|
90 |
+
transforms.ToTensor(),
|
91 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
92 |
+
])),
|
93 |
+
batch_size=config["batch_size"],
|
94 |
+
num_workers=config["num_workers"],
|
95 |
+
shuffle=False,
|
96 |
+
pin_memory=True,
|
97 |
+
persistent_workers=True,
|
98 |
+
pin_memory_device="cuda",
|
99 |
+
)
|
100 |
+
|
101 |
+
# Model
|
102 |
+
model, head = Model()
|
103 |
+
model = model.to(device)
|
104 |
+
criterion = nn.CrossEntropyLoss()
|
105 |
+
pre_optimizer = optim.AdamW(
|
106 |
+
head.parameters(),
|
107 |
+
lr=0.05,
|
108 |
+
weight_decay=0.01,
|
109 |
+
)
|
110 |
+
optimizer = optim.AdamW(
|
111 |
+
model.parameters(),
|
112 |
+
lr=config["learning_rate"],
|
113 |
+
weight_decay=config["weight_decay"],
|
114 |
+
)
|
115 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
116 |
+
optimizer,
|
117 |
+
T_max=config["epochs"],
|
118 |
+
eta_min=config["save_learning_rate"],
|
119 |
+
)
|
120 |
+
|
121 |
+
|
122 |
+
|
123 |
+
|
124 |
+
# Training
|
125 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
126 |
+
model.train()
|
127 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(train_loader),
|
128 |
+
total=len(dataset) // config["batch_size"]):
|
129 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
130 |
+
optimizer.zero_grad()
|
131 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
132 |
+
outputs = model(inputs)
|
133 |
+
loss = criterion(outputs, targets)
|
134 |
+
loss.backward()
|
135 |
+
optimizer.step()
|
136 |
+
if scheduler is not None:
|
137 |
+
scheduler.step()
|
138 |
+
|
139 |
+
# test
|
140 |
+
@torch.no_grad()
|
141 |
+
def test(model=model):
|
142 |
+
model.eval()
|
143 |
+
all_targets = []
|
144 |
+
all_predicts = []
|
145 |
+
test_loss = 0
|
146 |
+
correct = 0
|
147 |
+
total = 0
|
148 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(test_loader),
|
149 |
+
total=len(test_loader.dataset) // config["batch_size"]):
|
150 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
151 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
152 |
+
outputs = model(inputs)
|
153 |
+
loss = criterion(outputs, targets)
|
154 |
+
# to logging losses
|
155 |
+
all_targets.extend(targets.flatten().tolist())
|
156 |
+
test_loss += loss.item()
|
157 |
+
_, predicts = outputs.max(1)
|
158 |
+
all_predicts.extend(predicts.flatten().tolist())
|
159 |
+
total += targets.size(0)
|
160 |
+
correct += predicts.eq(targets).sum().item()
|
161 |
+
loss = test_loss / (batch_idx + 1)
|
162 |
+
acc = correct / total
|
163 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}\n")
|
164 |
+
model.train()
|
165 |
+
return loss, acc, all_targets, all_predicts
|
166 |
+
|
167 |
+
# save train
|
168 |
+
def save_train(model=model, optimizer=optimizer):
|
169 |
+
model.train()
|
170 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
171 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
172 |
+
optimizer.zero_grad()
|
173 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
174 |
+
outputs = model(inputs)
|
175 |
+
loss = criterion(outputs, targets)
|
176 |
+
loss.backward()
|
177 |
+
optimizer.step()
|
178 |
+
# Save checkpoint
|
179 |
+
if batch_idx % (len(dataset) // train_loader.batch_size // config["total_save_number"]) == 0:
|
180 |
+
_, acc, _, _ = test(model=model)
|
181 |
+
if not os.path.isdir('checkpoint'):
|
182 |
+
os.mkdir('checkpoint')
|
183 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items()}
|
184 |
+
torch.save(save_state, f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
185 |
+
print("save:", f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
186 |
+
|
187 |
+
|
188 |
+
|
189 |
+
|
190 |
+
# main
|
191 |
+
if __name__ == '__main__':
|
192 |
+
test(model=model)
|
193 |
+
for _ in range(1):
|
194 |
+
train(model=model, optimizer=pre_optimizer)
|
195 |
+
test(model=model)
|
196 |
+
for epoch in range(config["epochs"]):
|
197 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
198 |
+
test(model=model)
|
199 |
+
save_train(model=model, optimizer=optimizer)
|
dataset/cifar10_resnet18/model.py
ADDED
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import timm
|
3 |
+
|
4 |
+
|
5 |
+
def Model():
|
6 |
+
model = timm.create_model("resnet18", pretrained=True)
|
7 |
+
model.fc = nn.Linear(512, 10)
|
8 |
+
return model, model.fc
|
9 |
+
|
10 |
+
|
11 |
+
if __name__ == "__main__":
|
12 |
+
model, _ = Model()
|
13 |
+
print(model)
|
14 |
+
num_param = 0
|
15 |
+
for v in model.parameters():
|
16 |
+
num_param += v.numel()
|
17 |
+
print("num_param:", num_param)
|
dataset/cifar10_resnet18/test.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
if __name__ == "__main__":
|
4 |
+
from train import *
|
5 |
+
else: # relative import
|
6 |
+
from .train import *
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
|
11 |
+
try:
|
12 |
+
test_item = sys.argv[1]
|
13 |
+
except IndexError:
|
14 |
+
assert __name__ == "__main__"
|
15 |
+
test_item = "./checkpoint"
|
16 |
+
test_items = []
|
17 |
+
if os.path.isdir(test_item):
|
18 |
+
for item in os.listdir(test_item):
|
19 |
+
item = os.path.join(test_item, item)
|
20 |
+
test_items.append(item)
|
21 |
+
elif os.path.isfile(test_item):
|
22 |
+
test_items.append(test_item)
|
23 |
+
|
24 |
+
|
25 |
+
for item in test_items:
|
26 |
+
state = torch.load(item, map_location="cpu")
|
27 |
+
model.load_state_dict({key: value.to(torch.float32).to(device) for key, value in state.items()})
|
28 |
+
loss, acc, all_targets, all_predicts = test(model=model)
|
dataset/cifar10_resnet18/train.py
ADDED
@@ -0,0 +1,191 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
seed = SEED = 20
|
6 |
+
torch.manual_seed(seed)
|
7 |
+
torch.cuda.manual_seed(seed)
|
8 |
+
torch.cuda.manual_seed_all(seed)
|
9 |
+
torch.backends.cudnn.deterministic = True
|
10 |
+
torch.backends.cudnn.benchmark = True
|
11 |
+
np.random.seed(seed)
|
12 |
+
random.seed(seed)
|
13 |
+
|
14 |
+
|
15 |
+
try: # relative import
|
16 |
+
from model import Model
|
17 |
+
except ImportError:
|
18 |
+
from .model import Model
|
19 |
+
|
20 |
+
# import
|
21 |
+
import torch.nn as nn
|
22 |
+
from torch import optim
|
23 |
+
from torch.optim import lr_scheduler
|
24 |
+
from torch.utils.data import DataLoader
|
25 |
+
import torchvision.transforms as transforms
|
26 |
+
from torchvision.datasets import CIFAR10 as Dataset
|
27 |
+
from tqdm.auto import tqdm
|
28 |
+
import os
|
29 |
+
import warnings
|
30 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
31 |
+
|
32 |
+
# load additional config
|
33 |
+
import json
|
34 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
35 |
+
with open(config_file, "r") as f:
|
36 |
+
additional_config = json.load(f)
|
37 |
+
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
# config
|
42 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
43 |
+
config = {
|
44 |
+
"dataset_root": "from_additional_config",
|
45 |
+
"batch_size": 500 if __name__ == "__main__" else 200,
|
46 |
+
"num_workers": 32,
|
47 |
+
"learning_rate": 3e-3,
|
48 |
+
"weight_decay": 0.1,
|
49 |
+
"epochs": 50,
|
50 |
+
"save_learning_rate": 1e-5,
|
51 |
+
"total_save_number": 50,
|
52 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
53 |
+
}
|
54 |
+
config.update(additional_config)
|
55 |
+
|
56 |
+
|
57 |
+
|
58 |
+
|
59 |
+
# Data
|
60 |
+
dataset = Dataset(
|
61 |
+
root=config["dataset_root"],
|
62 |
+
download=True,
|
63 |
+
train=True,
|
64 |
+
transform=transforms.Compose([
|
65 |
+
transforms.Resize(64),
|
66 |
+
transforms.RandomCrop(64),
|
67 |
+
transforms.RandomHorizontalFlip(),
|
68 |
+
transforms.AutoAugment(policy=transforms.AutoAugmentPolicy("cifar10")),
|
69 |
+
transforms.ToTensor(),
|
70 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
71 |
+
])
|
72 |
+
)
|
73 |
+
train_loader = DataLoader(
|
74 |
+
dataset=dataset,
|
75 |
+
batch_size=config["batch_size"],
|
76 |
+
num_workers=config["num_workers"],
|
77 |
+
shuffle=True,
|
78 |
+
drop_last=True,
|
79 |
+
pin_memory=True,
|
80 |
+
persistent_workers=True,
|
81 |
+
)
|
82 |
+
test_loader = DataLoader(
|
83 |
+
dataset=Dataset(
|
84 |
+
root=config["dataset_root"],
|
85 |
+
download=True,
|
86 |
+
train=False,
|
87 |
+
transform=transforms.Compose([
|
88 |
+
transforms.Resize(64),
|
89 |
+
transforms.CenterCrop(64),
|
90 |
+
transforms.ToTensor(),
|
91 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
92 |
+
])),
|
93 |
+
batch_size=config["batch_size"],
|
94 |
+
num_workers=config["num_workers"],
|
95 |
+
shuffle=False,
|
96 |
+
pin_memory=True,
|
97 |
+
persistent_workers=True,
|
98 |
+
pin_memory_device="cuda",
|
99 |
+
)
|
100 |
+
|
101 |
+
# Model
|
102 |
+
model, head = Model()
|
103 |
+
model = model.to(device)
|
104 |
+
criterion = nn.CrossEntropyLoss()
|
105 |
+
optimizer = optim.AdamW(
|
106 |
+
model.parameters(),
|
107 |
+
lr=config["learning_rate"],
|
108 |
+
weight_decay=config["weight_decay"],
|
109 |
+
)
|
110 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
111 |
+
optimizer,
|
112 |
+
T_max=config["epochs"],
|
113 |
+
eta_min=config["save_learning_rate"],
|
114 |
+
)
|
115 |
+
|
116 |
+
|
117 |
+
|
118 |
+
|
119 |
+
# Training
|
120 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
121 |
+
model.train()
|
122 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(train_loader),
|
123 |
+
total=len(dataset) // config["batch_size"]):
|
124 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
125 |
+
optimizer.zero_grad()
|
126 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
127 |
+
outputs = model(inputs)
|
128 |
+
loss = criterion(outputs, targets)
|
129 |
+
loss.backward()
|
130 |
+
optimizer.step()
|
131 |
+
if scheduler is not None:
|
132 |
+
scheduler.step()
|
133 |
+
|
134 |
+
# test
|
135 |
+
@torch.no_grad()
|
136 |
+
def test(model=model):
|
137 |
+
model.eval()
|
138 |
+
all_targets = []
|
139 |
+
all_predicts = []
|
140 |
+
test_loss = 0
|
141 |
+
correct = 0
|
142 |
+
total = 0
|
143 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(test_loader),
|
144 |
+
total=len(test_loader.dataset) // config["batch_size"]):
|
145 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
146 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
147 |
+
outputs = model(inputs)
|
148 |
+
loss = criterion(outputs, targets)
|
149 |
+
# to logging losses
|
150 |
+
all_targets.extend(targets.flatten().tolist())
|
151 |
+
test_loss += loss.item()
|
152 |
+
_, predicts = outputs.max(1)
|
153 |
+
all_predicts.extend(predicts.flatten().tolist())
|
154 |
+
total += targets.size(0)
|
155 |
+
correct += predicts.eq(targets).sum().item()
|
156 |
+
loss = test_loss / (batch_idx + 1)
|
157 |
+
acc = correct / total
|
158 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}\n")
|
159 |
+
model.train()
|
160 |
+
return loss, acc, all_targets, all_predicts
|
161 |
+
|
162 |
+
# save train
|
163 |
+
def save_train(model=model, optimizer=optimizer):
|
164 |
+
model.train()
|
165 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
166 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
167 |
+
optimizer.zero_grad()
|
168 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
169 |
+
outputs = model(inputs)
|
170 |
+
loss = criterion(outputs, targets)
|
171 |
+
loss.backward()
|
172 |
+
optimizer.step()
|
173 |
+
# Save checkpoint
|
174 |
+
if batch_idx % (len(dataset) // train_loader.batch_size // config["total_save_number"]) == 0:
|
175 |
+
_, acc, _, _ = test(model=model)
|
176 |
+
if not os.path.isdir('checkpoint'):
|
177 |
+
os.mkdir('checkpoint')
|
178 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items()}
|
179 |
+
torch.save(save_state, f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
180 |
+
print("save:", f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
181 |
+
|
182 |
+
|
183 |
+
|
184 |
+
|
185 |
+
# main
|
186 |
+
if __name__ == '__main__':
|
187 |
+
test(model=model)
|
188 |
+
for epoch in range(config["epochs"]):
|
189 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
190 |
+
test(model=model)
|
191 |
+
save_train(model=model, optimizer=optimizer)
|
dataset/cifar10_vitbase/model.py
ADDED
@@ -0,0 +1,17 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch.nn as nn
|
2 |
+
import timm
|
3 |
+
|
4 |
+
|
5 |
+
def Model():
|
6 |
+
model = timm.create_model("vit_base_patch16_224", pretrained=True)
|
7 |
+
model.head = nn.Linear(768, 10)
|
8 |
+
return model, model.head
|
9 |
+
|
10 |
+
|
11 |
+
if __name__ == "__main__":
|
12 |
+
model, _ = Model()
|
13 |
+
print(model)
|
14 |
+
num_param = 0
|
15 |
+
for v in model.parameters():
|
16 |
+
num_param += v.numel()
|
17 |
+
print("num_param:", num_param)
|
dataset/cifar10_vitbase/test.py
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
if __name__ == "__main__":
|
4 |
+
from train import *
|
5 |
+
else: # relative import
|
6 |
+
from .train import *
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
|
11 |
+
try:
|
12 |
+
test_item = sys.argv[1]
|
13 |
+
except IndexError:
|
14 |
+
assert __name__ == "__main__"
|
15 |
+
test_item = "./checkpoint"
|
16 |
+
test_items = []
|
17 |
+
if os.path.isdir(test_item):
|
18 |
+
for item in os.listdir(test_item):
|
19 |
+
item = os.path.join(test_item, item)
|
20 |
+
test_items.append(item)
|
21 |
+
elif os.path.isfile(test_item):
|
22 |
+
test_items.append(test_item)
|
23 |
+
|
24 |
+
|
25 |
+
for item in test_items:
|
26 |
+
state = torch.load(item, map_location="cpu")
|
27 |
+
model.load_state_dict({key: value.to(torch.float32).to(device) for key, value in state.items()})
|
28 |
+
loss, acc, all_targets, all_predicts = test(model=model)
|
dataset/cifar10_vitbase/train.py
ADDED
@@ -0,0 +1,199 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
seed = SEED = 20
|
6 |
+
torch.manual_seed(seed)
|
7 |
+
torch.cuda.manual_seed(seed)
|
8 |
+
torch.cuda.manual_seed_all(seed)
|
9 |
+
torch.backends.cudnn.deterministic = True
|
10 |
+
torch.backends.cudnn.benchmark = True
|
11 |
+
np.random.seed(seed)
|
12 |
+
random.seed(seed)
|
13 |
+
|
14 |
+
|
15 |
+
try: # relative import
|
16 |
+
from model import Model
|
17 |
+
except ImportError:
|
18 |
+
from .model import Model
|
19 |
+
|
20 |
+
# import
|
21 |
+
import torch.nn as nn
|
22 |
+
from torch import optim
|
23 |
+
from torch.optim import lr_scheduler
|
24 |
+
from torch.utils.data import DataLoader
|
25 |
+
import torchvision.transforms as transforms
|
26 |
+
from torchvision.datasets import CIFAR10 as Dataset
|
27 |
+
from tqdm.auto import tqdm
|
28 |
+
import os
|
29 |
+
import warnings
|
30 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
31 |
+
|
32 |
+
# load additional config
|
33 |
+
import json
|
34 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
35 |
+
with open(config_file, "r") as f:
|
36 |
+
additional_config = json.load(f)
|
37 |
+
|
38 |
+
|
39 |
+
|
40 |
+
|
41 |
+
# config
|
42 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
43 |
+
config = {
|
44 |
+
"dataset_root": "from_additional_config",
|
45 |
+
"batch_size": 500 if __name__ == "__main__" else 200,
|
46 |
+
"num_workers": 32,
|
47 |
+
"learning_rate": 3e-5,
|
48 |
+
"weight_decay": 0.1,
|
49 |
+
"epochs": 7,
|
50 |
+
"save_learning_rate": 1e-5,
|
51 |
+
"total_save_number": 50,
|
52 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
53 |
+
}
|
54 |
+
config.update(additional_config)
|
55 |
+
|
56 |
+
|
57 |
+
|
58 |
+
|
59 |
+
# Data
|
60 |
+
dataset = Dataset(
|
61 |
+
root=config["dataset_root"],
|
62 |
+
download=True,
|
63 |
+
train=True,
|
64 |
+
transform=transforms.Compose([
|
65 |
+
transforms.Resize(224),
|
66 |
+
transforms.RandomCrop(224),
|
67 |
+
transforms.RandomHorizontalFlip(),
|
68 |
+
transforms.AutoAugment(policy=transforms.AutoAugmentPolicy("cifar10")),
|
69 |
+
transforms.ToTensor(),
|
70 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
71 |
+
])
|
72 |
+
)
|
73 |
+
train_loader = DataLoader(
|
74 |
+
dataset=dataset,
|
75 |
+
batch_size=config["batch_size"],
|
76 |
+
num_workers=config["num_workers"],
|
77 |
+
shuffle=True,
|
78 |
+
drop_last=True,
|
79 |
+
pin_memory=True,
|
80 |
+
persistent_workers=True,
|
81 |
+
)
|
82 |
+
test_loader = DataLoader(
|
83 |
+
dataset=Dataset(
|
84 |
+
root=config["dataset_root"],
|
85 |
+
download=True,
|
86 |
+
train=False,
|
87 |
+
transform=transforms.Compose([
|
88 |
+
transforms.Resize(224),
|
89 |
+
transforms.CenterCrop(224),
|
90 |
+
transforms.ToTensor(),
|
91 |
+
transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)),
|
92 |
+
])),
|
93 |
+
batch_size=config["batch_size"],
|
94 |
+
num_workers=config["num_workers"],
|
95 |
+
shuffle=False,
|
96 |
+
pin_memory=True,
|
97 |
+
persistent_workers=True,
|
98 |
+
pin_memory_device="cuda",
|
99 |
+
)
|
100 |
+
|
101 |
+
# Model
|
102 |
+
model, head = Model()
|
103 |
+
model = model.to(device)
|
104 |
+
criterion = nn.CrossEntropyLoss()
|
105 |
+
pre_optimizer = optim.AdamW(
|
106 |
+
head.parameters(),
|
107 |
+
lr=0.05,
|
108 |
+
weight_decay=0.01,
|
109 |
+
)
|
110 |
+
optimizer = optim.AdamW(
|
111 |
+
model.parameters(),
|
112 |
+
lr=config["learning_rate"],
|
113 |
+
weight_decay=config["weight_decay"],
|
114 |
+
)
|
115 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
116 |
+
optimizer,
|
117 |
+
T_max=config["epochs"],
|
118 |
+
eta_min=config["save_learning_rate"],
|
119 |
+
)
|
120 |
+
|
121 |
+
|
122 |
+
|
123 |
+
|
124 |
+
# Training
|
125 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
126 |
+
model.train()
|
127 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(train_loader),
|
128 |
+
total=len(dataset) // config["batch_size"]):
|
129 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
130 |
+
optimizer.zero_grad()
|
131 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
132 |
+
outputs = model(inputs)
|
133 |
+
loss = criterion(outputs, targets)
|
134 |
+
loss.backward()
|
135 |
+
optimizer.step()
|
136 |
+
if scheduler is not None:
|
137 |
+
scheduler.step()
|
138 |
+
|
139 |
+
# test
|
140 |
+
@torch.no_grad()
|
141 |
+
def test(model=model):
|
142 |
+
model.eval()
|
143 |
+
all_targets = []
|
144 |
+
all_predicts = []
|
145 |
+
test_loss = 0
|
146 |
+
correct = 0
|
147 |
+
total = 0
|
148 |
+
for batch_idx, (inputs, targets) in tqdm(enumerate(test_loader),
|
149 |
+
total=len(test_loader.dataset) // config["batch_size"]):
|
150 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
151 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
152 |
+
outputs = model(inputs)
|
153 |
+
loss = criterion(outputs, targets)
|
154 |
+
# to logging losses
|
155 |
+
all_targets.extend(targets.flatten().tolist())
|
156 |
+
test_loss += loss.item()
|
157 |
+
_, predicts = outputs.max(1)
|
158 |
+
all_predicts.extend(predicts.flatten().tolist())
|
159 |
+
total += targets.size(0)
|
160 |
+
correct += predicts.eq(targets).sum().item()
|
161 |
+
loss = test_loss / (batch_idx + 1)
|
162 |
+
acc = correct / total
|
163 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}\n")
|
164 |
+
model.train()
|
165 |
+
return loss, acc, all_targets, all_predicts
|
166 |
+
|
167 |
+
# save train
|
168 |
+
def save_train(model=model, optimizer=optimizer):
|
169 |
+
model.train()
|
170 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
171 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
172 |
+
optimizer.zero_grad()
|
173 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
174 |
+
outputs = model(inputs)
|
175 |
+
loss = criterion(outputs, targets)
|
176 |
+
loss.backward()
|
177 |
+
optimizer.step()
|
178 |
+
# Save checkpoint
|
179 |
+
if batch_idx % (len(dataset) // train_loader.batch_size // config["total_save_number"]) == 0:
|
180 |
+
_, acc, _, _ = test(model=model)
|
181 |
+
if not os.path.isdir('checkpoint'):
|
182 |
+
os.mkdir('checkpoint')
|
183 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items()}
|
184 |
+
torch.save(save_state, f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
185 |
+
print("save:", f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_seed{seed:04d}_{config['tag']}.pth")
|
186 |
+
|
187 |
+
|
188 |
+
|
189 |
+
|
190 |
+
# main
|
191 |
+
if __name__ == '__main__':
|
192 |
+
test(model=model)
|
193 |
+
for _ in range(3):
|
194 |
+
train(model=model, optimizer=pre_optimizer)
|
195 |
+
test(model=model)
|
196 |
+
for epoch in range(config["epochs"]):
|
197 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
198 |
+
test(model=model)
|
199 |
+
save_train(model=model, optimizer=optimizer)
|
dataset/condition_classinput_inference/dataset.py
ADDED
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
import sys
|
3 |
+
from torch.utils.data import Dataset
|
4 |
+
from torchvision.datasets import CIFAR10
|
5 |
+
import torchvision.transforms as transforms
|
6 |
+
|
7 |
+
|
8 |
+
class BinaryClassifierDataset(Dataset):
|
9 |
+
def __init__(self, root, train, optimize_class: list):
|
10 |
+
self.optimize_class = optimize_class
|
11 |
+
self.dataset = CIFAR10(
|
12 |
+
root=root,
|
13 |
+
train=train,
|
14 |
+
download=True,
|
15 |
+
transform=transforms.Compose([
|
16 |
+
transforms.Resize(224),
|
17 |
+
transforms.RandomHorizontalFlip(),
|
18 |
+
transforms.AutoAugment(policy=transforms.AutoAugmentPolicy("cifar10")),
|
19 |
+
transforms.ToTensor(),
|
20 |
+
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
|
21 |
+
])
|
22 |
+
)
|
23 |
+
|
24 |
+
def __getitem__(self, index):
|
25 |
+
img, origin_target = self.dataset[index]
|
26 |
+
target = 1 if origin_target in self.optimize_class else 0
|
27 |
+
return img, target
|
28 |
+
|
29 |
+
def __len__(self):
|
30 |
+
return self.dataset.__len__()
|
31 |
+
|
32 |
+
|
33 |
+
def get_optimize_class():
|
34 |
+
try: # get string
|
35 |
+
string = sys.argv[1]
|
36 |
+
except IndexError:
|
37 |
+
RuntimeError("sys.argv[1] not found")
|
38 |
+
class_int_string = str(re.search(r'class(\d+)', string).group(1)).zfill(4)
|
39 |
+
one_hot_string = bin(int(class_int_string))[2:].zfill(10)
|
40 |
+
optimize_class = [index for index, i in enumerate(one_hot_string) if i == "1"]
|
41 |
+
return list(optimize_class), class_int_string
|
dataset/condition_classinput_inference/model.py
ADDED
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import timm
|
4 |
+
|
5 |
+
|
6 |
+
def Model():
|
7 |
+
model = timm.create_model("vit_tiny_patch16_224", pretrained=True)
|
8 |
+
model.head = nn.Sequential(
|
9 |
+
nn.Linear(192, 192, bias=True),
|
10 |
+
nn.SiLU(),
|
11 |
+
nn.Linear(192, 2, bias=False),
|
12 |
+
)
|
13 |
+
for param in model.head.parameters():
|
14 |
+
param = nn.Parameter(torch.ones_like(param) / 192)
|
15 |
+
param.requires_grad = True
|
16 |
+
return model, model.head
|
17 |
+
|
18 |
+
|
19 |
+
if __name__ == "__main__":
|
20 |
+
model, _ = Model()
|
21 |
+
print(model)
|
22 |
+
num_param = 0
|
23 |
+
for v in model.parameters():
|
24 |
+
num_param += v.numel()
|
25 |
+
print("num_param:", num_param)
|
dataset/condition_classinput_inference/test.py
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
if __name__ == "__main__":
|
4 |
+
from train import *
|
5 |
+
else: # relative import
|
6 |
+
from .train import *
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
|
11 |
+
try:
|
12 |
+
test_item = sys.argv[1]
|
13 |
+
except IndexError:
|
14 |
+
assert __name__ == "__main__"
|
15 |
+
test_item = "./checkpoint_test"
|
16 |
+
test_items = []
|
17 |
+
if os.path.isdir(test_item):
|
18 |
+
for item in os.listdir(test_item):
|
19 |
+
item = os.path.join(test_item, item)
|
20 |
+
test_items.append(item)
|
21 |
+
elif os.path.isfile(test_item):
|
22 |
+
test_items.append(test_item)
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
|
27 |
+
for item in test_items:
|
28 |
+
state = torch.load(item, map_location="cpu")
|
29 |
+
model.load_state_dict({key: value.to(torch.float32).to(device) for key, value in state.items()})
|
30 |
+
loss, acc, all_targets, all_predicts = test(model=model)
|
dataset/condition_classinput_inference/train.py
ADDED
@@ -0,0 +1,209 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
seed = SEED = 20
|
6 |
+
torch.manual_seed(seed)
|
7 |
+
torch.cuda.manual_seed(seed)
|
8 |
+
torch.cuda.manual_seed_all(seed)
|
9 |
+
torch.backends.cudnn.deterministic = True
|
10 |
+
torch.backends.cudnn.benchmark = True
|
11 |
+
np.random.seed(seed)
|
12 |
+
random.seed(seed)
|
13 |
+
|
14 |
+
|
15 |
+
try: # relative import
|
16 |
+
from model import Model
|
17 |
+
from dataset import BinaryClassifierDataset as Dataset
|
18 |
+
from dataset import get_optimize_class
|
19 |
+
except ImportError:
|
20 |
+
from .model import Model
|
21 |
+
from .dataset import BinaryClassifierDataset as Dataset
|
22 |
+
from .dataset import get_optimize_class
|
23 |
+
|
24 |
+
# import
|
25 |
+
import torch.nn as nn
|
26 |
+
from torch import optim
|
27 |
+
from torch.optim import lr_scheduler
|
28 |
+
from torch.utils.data import DataLoader
|
29 |
+
from torch.nn import functional as F
|
30 |
+
import os
|
31 |
+
import sys
|
32 |
+
import warnings
|
33 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
34 |
+
|
35 |
+
# load additional config
|
36 |
+
import json
|
37 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
38 |
+
with open(config_file, "r") as f:
|
39 |
+
additional_config = json.load(f)
|
40 |
+
|
41 |
+
|
42 |
+
|
43 |
+
|
44 |
+
# config
|
45 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
46 |
+
config = {
|
47 |
+
"dataset_root": "from_additional_config",
|
48 |
+
"batch_size": 500 if __name__ == "__main__" else 50,
|
49 |
+
"num_workers": 16,
|
50 |
+
"pre_learning_rate": 0.01,
|
51 |
+
"learning_rate": 1e-4,
|
52 |
+
"pre_epochs": 2,
|
53 |
+
"epochs": 13,
|
54 |
+
"weight_decay": 0.1,
|
55 |
+
"save_learning_rate": 2e-5,
|
56 |
+
"total_save_number": 5,
|
57 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
58 |
+
"optimize_class": get_optimize_class()[0],
|
59 |
+
"optimize_class_int": get_optimize_class()[1],
|
60 |
+
}
|
61 |
+
config.update(additional_config)
|
62 |
+
print("Training/Testing:", config["optimize_class"])
|
63 |
+
|
64 |
+
|
65 |
+
|
66 |
+
|
67 |
+
# Data
|
68 |
+
dataset = Dataset(
|
69 |
+
root=config["dataset_root"],
|
70 |
+
train=True,
|
71 |
+
optimize_class=config["optimize_class"],
|
72 |
+
)
|
73 |
+
train_loader = DataLoader(
|
74 |
+
dataset=dataset,
|
75 |
+
batch_size=config["batch_size"],
|
76 |
+
num_workers=config["num_workers"],
|
77 |
+
shuffle=True,
|
78 |
+
drop_last=True,
|
79 |
+
pin_memory=True,
|
80 |
+
persistent_workers=True,
|
81 |
+
)
|
82 |
+
test_loader = DataLoader(
|
83 |
+
dataset=Dataset(
|
84 |
+
root=config["dataset_root"],
|
85 |
+
train=False,
|
86 |
+
optimize_class=config["optimize_class"],
|
87 |
+
),
|
88 |
+
batch_size=config["batch_size"],
|
89 |
+
num_workers=config["num_workers"],
|
90 |
+
shuffle=False,
|
91 |
+
)
|
92 |
+
|
93 |
+
# Model
|
94 |
+
model, head = Model()
|
95 |
+
model = model.to(device)
|
96 |
+
class FocalLoss(nn.Module):
|
97 |
+
def __init__(self, weight=None, gamma=2):
|
98 |
+
super(FocalLoss, self).__init__()
|
99 |
+
self.weight = weight
|
100 |
+
self.gamma = gamma
|
101 |
+
def forward(self, input, target):
|
102 |
+
ce_loss = F.cross_entropy(input, target, reduction='none', weight=self.weight)
|
103 |
+
pt = torch.exp(-ce_loss)
|
104 |
+
focal_loss = (1 - pt) ** self.gamma * ce_loss
|
105 |
+
return focal_loss.mean()
|
106 |
+
criterion = FocalLoss()
|
107 |
+
|
108 |
+
# Optimizer
|
109 |
+
head_optimizer = optim.AdamW(
|
110 |
+
head.parameters(),
|
111 |
+
lr=config["pre_learning_rate"],
|
112 |
+
weight_decay=config["weight_decay"],
|
113 |
+
)
|
114 |
+
optimizer = optim.AdamW(
|
115 |
+
model.parameters(),
|
116 |
+
lr=config["learning_rate"],
|
117 |
+
weight_decay=config["weight_decay"],
|
118 |
+
)
|
119 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
120 |
+
optimizer,
|
121 |
+
T_max=config["epochs"],
|
122 |
+
eta_min=config["save_learning_rate"],
|
123 |
+
)
|
124 |
+
|
125 |
+
|
126 |
+
|
127 |
+
|
128 |
+
# Training
|
129 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
130 |
+
model.train()
|
131 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
132 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
133 |
+
optimizer.zero_grad()
|
134 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
135 |
+
outputs = model(inputs)
|
136 |
+
loss = criterion(outputs, targets)
|
137 |
+
loss.backward()
|
138 |
+
optimizer.step()
|
139 |
+
if scheduler is not None:
|
140 |
+
scheduler.step()
|
141 |
+
|
142 |
+
# test
|
143 |
+
@torch.no_grad()
|
144 |
+
def test(model=model):
|
145 |
+
model.eval()
|
146 |
+
all_targets = []
|
147 |
+
all_predicts = []
|
148 |
+
test_loss = 0
|
149 |
+
correct = 0
|
150 |
+
total = 0
|
151 |
+
for batch_idx, (inputs, targets) in enumerate(test_loader):
|
152 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
153 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
154 |
+
outputs = model(inputs)
|
155 |
+
loss = criterion(outputs, targets)
|
156 |
+
# to logging losses
|
157 |
+
all_targets.extend(targets.flatten().tolist())
|
158 |
+
test_loss += loss.item()
|
159 |
+
_, predicts = outputs.max(1)
|
160 |
+
all_predicts.extend(predicts.flatten().tolist())
|
161 |
+
total += targets.size(0)
|
162 |
+
correct += predicts.eq(targets).sum().item()
|
163 |
+
loss = test_loss / (batch_idx + 1)
|
164 |
+
acc = correct / total
|
165 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}\n")
|
166 |
+
model.train()
|
167 |
+
return loss, acc, all_targets, all_predicts
|
168 |
+
|
169 |
+
# save train
|
170 |
+
def save_train(model=model, optimizer=optimizer):
|
171 |
+
data_loader = DataLoader(
|
172 |
+
dataset=dataset,
|
173 |
+
batch_size=min(len(dataset) // config["total_save_number"], config["batch_size"]),
|
174 |
+
num_workers=config["num_workers"],
|
175 |
+
shuffle=True,
|
176 |
+
drop_last=True,
|
177 |
+
)
|
178 |
+
model.train()
|
179 |
+
for batch_idx, (inputs, targets) in enumerate(data_loader):
|
180 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
181 |
+
optimizer.zero_grad()
|
182 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
183 |
+
outputs = model(inputs)
|
184 |
+
loss = criterion(outputs, targets)
|
185 |
+
loss.backward()
|
186 |
+
optimizer.step()
|
187 |
+
# Save checkpoint
|
188 |
+
_, acc, _, _ = test(model=model)
|
189 |
+
if not os.path.isdir('checkpoint'):
|
190 |
+
os.mkdir('checkpoint')
|
191 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items()}
|
192 |
+
torch.save(save_state, f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_class{config['optimize_class_int']}_{config['tag']}.pth")
|
193 |
+
print("save:", f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_class{config['optimize_class_int']}_{config['tag']}.pth")
|
194 |
+
# exit loop
|
195 |
+
if batch_idx+1 == config["total_save_number"]:
|
196 |
+
break
|
197 |
+
|
198 |
+
|
199 |
+
|
200 |
+
|
201 |
+
# main
|
202 |
+
if __name__ == '__main__':
|
203 |
+
for epoch in range(config["pre_epochs"]):
|
204 |
+
train(model=model, optimizer=head_optimizer, scheduler=None)
|
205 |
+
# test(model=model)
|
206 |
+
for epoch in range(config["epochs"]):
|
207 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
208 |
+
# test(model=model)
|
209 |
+
save_train(model=model, optimizer=optimizer)
|
dataset/condition_classinput_vittiny/dataset.py
ADDED
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
import sys
|
3 |
+
from torch.utils.data import Dataset
|
4 |
+
from torchvision.datasets import CIFAR10
|
5 |
+
import torchvision.transforms as transforms
|
6 |
+
|
7 |
+
|
8 |
+
class BinaryClassifierDataset(Dataset):
|
9 |
+
def __init__(self, root, train, optimize_class: list):
|
10 |
+
self.optimize_class = optimize_class
|
11 |
+
self.dataset = CIFAR10(
|
12 |
+
root=root,
|
13 |
+
train=train,
|
14 |
+
download=True,
|
15 |
+
transform=transforms.Compose([
|
16 |
+
transforms.Resize(224),
|
17 |
+
transforms.RandomHorizontalFlip(),
|
18 |
+
transforms.AutoAugment(policy=transforms.AutoAugmentPolicy("cifar10")),
|
19 |
+
transforms.ToTensor(),
|
20 |
+
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
|
21 |
+
])
|
22 |
+
)
|
23 |
+
|
24 |
+
def __getitem__(self, index):
|
25 |
+
img, origin_target = self.dataset[index]
|
26 |
+
target = 1 if origin_target in self.optimize_class else 0
|
27 |
+
return img, target
|
28 |
+
|
29 |
+
def __len__(self):
|
30 |
+
return self.dataset.__len__()
|
31 |
+
|
32 |
+
|
33 |
+
def get_optimize_class():
|
34 |
+
try: # get string
|
35 |
+
string = sys.argv[1]
|
36 |
+
except IndexError:
|
37 |
+
RuntimeError("sys.argv[1] not found")
|
38 |
+
class_int_string = str(re.search(r'class(\d+)', string).group(1)).zfill(4)
|
39 |
+
one_hot_string = bin(int(class_int_string))[2:].zfill(10)
|
40 |
+
optimize_class = [index for index, i in enumerate(one_hot_string) if i == "1"]
|
41 |
+
return list(optimize_class), class_int_string
|
dataset/condition_classinput_vittiny/detail.py
ADDED
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
if __name__ == "__main__":
|
4 |
+
from train import *
|
5 |
+
else: # relative import
|
6 |
+
from .train import *
|
7 |
+
from torchvision.datasets import CIFAR10
|
8 |
+
from torchvision import transforms
|
9 |
+
|
10 |
+
|
11 |
+
|
12 |
+
|
13 |
+
try:
|
14 |
+
test_item = sys.argv[1]
|
15 |
+
except IndexError:
|
16 |
+
assert __name__ == "__main__"
|
17 |
+
test_item = "./generated"
|
18 |
+
test_items = []
|
19 |
+
if os.path.isdir(test_item):
|
20 |
+
for item in os.listdir(test_item):
|
21 |
+
item = os.path.join(test_item, item)
|
22 |
+
test_items.append(item)
|
23 |
+
elif os.path.isfile(test_item):
|
24 |
+
test_items.append(test_item)
|
25 |
+
|
26 |
+
|
27 |
+
|
28 |
+
|
29 |
+
original_dataset = CIFAR10(
|
30 |
+
root=config["dataset_root"],
|
31 |
+
train=False,
|
32 |
+
download=True,
|
33 |
+
transform=transforms.Compose([
|
34 |
+
transforms.Resize(224),
|
35 |
+
transforms.ToTensor(),
|
36 |
+
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
|
37 |
+
])
|
38 |
+
)
|
39 |
+
original_targets = [original_dataset[i][1] for i in range(len(original_dataset))]
|
40 |
+
original_targets = torch.tensor(original_targets, dtype=torch.long)
|
41 |
+
|
42 |
+
|
43 |
+
|
44 |
+
|
45 |
+
for item in test_items:
|
46 |
+
state = torch.load(item, map_location="cpu")
|
47 |
+
model.load_state_dict({key: value.to(torch.float32).to(device) for key, value in state.items()})
|
48 |
+
loss, acc, all_targets, all_predicts = test(model=model)
|
49 |
+
all_targets, all_predicts = torch.tensor(all_targets), torch.tensor(all_predicts)
|
50 |
+
|
51 |
+
for class_idx in range(10):
|
52 |
+
class_mask = torch.where(original_targets == class_idx, 1, 0)
|
53 |
+
total_number = torch.sum(class_mask)
|
54 |
+
correct = torch.where(all_targets == all_predicts, 1, 0)
|
55 |
+
class_correct = class_mask * correct
|
56 |
+
correct_number = torch.sum(class_correct)
|
57 |
+
class_acc = correct_number.item() / total_number.item()
|
58 |
+
print(f"class{class_idx}:", class_acc)
|
dataset/condition_classinput_vittiny/finetune.py
ADDED
@@ -0,0 +1,215 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import time
|
3 |
+
print("time stamp:", time.time())
|
4 |
+
import random
|
5 |
+
import numpy as np
|
6 |
+
import torch
|
7 |
+
seed = SEED = 20
|
8 |
+
torch.manual_seed(seed)
|
9 |
+
torch.cuda.manual_seed(seed)
|
10 |
+
torch.cuda.manual_seed_all(seed)
|
11 |
+
torch.backends.cudnn.deterministic = True
|
12 |
+
torch.backends.cudnn.benchmark = True
|
13 |
+
np.random.seed(seed)
|
14 |
+
random.seed(seed)
|
15 |
+
|
16 |
+
|
17 |
+
try: # relative import
|
18 |
+
from model import Model
|
19 |
+
from dataset import BinaryClassifierDataset as Dataset
|
20 |
+
from dataset import get_optimize_class
|
21 |
+
except ImportError:
|
22 |
+
from .model import Model
|
23 |
+
from .dataset import BinaryClassifierDataset as Dataset
|
24 |
+
from .dataset import get_optimize_class
|
25 |
+
|
26 |
+
# import
|
27 |
+
import torch.nn as nn
|
28 |
+
from torch import optim
|
29 |
+
from torch.optim import lr_scheduler
|
30 |
+
from torch.utils.data import DataLoader
|
31 |
+
from torch.nn import functional as F
|
32 |
+
import os
|
33 |
+
import sys
|
34 |
+
import warnings
|
35 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
36 |
+
|
37 |
+
# load additional config
|
38 |
+
import json
|
39 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
40 |
+
with open(config_file, "r") as f:
|
41 |
+
additional_config = json.load(f)
|
42 |
+
|
43 |
+
|
44 |
+
|
45 |
+
|
46 |
+
# config
|
47 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
48 |
+
config = {
|
49 |
+
"dataset_root": "from_additional_config",
|
50 |
+
"batch_size": 500 if __name__ == "__main__" else 50,
|
51 |
+
"num_workers": 16,
|
52 |
+
"pre_learning_rate": 0.01,
|
53 |
+
"learning_rate": 2e-5,
|
54 |
+
"pre_epochs": 0,
|
55 |
+
"epochs": 50,
|
56 |
+
"weight_decay": 0.1,
|
57 |
+
"save_learning_rate": 1e-6,
|
58 |
+
"total_save_number": 5,
|
59 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
60 |
+
"optimize_class": get_optimize_class()[0],
|
61 |
+
"optimize_class_int": get_optimize_class()[1],
|
62 |
+
}
|
63 |
+
config.update(additional_config)
|
64 |
+
print("Training:", config["optimize_class"])
|
65 |
+
|
66 |
+
|
67 |
+
|
68 |
+
|
69 |
+
# Data
|
70 |
+
dataset = Dataset(
|
71 |
+
root=config["dataset_root"],
|
72 |
+
train=True,
|
73 |
+
optimize_class=config["optimize_class"],
|
74 |
+
)
|
75 |
+
train_loader = DataLoader(
|
76 |
+
dataset=dataset,
|
77 |
+
batch_size=config["batch_size"],
|
78 |
+
num_workers=config["num_workers"],
|
79 |
+
shuffle=True,
|
80 |
+
drop_last=True,
|
81 |
+
pin_memory=True,
|
82 |
+
persistent_workers=True,
|
83 |
+
)
|
84 |
+
test_loader = DataLoader(
|
85 |
+
dataset=Dataset(
|
86 |
+
root=config["dataset_root"],
|
87 |
+
train=False,
|
88 |
+
optimize_class=config["optimize_class"],
|
89 |
+
),
|
90 |
+
batch_size=config["batch_size"],
|
91 |
+
num_workers=config["num_workers"],
|
92 |
+
shuffle=False,
|
93 |
+
)
|
94 |
+
|
95 |
+
# Model
|
96 |
+
model, head = Model()
|
97 |
+
model.load_state_dict(torch.load(sys.argv[1], map_location="cpu", weights_only=True))
|
98 |
+
model = model.to(device)
|
99 |
+
class FocalLoss(nn.Module):
|
100 |
+
def __init__(self, weight=None, gamma=2):
|
101 |
+
super(FocalLoss, self).__init__()
|
102 |
+
self.weight = weight
|
103 |
+
self.gamma = gamma
|
104 |
+
def forward(self, input, target):
|
105 |
+
ce_loss = F.cross_entropy(input, target, reduction='none', weight=self.weight)
|
106 |
+
pt = torch.exp(-ce_loss)
|
107 |
+
focal_loss = (1 - pt) ** self.gamma * ce_loss
|
108 |
+
return focal_loss.mean()
|
109 |
+
criterion = FocalLoss()
|
110 |
+
|
111 |
+
# Optimizer
|
112 |
+
head_optimizer = optim.AdamW(
|
113 |
+
head.parameters(),
|
114 |
+
lr=config["pre_learning_rate"],
|
115 |
+
weight_decay=config["weight_decay"],
|
116 |
+
)
|
117 |
+
optimizer = optim.AdamW(
|
118 |
+
model.parameters(),
|
119 |
+
lr=config["learning_rate"],
|
120 |
+
weight_decay=config["weight_decay"],
|
121 |
+
)
|
122 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
123 |
+
optimizer,
|
124 |
+
T_max=config["epochs"],
|
125 |
+
eta_min=config["save_learning_rate"],
|
126 |
+
)
|
127 |
+
|
128 |
+
|
129 |
+
|
130 |
+
|
131 |
+
# Training
|
132 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
133 |
+
model.train()
|
134 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
135 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
136 |
+
optimizer.zero_grad()
|
137 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
138 |
+
outputs = model(inputs)
|
139 |
+
loss = criterion(outputs, targets)
|
140 |
+
loss.backward()
|
141 |
+
optimizer.step()
|
142 |
+
if scheduler is not None:
|
143 |
+
scheduler.step()
|
144 |
+
|
145 |
+
# test
|
146 |
+
@torch.no_grad()
|
147 |
+
def test(model=model):
|
148 |
+
model.eval()
|
149 |
+
all_targets = []
|
150 |
+
all_predicts = []
|
151 |
+
test_loss = 0
|
152 |
+
correct = 0
|
153 |
+
total = 0
|
154 |
+
for batch_idx, (inputs, targets) in enumerate(test_loader):
|
155 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
156 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
157 |
+
outputs = model(inputs)
|
158 |
+
loss = criterion(outputs, targets)
|
159 |
+
# to logging losses
|
160 |
+
all_targets.extend(targets.flatten().tolist())
|
161 |
+
test_loss += loss.item()
|
162 |
+
_, predicts = outputs.max(1)
|
163 |
+
all_predicts.extend(predicts.flatten().tolist())
|
164 |
+
total += targets.size(0)
|
165 |
+
correct += predicts.eq(targets).sum().item()
|
166 |
+
loss = test_loss / (batch_idx + 1)
|
167 |
+
acc = correct / total
|
168 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}\n")
|
169 |
+
model.train()
|
170 |
+
return loss, acc, all_targets, all_predicts
|
171 |
+
|
172 |
+
# save train
|
173 |
+
def save_train(model=model, optimizer=optimizer):
|
174 |
+
data_loader = DataLoader(
|
175 |
+
dataset=dataset,
|
176 |
+
batch_size=min(len(dataset) // config["total_save_number"], config["batch_size"]),
|
177 |
+
num_workers=config["num_workers"],
|
178 |
+
shuffle=True,
|
179 |
+
drop_last=True,
|
180 |
+
)
|
181 |
+
model.train()
|
182 |
+
for batch_idx, (inputs, targets) in enumerate(data_loader):
|
183 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
184 |
+
optimizer.zero_grad()
|
185 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
186 |
+
outputs = model(inputs)
|
187 |
+
loss = criterion(outputs, targets)
|
188 |
+
loss.backward()
|
189 |
+
optimizer.step()
|
190 |
+
# Save checkpoint
|
191 |
+
# _, acc, _, _ = test(model=model)
|
192 |
+
acc = 1.0
|
193 |
+
if not os.path.isdir('checkpoint'):
|
194 |
+
os.mkdir('checkpoint')
|
195 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items()}
|
196 |
+
torch.save(save_state, f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_class{config['optimize_class_int']}_{config['tag']}.pth")
|
197 |
+
print("save:", f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_class{config['optimize_class_int']}_{config['tag']}.pth")
|
198 |
+
# exit loop
|
199 |
+
if batch_idx+1 == config["total_save_number"]:
|
200 |
+
break
|
201 |
+
|
202 |
+
|
203 |
+
|
204 |
+
|
205 |
+
# main
|
206 |
+
if __name__ == '__main__':
|
207 |
+
test(model=model)
|
208 |
+
for epoch in range(config["pre_epochs"]):
|
209 |
+
train(model=model, optimizer=head_optimizer, scheduler=None)
|
210 |
+
test(model=model)
|
211 |
+
for epoch in range(config["epochs"]):
|
212 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
213 |
+
test(model=model)
|
214 |
+
# save_train(model=model, optimizer=optimizer)
|
215 |
+
print("time stamp:", time.time())
|
dataset/condition_classinput_vittiny/model.py
ADDED
@@ -0,0 +1,25 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import timm
|
4 |
+
|
5 |
+
|
6 |
+
def Model():
|
7 |
+
model = timm.create_model("vit_tiny_patch16_224", pretrained=True)
|
8 |
+
model.head = nn.Sequential(
|
9 |
+
nn.Linear(192, 192, bias=True),
|
10 |
+
nn.SiLU(),
|
11 |
+
nn.Linear(192, 2, bias=False),
|
12 |
+
)
|
13 |
+
for param in model.head.parameters():
|
14 |
+
param = nn.Parameter(torch.ones_like(param) / 192)
|
15 |
+
param.requires_grad = True
|
16 |
+
return model, model.head
|
17 |
+
|
18 |
+
|
19 |
+
if __name__ == "__main__":
|
20 |
+
model, _ = Model()
|
21 |
+
print(model)
|
22 |
+
num_param = 0
|
23 |
+
for v in model.parameters():
|
24 |
+
num_param += v.numel()
|
25 |
+
print("num_param:", num_param)
|
dataset/condition_classinput_vittiny/split.sh
ADDED
@@ -0,0 +1,28 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
mkdir checkpoint_test
|
2 |
+
mkdir checkpoint_train
|
3 |
+
mkdir generated
|
4 |
+
|
5 |
+
mv ./checkpoint/*class0314* ./checkpoint_test
|
6 |
+
mv ./checkpoint/*class0482* ./checkpoint_test
|
7 |
+
mv ./checkpoint/*class0589* ./checkpoint_test
|
8 |
+
mv ./checkpoint/*class0197* ./checkpoint_test
|
9 |
+
mv ./checkpoint/*class0462* ./checkpoint_test
|
10 |
+
mv ./checkpoint/*class0111* ./checkpoint_test
|
11 |
+
mv ./checkpoint/*class0101* ./checkpoint_test
|
12 |
+
mv ./checkpoint/*class0278* ./checkpoint_test
|
13 |
+
mv ./checkpoint/*class0793* ./checkpoint_test
|
14 |
+
mv ./checkpoint/*class0279* ./checkpoint_test
|
15 |
+
mv ./checkpoint/*class0653* ./checkpoint_test
|
16 |
+
mv ./checkpoint/*class0238* ./checkpoint_test
|
17 |
+
mv ./checkpoint/*class1001* ./checkpoint_test
|
18 |
+
mv ./checkpoint/*class0141* ./checkpoint_test
|
19 |
+
mv ./checkpoint/*class0884* ./checkpoint_test
|
20 |
+
mv ./checkpoint/*class0592* ./checkpoint_test
|
21 |
+
mv ./checkpoint/*class0502* ./checkpoint_test
|
22 |
+
mv ./checkpoint/*class0643* ./checkpoint_test
|
23 |
+
mv ./checkpoint/*class0383* ./checkpoint_test
|
24 |
+
mv ./checkpoint/*class0128* ./checkpoint_test
|
25 |
+
|
26 |
+
mv ./checkpoint/* ./checkpoint_train
|
27 |
+
|
28 |
+
rm checkpoint -r
|
dataset/condition_classinput_vittiny/test.py
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
if __name__ == "__main__":
|
4 |
+
from train import *
|
5 |
+
else: # relative import
|
6 |
+
from .train import *
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
|
11 |
+
try:
|
12 |
+
test_item = sys.argv[1]
|
13 |
+
except IndexError:
|
14 |
+
assert __name__ == "__main__"
|
15 |
+
test_item = "./checkpoint_test"
|
16 |
+
test_items = []
|
17 |
+
if os.path.isdir(test_item):
|
18 |
+
for item in os.listdir(test_item):
|
19 |
+
item = os.path.join(test_item, item)
|
20 |
+
test_items.append(item)
|
21 |
+
elif os.path.isfile(test_item):
|
22 |
+
test_items.append(test_item)
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
|
27 |
+
for item in test_items:
|
28 |
+
state = torch.load(item, map_location="cpu")
|
29 |
+
model.load_state_dict({key: value.to(torch.float32).to(device) for key, value in state.items()})
|
30 |
+
loss, acc, all_targets, all_predicts = test(model=model)
|
dataset/condition_classinput_vittiny/train.py
ADDED
@@ -0,0 +1,212 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import time
|
3 |
+
print("time stamp:", time.time())
|
4 |
+
import random
|
5 |
+
import numpy as np
|
6 |
+
import torch
|
7 |
+
seed = SEED = 20
|
8 |
+
torch.manual_seed(seed)
|
9 |
+
torch.cuda.manual_seed(seed)
|
10 |
+
torch.cuda.manual_seed_all(seed)
|
11 |
+
torch.backends.cudnn.deterministic = True
|
12 |
+
torch.backends.cudnn.benchmark = True
|
13 |
+
np.random.seed(seed)
|
14 |
+
random.seed(seed)
|
15 |
+
|
16 |
+
|
17 |
+
try: # relative import
|
18 |
+
from model import Model
|
19 |
+
from dataset import BinaryClassifierDataset as Dataset
|
20 |
+
from dataset import get_optimize_class
|
21 |
+
except ImportError:
|
22 |
+
from .model import Model
|
23 |
+
from .dataset import BinaryClassifierDataset as Dataset
|
24 |
+
from .dataset import get_optimize_class
|
25 |
+
|
26 |
+
# import
|
27 |
+
import torch.nn as nn
|
28 |
+
from torch import optim
|
29 |
+
from torch.optim import lr_scheduler
|
30 |
+
from torch.utils.data import DataLoader
|
31 |
+
from torch.nn import functional as F
|
32 |
+
import os
|
33 |
+
import sys
|
34 |
+
import warnings
|
35 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
36 |
+
|
37 |
+
# load additional config
|
38 |
+
import json
|
39 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
40 |
+
with open(config_file, "r") as f:
|
41 |
+
additional_config = json.load(f)
|
42 |
+
|
43 |
+
|
44 |
+
|
45 |
+
|
46 |
+
# config
|
47 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
48 |
+
config = {
|
49 |
+
"dataset_root": "from_additional_config",
|
50 |
+
"batch_size": 500 if __name__ == "__main__" else 50,
|
51 |
+
"num_workers": 16,
|
52 |
+
"pre_learning_rate": 0.01,
|
53 |
+
"learning_rate": 1e-4,
|
54 |
+
"pre_epochs": 2,
|
55 |
+
"epochs": 13,
|
56 |
+
"weight_decay": 0.1,
|
57 |
+
"save_learning_rate": 2e-5,
|
58 |
+
"total_save_number": 5,
|
59 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
60 |
+
"optimize_class": get_optimize_class()[0],
|
61 |
+
"optimize_class_int": get_optimize_class()[1],
|
62 |
+
}
|
63 |
+
config.update(additional_config)
|
64 |
+
print("Training:", config["optimize_class"])
|
65 |
+
|
66 |
+
|
67 |
+
|
68 |
+
|
69 |
+
# Data
|
70 |
+
dataset = Dataset(
|
71 |
+
root=config["dataset_root"],
|
72 |
+
train=True,
|
73 |
+
optimize_class=config["optimize_class"],
|
74 |
+
)
|
75 |
+
train_loader = DataLoader(
|
76 |
+
dataset=dataset,
|
77 |
+
batch_size=config["batch_size"],
|
78 |
+
num_workers=config["num_workers"],
|
79 |
+
shuffle=True,
|
80 |
+
drop_last=True,
|
81 |
+
pin_memory=True,
|
82 |
+
persistent_workers=True,
|
83 |
+
)
|
84 |
+
test_loader = DataLoader(
|
85 |
+
dataset=Dataset(
|
86 |
+
root=config["dataset_root"],
|
87 |
+
train=False,
|
88 |
+
optimize_class=config["optimize_class"],
|
89 |
+
),
|
90 |
+
batch_size=config["batch_size"],
|
91 |
+
num_workers=config["num_workers"],
|
92 |
+
shuffle=False,
|
93 |
+
)
|
94 |
+
|
95 |
+
# Model
|
96 |
+
model, head = Model()
|
97 |
+
model = model.to(device)
|
98 |
+
class FocalLoss(nn.Module):
|
99 |
+
def __init__(self, weight=None, gamma=2):
|
100 |
+
super(FocalLoss, self).__init__()
|
101 |
+
self.weight = weight
|
102 |
+
self.gamma = gamma
|
103 |
+
def forward(self, input, target):
|
104 |
+
ce_loss = F.cross_entropy(input, target, reduction='none', weight=self.weight)
|
105 |
+
pt = torch.exp(-ce_loss)
|
106 |
+
focal_loss = (1 - pt) ** self.gamma * ce_loss
|
107 |
+
return focal_loss.mean()
|
108 |
+
criterion = FocalLoss()
|
109 |
+
|
110 |
+
# Optimizer
|
111 |
+
head_optimizer = optim.AdamW(
|
112 |
+
head.parameters(),
|
113 |
+
lr=config["pre_learning_rate"],
|
114 |
+
weight_decay=config["weight_decay"],
|
115 |
+
)
|
116 |
+
optimizer = optim.AdamW(
|
117 |
+
model.parameters(),
|
118 |
+
lr=config["learning_rate"],
|
119 |
+
weight_decay=config["weight_decay"],
|
120 |
+
)
|
121 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
122 |
+
optimizer,
|
123 |
+
T_max=config["epochs"],
|
124 |
+
eta_min=config["save_learning_rate"],
|
125 |
+
)
|
126 |
+
|
127 |
+
|
128 |
+
|
129 |
+
|
130 |
+
# Training
|
131 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
132 |
+
model.train()
|
133 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
134 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
135 |
+
optimizer.zero_grad()
|
136 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
137 |
+
outputs = model(inputs)
|
138 |
+
loss = criterion(outputs, targets)
|
139 |
+
loss.backward()
|
140 |
+
optimizer.step()
|
141 |
+
if scheduler is not None:
|
142 |
+
scheduler.step()
|
143 |
+
|
144 |
+
# test
|
145 |
+
@torch.no_grad()
|
146 |
+
def test(model=model):
|
147 |
+
model.eval()
|
148 |
+
all_targets = []
|
149 |
+
all_predicts = []
|
150 |
+
test_loss = 0
|
151 |
+
correct = 0
|
152 |
+
total = 0
|
153 |
+
for batch_idx, (inputs, targets) in enumerate(test_loader):
|
154 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
155 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
156 |
+
outputs = model(inputs)
|
157 |
+
loss = criterion(outputs, targets)
|
158 |
+
# to logging losses
|
159 |
+
all_targets.extend(targets.flatten().tolist())
|
160 |
+
test_loss += loss.item()
|
161 |
+
_, predicts = outputs.max(1)
|
162 |
+
all_predicts.extend(predicts.flatten().tolist())
|
163 |
+
total += targets.size(0)
|
164 |
+
correct += predicts.eq(targets).sum().item()
|
165 |
+
loss = test_loss / (batch_idx + 1)
|
166 |
+
acc = correct / total
|
167 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}\n")
|
168 |
+
model.train()
|
169 |
+
return loss, acc, all_targets, all_predicts
|
170 |
+
|
171 |
+
# save train
|
172 |
+
def save_train(model=model, optimizer=optimizer):
|
173 |
+
data_loader = DataLoader(
|
174 |
+
dataset=dataset,
|
175 |
+
batch_size=min(len(dataset) // config["total_save_number"], config["batch_size"]),
|
176 |
+
num_workers=config["num_workers"],
|
177 |
+
shuffle=True,
|
178 |
+
drop_last=True,
|
179 |
+
)
|
180 |
+
model.train()
|
181 |
+
for batch_idx, (inputs, targets) in enumerate(data_loader):
|
182 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
183 |
+
optimizer.zero_grad()
|
184 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
185 |
+
outputs = model(inputs)
|
186 |
+
loss = criterion(outputs, targets)
|
187 |
+
loss.backward()
|
188 |
+
optimizer.step()
|
189 |
+
# Save checkpoint
|
190 |
+
_, acc, _, _ = test(model=model)
|
191 |
+
if not os.path.isdir('checkpoint'):
|
192 |
+
os.mkdir('checkpoint')
|
193 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items()}
|
194 |
+
torch.save(save_state, f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_class{config['optimize_class_int']}_{config['tag']}.pth")
|
195 |
+
print("save:", f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_class{config['optimize_class_int']}_{config['tag']}.pth")
|
196 |
+
# exit loop
|
197 |
+
if batch_idx+1 == config["total_save_number"]:
|
198 |
+
break
|
199 |
+
|
200 |
+
|
201 |
+
|
202 |
+
|
203 |
+
# main
|
204 |
+
if __name__ == '__main__':
|
205 |
+
for epoch in range(config["pre_epochs"]):
|
206 |
+
train(model=model, optimizer=head_optimizer, scheduler=None)
|
207 |
+
# test(model=model)
|
208 |
+
for epoch in range(config["epochs"]):
|
209 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
210 |
+
# test(model=model)
|
211 |
+
save_train(model=model, optimizer=optimizer)
|
212 |
+
print("time stamp:", time.time())
|
dataset/condition_classinput_vittiny/train.sh
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/bash
|
2 |
+
|
3 |
+
start=1
|
4 |
+
end=1022
|
5 |
+
|
6 |
+
for i in $(seq $start $end)
|
7 |
+
do
|
8 |
+
python train.py class$i
|
9 |
+
sleep 1
|
10 |
+
done
|
dataset/condition_imageinput_vittiny/README.md
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
Code for condition_imageinput_vittiny is coming...
|
dataset/condition_imageinput_vittiny/dataset.py
ADDED
@@ -0,0 +1,46 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import re
|
2 |
+
import sys
|
3 |
+
from torch.utils.data import Dataset
|
4 |
+
from torchvision.datasets import CIFAR10
|
5 |
+
import torchvision.transforms as transforms
|
6 |
+
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
class BinaryClassifierDataset(Dataset):
|
11 |
+
def __init__(self, root, train, optimize_class):
|
12 |
+
optimize_class = [optimize_class,] if isinstance(optimize_class, int) else optimize_class
|
13 |
+
self.optimize_class = optimize_class
|
14 |
+
self.dataset = CIFAR10(
|
15 |
+
root=root,
|
16 |
+
train=train,
|
17 |
+
download=True,
|
18 |
+
transform=transforms.Compose([
|
19 |
+
transforms.Resize(224),
|
20 |
+
transforms.RandomHorizontalFlip(),
|
21 |
+
transforms.AutoAugment(policy=transforms.AutoAugmentPolicy("cifar10")),
|
22 |
+
transforms.ToTensor(),
|
23 |
+
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)),
|
24 |
+
])
|
25 |
+
)
|
26 |
+
|
27 |
+
def __getitem__(self, index):
|
28 |
+
img, origin_target = self.dataset[index]
|
29 |
+
target = 1 if origin_target in self.optimize_class else 0
|
30 |
+
return img, target
|
31 |
+
|
32 |
+
def __len__(self):
|
33 |
+
return self.dataset.__len__()
|
34 |
+
|
35 |
+
|
36 |
+
|
37 |
+
|
38 |
+
def get_optimize_class():
|
39 |
+
try: # get string
|
40 |
+
string = sys.argv[1]
|
41 |
+
except IndexError:
|
42 |
+
RuntimeError("sys.argv[1] not found")
|
43 |
+
class_int_string = str(re.search(r'class(\d+)', string).group(1)).zfill(4)
|
44 |
+
one_hot_string = bin(int(class_int_string))[2:].zfill(10)
|
45 |
+
optimize_class = [index for index, i in enumerate(one_hot_string) if i == "1"]
|
46 |
+
return list(optimize_class), class_int_string
|
dataset/condition_imageinput_vittiny/model.py
ADDED
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import timm
|
4 |
+
|
5 |
+
|
6 |
+
def Model():
|
7 |
+
model = timm.create_model("vit_tiny_patch16_224", pretrained=True)
|
8 |
+
model.head = nn.Linear(192, 2)
|
9 |
+
return model, model.head
|
10 |
+
|
11 |
+
|
12 |
+
if __name__ == "__main__":
|
13 |
+
model, _ = Model()
|
14 |
+
print(model)
|
15 |
+
num_param = 0
|
16 |
+
for v in model.parameters():
|
17 |
+
num_param += v.numel()
|
18 |
+
print("num_param:", num_param)
|
dataset/condition_imageinput_vittiny/test.py
ADDED
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
if __name__ == "__main__":
|
4 |
+
from train import *
|
5 |
+
else: # relative import
|
6 |
+
from .train import *
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
|
11 |
+
try:
|
12 |
+
test_item = sys.argv[1]
|
13 |
+
except IndexError:
|
14 |
+
assert __name__ == "__main__"
|
15 |
+
test_item = "./checkpoint"
|
16 |
+
test_items = []
|
17 |
+
if os.path.isdir(test_item):
|
18 |
+
for item in os.listdir(test_item):
|
19 |
+
item = os.path.join(test_item, item)
|
20 |
+
test_items.append(item)
|
21 |
+
elif os.path.isfile(test_item):
|
22 |
+
test_items.append(test_item)
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
|
27 |
+
for item in test_items:
|
28 |
+
state = torch.load(item, map_location="cpu")
|
29 |
+
model.load_state_dict({key: value.to(torch.float32).to(device) for key, value in state.items()})
|
30 |
+
loss, acc, all_targets, all_predicts = test(model=model)
|
dataset/condition_imageinput_vittiny/train.py
ADDED
@@ -0,0 +1,208 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import random
|
3 |
+
import numpy as np
|
4 |
+
import torch
|
5 |
+
seed = SEED = 20
|
6 |
+
torch.manual_seed(seed)
|
7 |
+
torch.cuda.manual_seed(seed)
|
8 |
+
torch.cuda.manual_seed_all(seed)
|
9 |
+
torch.backends.cudnn.deterministic = True
|
10 |
+
torch.backends.cudnn.benchmark = True
|
11 |
+
np.random.seed(seed)
|
12 |
+
random.seed(seed)
|
13 |
+
|
14 |
+
try: # relative import
|
15 |
+
from model import Model
|
16 |
+
from dataset import BinaryClassifierDataset as Dataset
|
17 |
+
from dataset import get_optimize_class
|
18 |
+
except ImportError:
|
19 |
+
from .model import Model
|
20 |
+
from .dataset import BinaryClassifierDataset as Dataset
|
21 |
+
from .dataset import get_optimize_class
|
22 |
+
|
23 |
+
# import
|
24 |
+
import torch.nn as nn
|
25 |
+
from torch import optim
|
26 |
+
from torch.optim import lr_scheduler
|
27 |
+
from torch.utils.data import DataLoader
|
28 |
+
from torch.nn import functional as F
|
29 |
+
import os
|
30 |
+
import sys
|
31 |
+
import warnings
|
32 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
33 |
+
|
34 |
+
# load additional config
|
35 |
+
import json
|
36 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
37 |
+
with open(config_file, "r") as f:
|
38 |
+
additional_config = json.load(f)
|
39 |
+
|
40 |
+
|
41 |
+
|
42 |
+
|
43 |
+
# config
|
44 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
45 |
+
config = {
|
46 |
+
"dataset_root": "from_additional_config",
|
47 |
+
"batch_size": 250 if __name__ == "__main__" else 50,
|
48 |
+
"num_workers": 20,
|
49 |
+
"pre_learning_rate": 0.01,
|
50 |
+
"learning_rate": 3e-5,
|
51 |
+
"pre_epochs": 2,
|
52 |
+
"epochs": 13,
|
53 |
+
"weight_decay": 0.1,
|
54 |
+
"save_learning_rate": 1e-5,
|
55 |
+
"total_save_number": 10,
|
56 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
57 |
+
"optimize_class": get_optimize_class()[0],
|
58 |
+
"optimize_class_int": get_optimize_class()[1],
|
59 |
+
}
|
60 |
+
config.update(additional_config)
|
61 |
+
print("Training:", config["optimize_class"])
|
62 |
+
|
63 |
+
|
64 |
+
|
65 |
+
|
66 |
+
# Data
|
67 |
+
dataset = Dataset(
|
68 |
+
root=config["dataset_root"],
|
69 |
+
train=True,
|
70 |
+
optimize_class=config["optimize_class"],
|
71 |
+
)
|
72 |
+
train_loader = DataLoader(
|
73 |
+
dataset=dataset,
|
74 |
+
batch_size=config["batch_size"],
|
75 |
+
num_workers=config["num_workers"],
|
76 |
+
shuffle=True,
|
77 |
+
drop_last=True,
|
78 |
+
pin_memory=True,
|
79 |
+
persistent_workers=True,
|
80 |
+
)
|
81 |
+
test_loader = DataLoader(
|
82 |
+
dataset=Dataset(
|
83 |
+
root=config["dataset_root"],
|
84 |
+
train=False,
|
85 |
+
optimize_class=config["optimize_class"],
|
86 |
+
),
|
87 |
+
batch_size=config["batch_size"],
|
88 |
+
num_workers=config["num_workers"],
|
89 |
+
shuffle=False,
|
90 |
+
)
|
91 |
+
|
92 |
+
# Model
|
93 |
+
model, head = Model()
|
94 |
+
model = model.to(device)
|
95 |
+
class FocalLoss(nn.Module):
|
96 |
+
def __init__(self, weight=None, gamma=2):
|
97 |
+
super(FocalLoss, self).__init__()
|
98 |
+
self.weight = weight
|
99 |
+
self.gamma = gamma
|
100 |
+
def forward(self, input, target):
|
101 |
+
ce_loss = F.cross_entropy(input, target, reduction='none', weight=self.weight)
|
102 |
+
pt = torch.exp(-ce_loss)
|
103 |
+
focal_loss = (1 - pt) ** self.gamma * ce_loss
|
104 |
+
return focal_loss.mean()
|
105 |
+
criterion = FocalLoss()
|
106 |
+
|
107 |
+
# Optimizer
|
108 |
+
head_optimizer = optim.AdamW(
|
109 |
+
head.parameters(),
|
110 |
+
lr=config["pre_learning_rate"],
|
111 |
+
weight_decay=config["weight_decay"],
|
112 |
+
)
|
113 |
+
optimizer = optim.AdamW(
|
114 |
+
model.parameters(),
|
115 |
+
lr=config["learning_rate"],
|
116 |
+
weight_decay=config["weight_decay"],
|
117 |
+
)
|
118 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
119 |
+
optimizer,
|
120 |
+
T_max=config["epochs"],
|
121 |
+
eta_min=config["save_learning_rate"],
|
122 |
+
)
|
123 |
+
|
124 |
+
|
125 |
+
|
126 |
+
|
127 |
+
# Training
|
128 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
129 |
+
model.train()
|
130 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
131 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
132 |
+
optimizer.zero_grad()
|
133 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
134 |
+
outputs = model(inputs)
|
135 |
+
loss = criterion(outputs, targets)
|
136 |
+
loss.backward()
|
137 |
+
optimizer.step()
|
138 |
+
if scheduler is not None:
|
139 |
+
scheduler.step()
|
140 |
+
|
141 |
+
# test
|
142 |
+
@torch.no_grad()
|
143 |
+
def test(model=model):
|
144 |
+
model.eval()
|
145 |
+
all_targets = []
|
146 |
+
all_predicts = []
|
147 |
+
test_loss = 0
|
148 |
+
correct = 0
|
149 |
+
total = 0
|
150 |
+
for batch_idx, (inputs, targets) in enumerate(test_loader):
|
151 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
152 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
153 |
+
outputs = model(inputs)
|
154 |
+
loss = criterion(outputs, targets)
|
155 |
+
# to logging losses
|
156 |
+
all_targets.extend(targets.flatten().tolist())
|
157 |
+
test_loss += loss.item()
|
158 |
+
_, predicts = outputs.max(1)
|
159 |
+
all_predicts.extend(predicts.flatten().tolist())
|
160 |
+
total += targets.size(0)
|
161 |
+
correct += predicts.eq(targets).sum().item()
|
162 |
+
loss = test_loss / (batch_idx + 1)
|
163 |
+
acc = correct / total
|
164 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}\n")
|
165 |
+
model.train()
|
166 |
+
return loss, acc, all_targets, all_predicts
|
167 |
+
|
168 |
+
# save train
|
169 |
+
def save_train(model=model, optimizer=optimizer):
|
170 |
+
data_loader = DataLoader(
|
171 |
+
dataset=dataset,
|
172 |
+
batch_size=min(len(dataset) // config["total_save_number"], config["batch_size"]),
|
173 |
+
num_workers=config["num_workers"],
|
174 |
+
shuffle=True,
|
175 |
+
drop_last=True,
|
176 |
+
)
|
177 |
+
model.train()
|
178 |
+
for batch_idx, (inputs, targets) in enumerate(data_loader):
|
179 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
180 |
+
optimizer.zero_grad()
|
181 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
182 |
+
outputs = model(inputs)
|
183 |
+
loss = criterion(outputs, targets)
|
184 |
+
loss.backward()
|
185 |
+
optimizer.step()
|
186 |
+
# Save checkpoint
|
187 |
+
_, acc, _, _ = test(model=model)
|
188 |
+
if not os.path.isdir('checkpoint'):
|
189 |
+
os.mkdir('checkpoint')
|
190 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items()}
|
191 |
+
torch.save(save_state, f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_class{config['optimize_class_int']}_{config['tag']}.pth")
|
192 |
+
print("save:", f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_class{config['optimize_class_int']}_{config['tag']}.pth")
|
193 |
+
# exit loop
|
194 |
+
if batch_idx+1 == config["total_save_number"]:
|
195 |
+
break
|
196 |
+
|
197 |
+
|
198 |
+
|
199 |
+
|
200 |
+
# main
|
201 |
+
if __name__ == '__main__':
|
202 |
+
for epoch in range(config["pre_epochs"]):
|
203 |
+
train(model=model, optimizer=head_optimizer, scheduler=None)
|
204 |
+
test(model=model)
|
205 |
+
for epoch in range(config["epochs"]):
|
206 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
207 |
+
test(model=model)
|
208 |
+
save_train(model=model, optimizer=optimizer)
|
dataset/condition_imageinput_vittiny/train.sh
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/bash
|
2 |
+
|
3 |
+
start=0
|
4 |
+
end=9
|
5 |
+
|
6 |
+
for i in $(seq $start $end)
|
7 |
+
do
|
8 |
+
power=$((2**i))
|
9 |
+
CUDA_VISIBLE_DEVICES=5 python train.py class$power
|
10 |
+
sleep 1
|
11 |
+
done
|
dataset/condition_permutation_vittiny/model.py
ADDED
@@ -0,0 +1,18 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import timm
|
4 |
+
|
5 |
+
|
6 |
+
def Model():
|
7 |
+
model = timm.create_model("vit_tiny_patch16_224", pretrained=False)
|
8 |
+
model.head = nn.Linear(192, 10)
|
9 |
+
return model, model.head
|
10 |
+
|
11 |
+
|
12 |
+
if __name__ == "__main__":
|
13 |
+
model, _ = Model()
|
14 |
+
print(model)
|
15 |
+
num_param = 0
|
16 |
+
for v in model.parameters():
|
17 |
+
num_param += v.numel()
|
18 |
+
print("num_param:", num_param)
|
dataset/condition_permutation_vittiny/test.py
ADDED
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
if __name__ == "__main__":
|
4 |
+
from train import *
|
5 |
+
else: # relative import
|
6 |
+
from .train import *
|
7 |
+
|
8 |
+
|
9 |
+
|
10 |
+
|
11 |
+
try:
|
12 |
+
test_item = sys.argv[1]
|
13 |
+
except IndexError:
|
14 |
+
assert __name__ == "__main__"
|
15 |
+
test_item = "./checkpoint"
|
16 |
+
test_items = []
|
17 |
+
if os.path.isdir(test_item):
|
18 |
+
for item in os.listdir(test_item):
|
19 |
+
item = os.path.join(test_item, item)
|
20 |
+
test_items.append(item)
|
21 |
+
elif os.path.isfile(test_item):
|
22 |
+
test_items.append(test_item)
|
23 |
+
|
24 |
+
|
25 |
+
|
26 |
+
|
27 |
+
for item in test_items:
|
28 |
+
print(f"testing: {item}")
|
29 |
+
state = torch.load(item, map_location="cpu")
|
30 |
+
model.load_state_dict({key: value.to(torch.float32).to(device) for key, value in state.items()})
|
31 |
+
loss, acc, all_targets, all_predicts = test(model=model)
|
dataset/condition_permutation_vittiny/train.py
ADDED
@@ -0,0 +1,210 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# set global seed
|
2 |
+
import time
|
3 |
+
print("time stamp:", time.time())
|
4 |
+
import random
|
5 |
+
import numpy as np
|
6 |
+
import torch
|
7 |
+
import re
|
8 |
+
import sys
|
9 |
+
if __name__ == "__main__":
|
10 |
+
def get_permutation_state():
|
11 |
+
try: # get string
|
12 |
+
string = sys.argv[1]
|
13 |
+
except IndexError:
|
14 |
+
RuntimeError("sys.argv[1] not found")
|
15 |
+
class_int_string = str(re.search(r'class(\d+)', string).group(1)).zfill(4)
|
16 |
+
return int(class_int_string)
|
17 |
+
seed = SEED = get_permutation_state()
|
18 |
+
else: # when testing
|
19 |
+
seed = SEED = 0
|
20 |
+
torch.manual_seed(seed)
|
21 |
+
torch.cuda.manual_seed(seed)
|
22 |
+
torch.cuda.manual_seed_all(seed)
|
23 |
+
torch.backends.cudnn.deterministic = True
|
24 |
+
torch.backends.cudnn.benchmark = True
|
25 |
+
np.random.seed(seed)
|
26 |
+
random.seed(seed)
|
27 |
+
print("Seed:", SEED)
|
28 |
+
|
29 |
+
try: # relative import
|
30 |
+
from model import Model
|
31 |
+
except ImportError:
|
32 |
+
from .model import Model
|
33 |
+
|
34 |
+
# import
|
35 |
+
import torch.nn as nn
|
36 |
+
from torch import optim
|
37 |
+
from torch.optim import lr_scheduler
|
38 |
+
from torch.utils.data import DataLoader
|
39 |
+
from torchvision.datasets import CIFAR10 as Dataset
|
40 |
+
from torchvision import transforms
|
41 |
+
from torch.nn import functional as F
|
42 |
+
import warnings
|
43 |
+
warnings.filterwarnings("ignore", category=UserWarning)
|
44 |
+
|
45 |
+
# load additional config
|
46 |
+
import os
|
47 |
+
import json
|
48 |
+
config_file = os.path.join(os.path.dirname(os.path.dirname(__file__)), "config.json")
|
49 |
+
with open(config_file, "r") as f:
|
50 |
+
additional_config = json.load(f)
|
51 |
+
|
52 |
+
|
53 |
+
|
54 |
+
|
55 |
+
# config
|
56 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
57 |
+
config = {
|
58 |
+
"dataset_root": "from_additional_config",
|
59 |
+
"batch_size": 250 if __name__ == "__main__" else 50,
|
60 |
+
"num_workers": 16,
|
61 |
+
"learning_rate": 5e-3,
|
62 |
+
"epochs": 200,
|
63 |
+
"weight_decay": 0.1,
|
64 |
+
"save_learning_rate": 2e-5,
|
65 |
+
"total_save_number": 5,
|
66 |
+
"tag": os.path.basename(os.path.dirname(__file__)),
|
67 |
+
}
|
68 |
+
config.update(additional_config)
|
69 |
+
|
70 |
+
|
71 |
+
|
72 |
+
|
73 |
+
# Data
|
74 |
+
dataset = Dataset(
|
75 |
+
root=config["dataset_root"],
|
76 |
+
train=True,
|
77 |
+
download=True,
|
78 |
+
transform=transforms.Compose([
|
79 |
+
transforms.Resize(224),
|
80 |
+
transforms.RandomCrop(224, padding=32),
|
81 |
+
transforms.RandomHorizontalFlip(),
|
82 |
+
transforms.AutoAugment(policy=transforms.AutoAugmentPolicy("cifar10")),
|
83 |
+
transforms.ToTensor(),
|
84 |
+
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2471, 0.2435, 0.2616)),
|
85 |
+
])
|
86 |
+
)
|
87 |
+
train_loader = DataLoader(
|
88 |
+
dataset=dataset,
|
89 |
+
batch_size=config["batch_size"],
|
90 |
+
num_workers=config["num_workers"],
|
91 |
+
shuffle=True,
|
92 |
+
drop_last=True,
|
93 |
+
pin_memory=True,
|
94 |
+
persistent_workers=True,
|
95 |
+
)
|
96 |
+
test_loader = DataLoader(
|
97 |
+
dataset=Dataset(
|
98 |
+
root=config["dataset_root"],
|
99 |
+
train=False,
|
100 |
+
download=True,
|
101 |
+
transform=transforms.Compose([
|
102 |
+
transforms.Resize(224),
|
103 |
+
transforms.ToTensor(),
|
104 |
+
transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2471, 0.2435, 0.2616)),
|
105 |
+
])),
|
106 |
+
batch_size=config["batch_size"],
|
107 |
+
num_workers=config["num_workers"],
|
108 |
+
shuffle=False,
|
109 |
+
)
|
110 |
+
|
111 |
+
# Model
|
112 |
+
model, head = Model()
|
113 |
+
model = model.to(device)
|
114 |
+
criterion = nn.CrossEntropyLoss()
|
115 |
+
|
116 |
+
# Optimizer
|
117 |
+
optimizer = optim.AdamW(
|
118 |
+
model.parameters(),
|
119 |
+
lr=config["learning_rate"],
|
120 |
+
weight_decay=config["weight_decay"],
|
121 |
+
)
|
122 |
+
scheduler = lr_scheduler.CosineAnnealingLR(
|
123 |
+
optimizer,
|
124 |
+
T_max=config["epochs"],
|
125 |
+
eta_min=config["save_learning_rate"],
|
126 |
+
)
|
127 |
+
|
128 |
+
|
129 |
+
|
130 |
+
|
131 |
+
# Training
|
132 |
+
def train(model=model, optimizer=optimizer, scheduler=scheduler):
|
133 |
+
model.train()
|
134 |
+
for batch_idx, (inputs, targets) in enumerate(train_loader):
|
135 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
136 |
+
optimizer.zero_grad()
|
137 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
138 |
+
outputs = model(inputs)
|
139 |
+
loss = criterion(outputs, targets)
|
140 |
+
loss.backward()
|
141 |
+
optimizer.step()
|
142 |
+
if scheduler is not None:
|
143 |
+
scheduler.step()
|
144 |
+
|
145 |
+
# test
|
146 |
+
@torch.no_grad()
|
147 |
+
def test(model=model):
|
148 |
+
model.eval()
|
149 |
+
all_targets = []
|
150 |
+
all_predicts = []
|
151 |
+
test_loss = 0
|
152 |
+
correct = 0
|
153 |
+
total = 0
|
154 |
+
for batch_idx, (inputs, targets) in enumerate(test_loader):
|
155 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
156 |
+
with torch.cuda.amp.autocast(enabled=True, dtype=torch.bfloat16):
|
157 |
+
outputs = model(inputs)
|
158 |
+
loss = criterion(outputs, targets)
|
159 |
+
# to logging losses
|
160 |
+
all_targets.extend(targets.flatten().tolist())
|
161 |
+
test_loss += loss.item()
|
162 |
+
_, predicts = outputs.max(1)
|
163 |
+
all_predicts.extend(predicts.flatten().tolist())
|
164 |
+
total += targets.size(0)
|
165 |
+
correct += predicts.eq(targets).sum().item()
|
166 |
+
loss = test_loss / (batch_idx + 1)
|
167 |
+
acc = correct / total
|
168 |
+
print(f"Loss: {loss:.4f} | Acc: {acc:.4f}\n")
|
169 |
+
model.train()
|
170 |
+
return loss, acc, all_targets, all_predicts
|
171 |
+
|
172 |
+
# save train
|
173 |
+
def save_train(model=model, optimizer=optimizer):
|
174 |
+
data_loader = DataLoader(
|
175 |
+
dataset=dataset,
|
176 |
+
batch_size=min(len(dataset) // config["total_save_number"], config["batch_size"]),
|
177 |
+
num_workers=config["num_workers"],
|
178 |
+
shuffle=True,
|
179 |
+
drop_last=True,
|
180 |
+
)
|
181 |
+
model.train()
|
182 |
+
for batch_idx, (inputs, targets) in enumerate(data_loader):
|
183 |
+
inputs, targets = inputs.to(device), targets.to(device)
|
184 |
+
optimizer.zero_grad()
|
185 |
+
with torch.cuda.amp.autocast(enabled=False, dtype=torch.bfloat16):
|
186 |
+
outputs = model(inputs)
|
187 |
+
loss = criterion(outputs, targets)
|
188 |
+
loss.backward()
|
189 |
+
optimizer.step()
|
190 |
+
# Save checkpoint
|
191 |
+
_, acc, _, _ = test(model=model)
|
192 |
+
if not os.path.isdir('checkpoint'):
|
193 |
+
os.mkdir('checkpoint')
|
194 |
+
save_state = {key: value.cpu().to(torch.float32) for key, value in model.state_dict().items()}
|
195 |
+
torch.save(save_state, f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_class{SEED:04d}_{config['tag']}.pth")
|
196 |
+
print("save:", f"checkpoint/{str(batch_idx).zfill(4)}_acc{acc:.4f}_class{SEED:04d}_{config['tag']}.pth")
|
197 |
+
# exit loop
|
198 |
+
if batch_idx+1 == config["total_save_number"]:
|
199 |
+
break
|
200 |
+
|
201 |
+
|
202 |
+
|
203 |
+
|
204 |
+
# main
|
205 |
+
if __name__ == '__main__':
|
206 |
+
for epoch in range(config["epochs"]):
|
207 |
+
train(model=model, optimizer=optimizer, scheduler=scheduler)
|
208 |
+
test(model=model)
|
209 |
+
save_train(model=model, optimizer=optimizer)
|
210 |
+
print("time stamp:", time.time())
|
dataset/condition_permutation_vittiny/train.sh
ADDED
@@ -0,0 +1,10 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/bin/bash
|
2 |
+
|
3 |
+
start=0
|
4 |
+
end=19
|
5 |
+
|
6 |
+
for i in $(seq $start $end)
|
7 |
+
do
|
8 |
+
python train.py class$i
|
9 |
+
sleep 1
|
10 |
+
done
|
dataset/config.json
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
{"dataset_root": "path_to_your_dataset", "imagenet_root": {"train": null, "test": null}, "dora_root": "/home/wangkai/arpgen/DoRA/commonsense_reasoning", "dora_env_name": "dora_llama"}
|
dataset/dataset.py
ADDED
@@ -0,0 +1,327 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import einops
|
3 |
+
from torch.utils.data import Dataset
|
4 |
+
from torchvision.datasets import CIFAR10
|
5 |
+
from torchvision import transforms
|
6 |
+
import os
|
7 |
+
import math
|
8 |
+
import random
|
9 |
+
import json
|
10 |
+
from abc import ABC
|
11 |
+
import pickle
|
12 |
+
|
13 |
+
|
14 |
+
|
15 |
+
|
16 |
+
def pad_to_length(x, common_factor, **config):
|
17 |
+
if x.numel() % common_factor == 0:
|
18 |
+
return x.flatten()
|
19 |
+
# print(f"padding {x.shape} according to {common_factor}")
|
20 |
+
full_length = (x.numel() // common_factor + 1) * common_factor
|
21 |
+
padding_length = full_length - len(x.flatten())
|
22 |
+
padding = torch.full([padding_length, ], dtype=x.dtype, device=x.device, fill_value=config["fill_value"])
|
23 |
+
x = torch.cat((x.flatten(), padding), dim=0)
|
24 |
+
return x
|
25 |
+
|
26 |
+
def layer_to_token(x, common_factor, **config):
|
27 |
+
if config["granularity"] == 2: # split by output
|
28 |
+
if x.numel() <= common_factor:
|
29 |
+
return pad_to_length(x.flatten(), common_factor, **config)[None]
|
30 |
+
dim2 = x[0].numel()
|
31 |
+
dim1 = x.shape[0]
|
32 |
+
if dim2 <= common_factor:
|
33 |
+
i = int(dim1 / (common_factor / dim2))
|
34 |
+
while True:
|
35 |
+
if dim1 % i == 0 and dim2 * (dim1 // i) <= common_factor:
|
36 |
+
output = x.view(-1, dim2 * (dim1 // i))
|
37 |
+
output = [pad_to_length(item, common_factor, **config) for item in output]
|
38 |
+
return torch.stack(output, dim=0)
|
39 |
+
i += 1
|
40 |
+
else: # dim2 > common_factor
|
41 |
+
output = [layer_to_token(item, common_factor, **config) for item in x]
|
42 |
+
return torch.cat(output, dim=0)
|
43 |
+
elif config["granularity"] == 1: # split by layer
|
44 |
+
return pad_to_length(x.flatten(), common_factor, **config).view(-1, common_factor)
|
45 |
+
elif config["granularity"] == 0: # flatten directly
|
46 |
+
return x.flatten()
|
47 |
+
else: # NotImplementedError
|
48 |
+
raise NotImplementedError("granularity: 0: flatten directly, 1: split by layer, 2: split by output dim")
|
49 |
+
|
50 |
+
|
51 |
+
def token_to_layer(tokens, shape, **config):
|
52 |
+
common_factor = tokens.shape[-1]
|
53 |
+
if config["granularity"] == 2: # split by output
|
54 |
+
num_element = math.prod(shape)
|
55 |
+
if num_element <= common_factor:
|
56 |
+
param = tokens[0][:num_element].view(shape)
|
57 |
+
tokens = tokens[1:]
|
58 |
+
return param, tokens
|
59 |
+
dim2 = num_element // shape[0]
|
60 |
+
dim1 = shape[0]
|
61 |
+
if dim2 <= common_factor:
|
62 |
+
i = int(dim1 / (common_factor / dim2))
|
63 |
+
while True:
|
64 |
+
if dim1 % i == 0 and dim2 * (dim1 // i) <= common_factor:
|
65 |
+
item_per_token = dim2 * (dim1 // i)
|
66 |
+
length = num_element // item_per_token
|
67 |
+
output = [item[:item_per_token] for item in tokens[:length]]
|
68 |
+
param = torch.cat(output, dim=0).view(shape)
|
69 |
+
tokens = tokens[length:]
|
70 |
+
return param, tokens
|
71 |
+
i += 1
|
72 |
+
else: # dim2 > common_factor
|
73 |
+
output = []
|
74 |
+
for i in range(shape[0]):
|
75 |
+
param, tokens = token_to_layer(tokens, shape[1:], **config)
|
76 |
+
output.append(param.flatten())
|
77 |
+
param = torch.cat(output, dim=0).view(shape)
|
78 |
+
return param, tokens
|
79 |
+
elif config["granularity"] == 1: # split by layer
|
80 |
+
num_element = math.prod(shape)
|
81 |
+
token_num = num_element // common_factor if num_element % common_factor == 0 \
|
82 |
+
else num_element // common_factor + 1
|
83 |
+
param = tokens.flatten()[:num_element].view(shape)
|
84 |
+
tokens = tokens[token_num:]
|
85 |
+
return param, tokens
|
86 |
+
elif config["granularity"] == 0: # flatten directly
|
87 |
+
num_element = math.prod(shape)
|
88 |
+
param = tokens.flatten()[:num_element].view(shape)
|
89 |
+
tokens = pad_to_length(tokens.flatten()[num_element:],
|
90 |
+
common_factor, fill_value=torch.nan).view(-1, common_factor)
|
91 |
+
return param, tokens
|
92 |
+
else: # NotImplementedError
|
93 |
+
raise NotImplementedError("granularity: 0: flatten directly, 1: split by layer, 2: split by output dim")
|
94 |
+
|
95 |
+
|
96 |
+
def positional_embedding_2d(dim1, dim2, d_model):
|
97 |
+
assert d_model % 4 == 0, f"Cannot use sin/cos positional encoding with odd dimension {d_model}"
|
98 |
+
pe = torch.zeros(d_model, dim1, dim2)
|
99 |
+
d_model = int(d_model / 2) # Each dimension use half of d_model
|
100 |
+
div_term = torch.exp(torch.arange(0., d_model, 2, dtype=torch.float32) * -(math.log(10000.0) / d_model))
|
101 |
+
pos_w = torch.arange(0., dim2).unsqueeze(1)
|
102 |
+
pos_h = torch.arange(0., dim1).unsqueeze(1)
|
103 |
+
pe[0:d_model:2, :, :] = torch.sin(pos_w * div_term).transpose(0, 1).unsqueeze(1).repeat(1, dim1, 1)
|
104 |
+
pe[1:d_model:2, :, :] = torch.cos(pos_w * div_term).transpose(0, 1).unsqueeze(1).repeat(1, dim1, 1)
|
105 |
+
pe[d_model::2, :, :] = torch.sin(pos_h * div_term).transpose(0, 1).unsqueeze(2).repeat(1, 1, dim2)
|
106 |
+
pe[d_model+1::2, :, :] = torch.cos(pos_h * div_term).transpose(0, 1).unsqueeze(2).repeat(1, 1, dim2)
|
107 |
+
return pe.permute(1, 2, 0)
|
108 |
+
|
109 |
+
|
110 |
+
def positional_embedding_1d(dim1, d_model):
|
111 |
+
pe = torch.zeros(dim1, d_model)
|
112 |
+
position = torch.arange(0, dim1, dtype=torch.float).unsqueeze(1)
|
113 |
+
div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))
|
114 |
+
pe[:, 0::2] = torch.sin(position * div_term)
|
115 |
+
pe[:, 1::2] = torch.cos(position * div_term)
|
116 |
+
return pe
|
117 |
+
|
118 |
+
|
119 |
+
|
120 |
+
|
121 |
+
class BaseDataset(Dataset, ABC):
|
122 |
+
data_path = None
|
123 |
+
generated_path = None
|
124 |
+
test_command = None
|
125 |
+
config = {
|
126 |
+
"fill_value": torch.nan,
|
127 |
+
"granularity": 1, # 0: flatten directly, 1: split by layer, 2: split by output
|
128 |
+
"pe_granularity": 2, # 0: no embedding, 1: 1d embedding, 2: 2d embedding
|
129 |
+
}
|
130 |
+
|
131 |
+
def __init__(self, checkpoint_path=None, dim_per_token=8192, **kwargs):
|
132 |
+
if not os.path.exists(self.data_path):
|
133 |
+
os.makedirs(self.data_path, exist_ok=False)
|
134 |
+
if self.generated_path is not None and not os.path.exists(os.path.dirname(self.generated_path)):
|
135 |
+
os.makedirs(os.path.dirname(self.generated_path))
|
136 |
+
self.config.update(kwargs)
|
137 |
+
checkpoint_path = self.data_path if checkpoint_path is None else checkpoint_path
|
138 |
+
assert os.path.exists(checkpoint_path)
|
139 |
+
self.dim_per_token = dim_per_token
|
140 |
+
self.structure = None # set in get_structure()
|
141 |
+
self.sequence_length = None # set in get_structure()
|
142 |
+
# load checkpoint_list
|
143 |
+
checkpoint_list = os.listdir(checkpoint_path)
|
144 |
+
self.checkpoint_list = list([os.path.join(checkpoint_path, item) for item in checkpoint_list])
|
145 |
+
self.length = self.real_length = len(self.checkpoint_list)
|
146 |
+
self.set_infinite_dataset()
|
147 |
+
# get structure
|
148 |
+
structure_cache_file = os.path.join(os.path.dirname(self.data_path), "structure.cache")
|
149 |
+
try: # try to load cache file
|
150 |
+
assert os.path.exists(structure_cache_file)
|
151 |
+
with open(structure_cache_file, "rb") as f:
|
152 |
+
print(f"Loading cache from {structure_cache_file}")
|
153 |
+
cache_file = pickle.load(f)
|
154 |
+
if len(self.checkpoint_list) != 0:
|
155 |
+
assert set(cache_file["checkpoint_list"]) == set(self.checkpoint_list)
|
156 |
+
self.structure = cache_file["structure"]
|
157 |
+
else: # empty checkpoint_list, only generate
|
158 |
+
print("Cannot find any trained checkpoint, loading cache file for generating!")
|
159 |
+
self.structure = cache_file["structure"]
|
160 |
+
fake_diction = {key: torch.zeros(item[0]) for key, item in self.structure.items()}
|
161 |
+
torch.save(fake_diction, os.path.join(checkpoint_path, "fake_checkpoint.pth"))
|
162 |
+
self.checkpoint_list.append(os.path.join(checkpoint_path, "fake_checkpoint.pth"))
|
163 |
+
self.length = self.real_length = len(self.checkpoint_list)
|
164 |
+
self.set_infinite_dataset()
|
165 |
+
os.system(f"rm {os.path.join(checkpoint_path, 'fake_checkpoint.pth')}")
|
166 |
+
except AssertionError: # recompute cache file
|
167 |
+
print("==> Organizing structure..")
|
168 |
+
self.structure = self.get_structure()
|
169 |
+
with open(structure_cache_file, "wb") as f:
|
170 |
+
pickle.dump({"structure": self.structure, "checkpoint_list": self.checkpoint_list}, f)
|
171 |
+
# get sequence_length
|
172 |
+
self.sequence_length = self.get_sequence_length()
|
173 |
+
|
174 |
+
def get_sequence_length(self):
|
175 |
+
fake_diction = {key: torch.zeros(item[0]) for key, item in self.structure.items()}
|
176 |
+
# get sequence_length
|
177 |
+
param = self.preprocess(fake_diction)
|
178 |
+
self.sequence_length = param.size(0)
|
179 |
+
return self.sequence_length
|
180 |
+
|
181 |
+
def get_structure(self):
|
182 |
+
# get structure
|
183 |
+
checkpoint_list = self.checkpoint_list
|
184 |
+
structures = [{} for _ in range(len(checkpoint_list))]
|
185 |
+
for i, checkpoint in enumerate(checkpoint_list):
|
186 |
+
diction = torch.load(checkpoint, map_location="cpu")
|
187 |
+
for key, value in diction.items():
|
188 |
+
if ("num_batches_tracked" in key) or (value.numel() == 1) or not torch.is_floating_point(value):
|
189 |
+
structures[i][key] = (value.shape, value, None)
|
190 |
+
elif "running_var" in key:
|
191 |
+
pre_mean = value.mean() * 0.95
|
192 |
+
value = torch.log(value / pre_mean + 0.05)
|
193 |
+
structures[i][key] = (value.shape, pre_mean, value.mean(), value.std())
|
194 |
+
else: # conv & linear
|
195 |
+
structures[i][key] = (value.shape, value.mean(), value.std())
|
196 |
+
final_structure = {}
|
197 |
+
structure_diction = torch.load(checkpoint_list[0], map_location="cpu")
|
198 |
+
for key, param in structure_diction.items():
|
199 |
+
if ("num_batches_tracked" in key) or (param.numel() == 1) or not torch.is_floating_point(param):
|
200 |
+
final_structure[key] = (param.shape, param, None)
|
201 |
+
elif "running_var" in key:
|
202 |
+
value = [param.shape, 0., 0., 0.]
|
203 |
+
for structure in structures:
|
204 |
+
for i in [1, 2, 3]:
|
205 |
+
value[i] += structure[key][i]
|
206 |
+
for i in [1, 2, 3]:
|
207 |
+
value[i] /= len(structures)
|
208 |
+
final_structure[key] = tuple(value)
|
209 |
+
else: # conv & linear
|
210 |
+
value = [param.shape, 0., 0.]
|
211 |
+
for structure in structures:
|
212 |
+
for i in [1, 2]:
|
213 |
+
value[i] += structure[key][i]
|
214 |
+
for i in [1, 2]:
|
215 |
+
value[i] /= len(structures)
|
216 |
+
final_structure[key] = tuple(value)
|
217 |
+
self.structure = final_structure
|
218 |
+
return self.structure
|
219 |
+
|
220 |
+
def set_infinite_dataset(self, max_num=None):
|
221 |
+
if max_num is None:
|
222 |
+
max_num = self.length * 1000000
|
223 |
+
self.length = max_num
|
224 |
+
return self
|
225 |
+
|
226 |
+
@property
|
227 |
+
def max_permutation_state(self):
|
228 |
+
return self.real_length
|
229 |
+
|
230 |
+
def get_position_embedding(self, positional_embedding_dim=None):
|
231 |
+
if positional_embedding_dim is None:
|
232 |
+
positional_embedding_dim = self.dim_per_token // 2
|
233 |
+
assert self.structure is not None, "run get_structure before get_position_embedding"
|
234 |
+
if self.config["pe_granularity"] == 2:
|
235 |
+
print("Use 2d positional embedding")
|
236 |
+
positional_embedding_index = []
|
237 |
+
for key, item in self.structure.items():
|
238 |
+
if ("num_batches_tracked" in key) or (item[-1] is None):
|
239 |
+
continue
|
240 |
+
else: # conv & linear
|
241 |
+
shape, *_ = item
|
242 |
+
fake_param = torch.ones(size=shape)
|
243 |
+
fake_param = layer_to_token(fake_param, self.dim_per_token, **self.config)
|
244 |
+
positional_embedding_index.append(list(range(fake_param.size(0))))
|
245 |
+
dim1 = len(positional_embedding_index)
|
246 |
+
dim2 = max([len(token_per_layer) for token_per_layer in positional_embedding_index])
|
247 |
+
full_pe = positional_embedding_2d(dim1, dim2, positional_embedding_dim)
|
248 |
+
positional_embedding = []
|
249 |
+
for layer_index, token_indexes in enumerate(positional_embedding_index):
|
250 |
+
for token_index in token_indexes:
|
251 |
+
this_pe = full_pe[layer_index, token_index]
|
252 |
+
positional_embedding.append(this_pe)
|
253 |
+
positional_embedding = torch.stack(positional_embedding)
|
254 |
+
return positional_embedding
|
255 |
+
elif self.config["pe_granularity"] == 1:
|
256 |
+
print("Use 1d positional embedding")
|
257 |
+
return positional_embedding_1d(self.sequence_length, positional_embedding_dim)
|
258 |
+
elif self.config["pe_granularity"] == 0:
|
259 |
+
print("Not use positional embedding")
|
260 |
+
return torch.zeros_like(self.__getitem__(0))
|
261 |
+
else: # NotImplementedError
|
262 |
+
raise NotImplementedError("pe_granularity: 0: no embedding, 1: 1d embedding, 2: 2d embedding")
|
263 |
+
|
264 |
+
def __len__(self):
|
265 |
+
return self.length
|
266 |
+
|
267 |
+
def __getitem__(self, index):
|
268 |
+
index = index % self.real_length
|
269 |
+
diction = torch.load(self.checkpoint_list[index], map_location="cpu")
|
270 |
+
param = self.preprocess(diction)
|
271 |
+
return param, index
|
272 |
+
|
273 |
+
def save_params(self, params, save_path):
|
274 |
+
diction = self.postprocess(params.cpu().to(torch.float32))
|
275 |
+
torch.save(diction, save_path)
|
276 |
+
|
277 |
+
def preprocess(self, diction: dict, **kwargs) -> torch.Tensor:
|
278 |
+
param_list = []
|
279 |
+
for key, value in diction.items():
|
280 |
+
if ("num_batches_tracked" in key) or (value.numel() == 1) or not torch.is_floating_point(value):
|
281 |
+
continue
|
282 |
+
elif "running_var" in key:
|
283 |
+
shape, pre_mean, mean, std = self.structure[key]
|
284 |
+
value = torch.log(value / pre_mean + 0.05)
|
285 |
+
else: # normal
|
286 |
+
shape, mean, std = self.structure[key]
|
287 |
+
value = (value - mean) / std
|
288 |
+
value = layer_to_token(value, self.dim_per_token, **self.config)
|
289 |
+
param_list.append(value)
|
290 |
+
param = torch.cat(param_list, dim=0)
|
291 |
+
if self.config["granularity"] == 0: # padding directly process tail
|
292 |
+
param = pad_to_length(param, self.dim_per_token, **self.config).view(-1, self.dim_per_token)
|
293 |
+
# print("Sequence length:", param.size(0))
|
294 |
+
return param.to(torch.float32)
|
295 |
+
|
296 |
+
def postprocess(self, params: torch.Tensor, **kwargs) -> dict:
|
297 |
+
diction = {}
|
298 |
+
params = params if len(params.shape) == 2 else params.squeeze(0)
|
299 |
+
for key, item in self.structure.items():
|
300 |
+
if ("num_batches_tracked" in key) or (item[-1] is None):
|
301 |
+
shape, mean, std = item
|
302 |
+
diction[key] = mean
|
303 |
+
continue
|
304 |
+
elif "running_var" in key:
|
305 |
+
shape, pre_mean, mean, std = item
|
306 |
+
else: # conv & linear
|
307 |
+
shape, mean, std = item
|
308 |
+
this_param, params = token_to_layer(params, shape, **self.config)
|
309 |
+
this_param = this_param * std + mean
|
310 |
+
if "running_var" in key:
|
311 |
+
this_param = torch.clip(torch.exp(this_param) - 0.05, min=0.001) * pre_mean
|
312 |
+
diction[key] = this_param
|
313 |
+
return diction
|
314 |
+
|
315 |
+
|
316 |
+
class ConditionalDataset(BaseDataset, ABC):
|
317 |
+
def _extract_condition(self, index: int):
|
318 |
+
name = self.checkpoint_list[index]
|
319 |
+
condition_list = os.path.basename(name).split("_")
|
320 |
+
return condition_list
|
321 |
+
|
322 |
+
def __getitem__(self, index):
|
323 |
+
index = index % self.real_length
|
324 |
+
diction = torch.load(self.checkpoint_list[index], map_location="cpu")
|
325 |
+
condition = self._extract_condition(index)
|
326 |
+
param = self.preprocess(diction)
|
327 |
+
return param, condition
|
dataset/downtask_detection/README.md
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
Code for segmentation is coming...
|
dataset/downtask_detection/test.sh
ADDED
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env bash
|
2 |
+
|
3 |
+
source /path/to/miniconda3/bin/activate /path/to/miniconda3/envs/environment
|
4 |
+
|
5 |
+
CLUSTER=True \
|
6 |
+
DETECTRON2_DATASETS="/path/to/" \
|
7 |
+
PYTHONPATH="$(dirname $0)/Detection":$PYTHONPATH \
|
8 |
+
python $(dirname $0)/Detection/tools/lazyconfig_train_net.py --config-file $(dirname $0)/Detection/projects/ViTDet/configs/COCO/our_vit_b_100ep.py --finetune "VIT_BASE_IN21K" \
|
9 |
+
--num-gpus 1 \
|
10 |
+
--fulltune \
|
11 |
+
--eval-only "train.init_checkpoint='$1'"
|
dataset/downtask_dora_r16/adapter_config.json
ADDED
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"Wdecompose_target_modules": null,
|
3 |
+
"base_model_name_or_path": "yahma/llama-7b-hf",
|
4 |
+
"bias": "none",
|
5 |
+
"dora_simple": true,
|
6 |
+
"enable_lora": null,
|
7 |
+
"fan_in_fan_out": false,
|
8 |
+
"inference_mode": true,
|
9 |
+
"lora_alpha": 32,
|
10 |
+
"lora_dropout": 0.05,
|
11 |
+
"merge_weights": false,
|
12 |
+
"modules_to_save": null,
|
13 |
+
"peft_type": "DORA",
|
14 |
+
"r": 16,
|
15 |
+
"target_modules": [
|
16 |
+
"q_proj",
|
17 |
+
"k_proj",
|
18 |
+
"v_proj",
|
19 |
+
"up_proj",
|
20 |
+
"down_proj"
|
21 |
+
],
|
22 |
+
"task_type": "CAUSAL_LM"
|
23 |
+
}
|