RobbiePasquale
commited on
Upload lightbulb_wm.py
Browse files- lightbulb_wm.py +1292 -0
lightbulb_wm.py
ADDED
@@ -0,0 +1,1292 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
|
2 |
+
import argparse
|
3 |
+
import math
|
4 |
+
import os
|
5 |
+
import torch
|
6 |
+
import torch.nn as nn
|
7 |
+
import torch.nn.functional as F
|
8 |
+
import torch.optim as optim
|
9 |
+
from torch.utils.data import DataLoader
|
10 |
+
import copy
|
11 |
+
from torch.optim.lr_scheduler import CosineAnnealingLR
|
12 |
+
from torch.amp import autocast, GradScaler
|
13 |
+
from datasets import load_dataset
|
14 |
+
from transformers import AutoTokenizer
|
15 |
+
|
16 |
+
|
17 |
+
# Set the device
|
18 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
19 |
+
|
20 |
+
|
21 |
+
def parse_args():
|
22 |
+
parser = argparse.ArgumentParser(description='Train World Model with Transformer outputs.')
|
23 |
+
parser.add_argument('--model_name', type=str, default='gpt2', help='Pretrained model name or path')
|
24 |
+
parser.add_argument('--dataset_name', type=str, default='wikitext', help='Dataset name from HuggingFace Datasets')
|
25 |
+
parser.add_argument('--dataset_config', type=str, default='wikitext-2-raw-v1', help='Dataset configuration name')
|
26 |
+
parser.add_argument('--batch_size', type=int, default=2, help='Batch size')
|
27 |
+
parser.add_argument('--num_epochs', type=int, default=3, help='Number of epochs')
|
28 |
+
parser.add_argument('--max_length', type=int, default=128, help='Maximum sequence length')
|
29 |
+
parser.add_argument('--mcts_iterations', type=int, default=5, help='Number of MCTS Iterations')
|
30 |
+
parser.add_argument('--mcts_exploration_constant', type=float, default=1.414, help='Learning rate')
|
31 |
+
parser.add_argument('--accumulation_steps', type=int, default=4, help='Gradient accumulation steps')
|
32 |
+
parser.add_argument('--learning_rate', type=float, default=1e-4, help='Learning rate')
|
33 |
+
parser.add_argument('--weight_decay', type=float, default=1e-2, help='Weight decay')
|
34 |
+
parser.add_argument('--alpha', type=float, default=0.1, help='Entropy regularization weight')
|
35 |
+
parser.add_argument('--beta', type=float, default=0.1, help='Variance regularization weight')
|
36 |
+
parser.add_argument('--max_grad_norm', type=float, default=1.0, help='Max gradient norm for clipping')
|
37 |
+
parser.add_argument('--save_dir', type=str, default='./models', help='Directory to save the models')
|
38 |
+
parser.add_argument('--temperature', type=float, default=1.0, help='Temperature parameter for entropy and variance')
|
39 |
+
parser.add_argument('--transformer_model_path', type=str, required=True, help='Path to the saved Transformer model')
|
40 |
+
args = parser.parse_args()
|
41 |
+
return args
|
42 |
+
|
43 |
+
|
44 |
+
def load_data(args, tokenizer):
|
45 |
+
# Load the dataset
|
46 |
+
dataset = load_dataset(args.dataset_name, args.dataset_config)
|
47 |
+
|
48 |
+
# Ensure the tokenizer has a padding token
|
49 |
+
if tokenizer.pad_token is None:
|
50 |
+
tokenizer.pad_token = tokenizer.eos_token
|
51 |
+
|
52 |
+
def tokenize_function(examples):
|
53 |
+
return tokenizer(examples['text'], truncation=True, max_length=args.max_length)
|
54 |
+
|
55 |
+
tokenized_datasets = dataset.map(
|
56 |
+
tokenize_function,
|
57 |
+
batched=True,
|
58 |
+
num_proc=4,
|
59 |
+
remove_columns=dataset['train'].column_names,
|
60 |
+
)
|
61 |
+
|
62 |
+
# Build inputs and labels for language modeling
|
63 |
+
block_size = args.max_length
|
64 |
+
|
65 |
+
def group_texts(examples):
|
66 |
+
# Concatenate all texts
|
67 |
+
concatenated_examples = {k: sum(examples[k], []) for k in examples.keys()}
|
68 |
+
total_length = len(concatenated_examples['input_ids'])
|
69 |
+
# We drop the small remainder
|
70 |
+
total_length = (total_length // block_size) * block_size
|
71 |
+
# Split by chunks of block_size
|
72 |
+
result = {
|
73 |
+
k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
|
74 |
+
for k, t in concatenated_examples.items()
|
75 |
+
}
|
76 |
+
result['labels'] = result['input_ids'].copy()
|
77 |
+
return result
|
78 |
+
|
79 |
+
lm_datasets = tokenized_datasets.map(
|
80 |
+
group_texts,
|
81 |
+
batched=True,
|
82 |
+
num_proc=4,
|
83 |
+
)
|
84 |
+
|
85 |
+
# Create DataLoader
|
86 |
+
train_dataset = lm_datasets['train']
|
87 |
+
eval_dataset = lm_datasets['validation'] if 'validation' in lm_datasets else lm_datasets['test']
|
88 |
+
|
89 |
+
data_collator = lambda data: {
|
90 |
+
'input_ids': torch.tensor([f['input_ids'] for f in data], dtype=torch.long),
|
91 |
+
'labels': torch.tensor([f['labels'] for f in data], dtype=torch.long)
|
92 |
+
}
|
93 |
+
|
94 |
+
train_loader = DataLoader(train_dataset, shuffle=True, batch_size=args.batch_size, collate_fn=data_collator)
|
95 |
+
eval_loader = DataLoader(eval_dataset, shuffle=False, batch_size=args.batch_size, collate_fn=data_collator)
|
96 |
+
|
97 |
+
return train_loader, eval_loader
|
98 |
+
|
99 |
+
|
100 |
+
def save_all_models(transformer_model, representation_network, dynamics_network, prediction_network, action_encoder, save_dir, epoch):
|
101 |
+
"""
|
102 |
+
Save all models to the specified directory.
|
103 |
+
|
104 |
+
Args:
|
105 |
+
transformer_model (nn.Module): Transformer model.
|
106 |
+
representation_network (nn.Module): Representation network.
|
107 |
+
dynamics_network (nn.Module): Dynamics network.
|
108 |
+
prediction_network (nn.Module): Prediction network.
|
109 |
+
action_encoder (nn.Module): Action encoder.
|
110 |
+
save_dir (str): Directory to save the models.
|
111 |
+
epoch (int): Current epoch number.
|
112 |
+
"""
|
113 |
+
os.makedirs(save_dir, exist_ok=True)
|
114 |
+
|
115 |
+
torch.save(transformer_model.state_dict(), os.path.join(save_dir, f'transformer_model_epoch_{epoch}.pt'))
|
116 |
+
torch.save(representation_network.state_dict(), os.path.join(save_dir, f'representation_network_epoch_{epoch}.pt'))
|
117 |
+
torch.save(dynamics_network.state_dict(), os.path.join(save_dir, f'dynamics_network_epoch_{epoch}.pt'))
|
118 |
+
torch.save(prediction_network.state_dict(), os.path.join(save_dir, f'prediction_network_epoch_{epoch}.pt'))
|
119 |
+
torch.save(action_encoder.state_dict(), os.path.join(save_dir, f'action_encoder_epoch_{epoch}.pt'))
|
120 |
+
|
121 |
+
print(f"All models saved for epoch {epoch}.")
|
122 |
+
|
123 |
+
|
124 |
+
class RotaryPositionalEncoding(nn.Module):
|
125 |
+
def __init__(self, d_model):
|
126 |
+
super(RotaryPositionalEncoding, self).__init__()
|
127 |
+
inv_freq = 1.0 / (10000 ** (torch.arange(0, d_model, 2).float() / d_model))
|
128 |
+
self.register_buffer('inv_freq', inv_freq)
|
129 |
+
|
130 |
+
def forward(self, x):
|
131 |
+
seq_len, batch_size, _ = x.size()
|
132 |
+
t = torch.arange(seq_len, device=x.device).type_as(self.inv_freq)
|
133 |
+
sinusoid_inp = torch.einsum("i,j->ij", t, self.inv_freq)
|
134 |
+
sin = sinusoid_inp.sin().unsqueeze(1) # (seq_len, 1, d_model/2)
|
135 |
+
cos = sinusoid_inp.cos().unsqueeze(1) # (seq_len, 1, d_model/2)
|
136 |
+
|
137 |
+
x1 = x[..., 0::2]
|
138 |
+
x2 = x[..., 1::2]
|
139 |
+
|
140 |
+
# Apply rotation
|
141 |
+
x_rotated = torch.zeros_like(x)
|
142 |
+
x_rotated[..., 0::2] = x1 * cos - x2 * sin
|
143 |
+
x_rotated[..., 1::2] = x1 * sin + x2 * cos
|
144 |
+
|
145 |
+
return x_rotated
|
146 |
+
|
147 |
+
|
148 |
+
class MultiHeadAttention(nn.Module):
|
149 |
+
def __init__(self, d_model, num_heads):
|
150 |
+
super(MultiHeadAttention, self).__init__()
|
151 |
+
assert d_model % num_heads == 0, "d_model must be divisible by num_heads"
|
152 |
+
self.d_k = d_model // num_heads
|
153 |
+
self.num_heads = num_heads
|
154 |
+
self.linear_q = nn.Linear(d_model, d_model)
|
155 |
+
self.linear_k = nn.Linear(d_model, d_model)
|
156 |
+
self.linear_v = nn.Linear(d_model, d_model)
|
157 |
+
self.linear_out = nn.Linear(d_model, d_model)
|
158 |
+
|
159 |
+
def forward(self, query, key, value, mask=None):
|
160 |
+
batch_size = query.size(0)
|
161 |
+
query = self.linear_q(query).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)
|
162 |
+
key = self.linear_k(key).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)
|
163 |
+
value = self.linear_v(value).view(batch_size, -1, self.num_heads, self.d_k).transpose(1, 2)
|
164 |
+
|
165 |
+
scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(self.d_k)
|
166 |
+
if mask is not None:
|
167 |
+
scores = scores.masked_fill(mask == 0, -1e4)
|
168 |
+
attn = F.softmax(scores, dim=-1)
|
169 |
+
output = torch.matmul(attn, value)
|
170 |
+
|
171 |
+
output = output.transpose(1, 2).contiguous().view(batch_size, -1, self.num_heads * self.d_k)
|
172 |
+
return self.linear_out(output)
|
173 |
+
|
174 |
+
|
175 |
+
class MoE(nn.Module):
|
176 |
+
def __init__(self, d_model, num_experts, d_ff, top_k=2, dropout=0.1):
|
177 |
+
super(MoE, self).__init__()
|
178 |
+
self.num_experts = num_experts
|
179 |
+
self.top_k = top_k
|
180 |
+
self.experts = nn.ModuleList([
|
181 |
+
nn.Sequential(
|
182 |
+
nn.Linear(d_model, d_ff),
|
183 |
+
nn.GELU() if i % 2 == 0 else nn.SiLU(),
|
184 |
+
nn.Linear(d_ff, d_model)
|
185 |
+
)
|
186 |
+
for i in range(num_experts)
|
187 |
+
])
|
188 |
+
self.gate = nn.Linear(d_model, num_experts)
|
189 |
+
self.dropout = nn.Dropout(dropout)
|
190 |
+
|
191 |
+
def forward(self, x):
|
192 |
+
batch_size, seq_len, d_model = x.size()
|
193 |
+
# Compute gating scores
|
194 |
+
gate_scores = self.gate(x) # (batch_size, seq_len, num_experts)
|
195 |
+
top_k_scores, top_k_indices = torch.topk(gate_scores, self.top_k, dim=-1) # (batch_size, seq_len, top_k)
|
196 |
+
top_k_scores = F.softmax(top_k_scores, dim=-1) # (batch_size, seq_len, top_k)
|
197 |
+
|
198 |
+
# Initialize output
|
199 |
+
output = torch.zeros_like(x)
|
200 |
+
|
201 |
+
# Flatten batch and sequence dimensions
|
202 |
+
x_flat = x.view(-1, d_model) # (batch_size * seq_len, d_model)
|
203 |
+
output_flat = output.view(-1, d_model)
|
204 |
+
top_k_indices_flat = top_k_indices.view(-1, self.top_k) # (batch_size * seq_len, top_k)
|
205 |
+
top_k_scores_flat = top_k_scores.view(-1, self.top_k) # (batch_size * seq_len, top_k)
|
206 |
+
|
207 |
+
for k in range(self.top_k):
|
208 |
+
expert_idx_flat = top_k_indices_flat[:, k] # (batch_size * seq_len)
|
209 |
+
expert_scores_flat = top_k_scores_flat[:, k] # (batch_size * seq_len)
|
210 |
+
for e in range(self.num_experts):
|
211 |
+
mask = (expert_idx_flat == e) # Boolean mask
|
212 |
+
if mask.any():
|
213 |
+
x_masked = x_flat[mask] # Select tokens for expert e
|
214 |
+
expert_output = self.experts[e](x_masked) # Apply expert e
|
215 |
+
output_flat[mask] += expert_scores_flat[mask].unsqueeze(-1) * expert_output
|
216 |
+
|
217 |
+
output = output_flat.view(batch_size, seq_len, d_model)
|
218 |
+
return self.dropout(output)
|
219 |
+
|
220 |
+
|
221 |
+
class TransformerBlock(nn.Module):
|
222 |
+
def __init__(self, d_model, num_heads, d_ff, num_experts, dropout=0.1, top_k=2):
|
223 |
+
super(TransformerBlock, self).__init__()
|
224 |
+
self.self_attention = MultiHeadAttention(d_model, num_heads)
|
225 |
+
self.norm1 = nn.LayerNorm(d_model)
|
226 |
+
self.cross_attention = MultiHeadAttention(d_model, num_heads)
|
227 |
+
self.norm2 = nn.LayerNorm(d_model)
|
228 |
+
self.moe = MoE(d_model, num_experts, d_ff, top_k, dropout)
|
229 |
+
self.norm3 = nn.LayerNorm(d_model)
|
230 |
+
|
231 |
+
def forward(self, x, mask=None, enc_output=None, enc_mask=None):
|
232 |
+
# Self-attention
|
233 |
+
attn_output = self.self_attention(x, x, x, mask)
|
234 |
+
x = self.norm1(x + attn_output)
|
235 |
+
# Cross-attention (only in decoder)
|
236 |
+
if enc_output is not None:
|
237 |
+
cross_attn_output = self.cross_attention(x, enc_output, enc_output, enc_mask)
|
238 |
+
x = self.norm2(x + cross_attn_output)
|
239 |
+
# Feedforward/MoE
|
240 |
+
moe_output = self.moe(x)
|
241 |
+
return self.norm3(x + moe_output)
|
242 |
+
|
243 |
+
|
244 |
+
class Transformer(nn.Module):
|
245 |
+
def __init__(self, input_dim, d_model, num_heads, num_layers, d_ff, num_experts, output_dim, dropout=0.1, top_k=2):
|
246 |
+
super(Transformer, self).__init__()
|
247 |
+
self.embedding = nn.Embedding(input_dim, d_model, padding_idx=input_dim - 1)
|
248 |
+
self.rotary_positional_encoding = RotaryPositionalEncoding(d_model)
|
249 |
+
self.encoder_layers = nn.ModuleList(
|
250 |
+
[TransformerBlock(d_model, num_heads, d_ff, num_experts, dropout, top_k) for _ in range(num_layers)]
|
251 |
+
)
|
252 |
+
self.decoder_layers = nn.ModuleList(
|
253 |
+
[TransformerBlock(d_model, num_heads, d_ff, num_experts, dropout, top_k) for _ in range(num_layers)]
|
254 |
+
)
|
255 |
+
self.output_layer = nn.Linear(d_model, output_dim)
|
256 |
+
self.d_model = d_model
|
257 |
+
|
258 |
+
def forward(self, src, tgt, src_mask=None, tgt_mask=None):
|
259 |
+
# Encoder
|
260 |
+
src = self.embedding(src) * math.sqrt(self.d_model)
|
261 |
+
src = src.transpose(0, 1) # (batch_size, seq_len, d_model) -> (seq_len, batch_size, d_model)
|
262 |
+
src = self.rotary_positional_encoding(src)
|
263 |
+
src = src.transpose(0, 1) # (seq_len, batch_size, d_model) -> (batch_size, seq_len, d_model)
|
264 |
+
for layer in self.encoder_layers:
|
265 |
+
src = layer(src, src_mask)
|
266 |
+
|
267 |
+
# Decoder
|
268 |
+
tgt = self.embedding(tgt) * math.sqrt(self.d_model)
|
269 |
+
tgt = tgt.transpose(0, 1)
|
270 |
+
tgt = self.rotary_positional_encoding(tgt)
|
271 |
+
tgt = tgt.transpose(0, 1)
|
272 |
+
for layer in self.decoder_layers:
|
273 |
+
tgt = layer(tgt, tgt_mask, src, src_mask)
|
274 |
+
output = self.output_layer(tgt)
|
275 |
+
return output
|
276 |
+
|
277 |
+
def generate(self, src, tokenizer, max_length=20, temperature=1.0):
|
278 |
+
"""
|
279 |
+
Generate sequences using differentiable sampling (Gumbel-Softmax).
|
280 |
+
|
281 |
+
Args:
|
282 |
+
src (torch.Tensor): Source input tensor of shape (batch_size, seq_len)
|
283 |
+
tokenizer (transformers.PreTrainedTokenizer): Tokenizer to access special tokens
|
284 |
+
max_length (int): Maximum length of the generated sequence
|
285 |
+
temperature (float): Temperature parameter for Gumbel-Softmax
|
286 |
+
|
287 |
+
Returns:
|
288 |
+
torch.Tensor: Generated sequences of shape (batch_size, max_length)
|
289 |
+
torch.Tensor: Entropy values for each time step
|
290 |
+
torch.Tensor: Variance values for each time step
|
291 |
+
"""
|
292 |
+
batch_size = src.size(0)
|
293 |
+
|
294 |
+
# Encode the source
|
295 |
+
src_enc = self.embedding(src) * math.sqrt(self.d_model)
|
296 |
+
src_enc = src_enc.transpose(0, 1)
|
297 |
+
src_enc = self.rotary_positional_encoding(src_enc)
|
298 |
+
src_enc = src_enc.transpose(0, 1)
|
299 |
+
for layer in self.encoder_layers:
|
300 |
+
src_enc = layer(src_enc)
|
301 |
+
|
302 |
+
# Initialize decoder input with <sos> tokens
|
303 |
+
tgt_seq = torch.full((batch_size, 1), tokenizer.bos_token_id, dtype=torch.long, device=src.device)
|
304 |
+
entropies = []
|
305 |
+
variances = []
|
306 |
+
|
307 |
+
for _ in range(max_length):
|
308 |
+
tgt_emb = self.embedding(tgt_seq) * math.sqrt(self.d_model)
|
309 |
+
tgt_emb = tgt_emb.transpose(0, 1)
|
310 |
+
tgt_emb = self.rotary_positional_encoding(tgt_emb)
|
311 |
+
tgt_emb = tgt_emb.transpose(0, 1)
|
312 |
+
tgt_dec = tgt_emb
|
313 |
+
for layer in self.decoder_layers:
|
314 |
+
tgt_dec = layer(tgt_dec, None, src_enc, None)
|
315 |
+
output = self.output_layer(tgt_dec) # (batch_size, seq_len, vocab_size)
|
316 |
+
logits = output[:, -1, :] # Get logits for the last time step
|
317 |
+
|
318 |
+
# Compute token probabilities
|
319 |
+
probs = F.softmax(logits / temperature, dim=-1) # (batch_size, vocab_size)
|
320 |
+
|
321 |
+
# Compute entropy
|
322 |
+
entropy = -torch.sum(probs * torch.log(probs + 1e-9), dim=-1) # (batch_size)
|
323 |
+
entropies.append(entropy)
|
324 |
+
|
325 |
+
# Sample token using Gumbel-Softmax
|
326 |
+
gumbel_noise = -torch.log(-torch.log(torch.rand_like(probs) + 1e-9) + 1e-9)
|
327 |
+
y = (logits + gumbel_noise) / temperature
|
328 |
+
y = F.softmax(y, dim=-1) # (batch_size, vocab_size)
|
329 |
+
|
330 |
+
# Compute variance
|
331 |
+
variance = torch.var(y, dim=-1) # (batch_size)
|
332 |
+
variances.append(variance)
|
333 |
+
|
334 |
+
# Get token indices (argmax for hard selection)
|
335 |
+
next_tokens = torch.argmax(y, dim=-1, keepdim=True) # (batch_size, 1)
|
336 |
+
tgt_seq = torch.cat([tgt_seq, next_tokens], dim=1)
|
337 |
+
|
338 |
+
# Stack entropies and variances
|
339 |
+
entropies = torch.stack(entropies, dim=1) # (batch_size, max_length)
|
340 |
+
variances = torch.stack(variances, dim=1) # (batch_size, max_length)
|
341 |
+
|
342 |
+
return tgt_seq[:, 1:], entropies, variances # Exclude the initial <sos> token
|
343 |
+
|
344 |
+
|
345 |
+
# Objective Functions
|
346 |
+
|
347 |
+
class InfoNCE_Loss(nn.Module):
|
348 |
+
def __init__(self, temperature=0.07):
|
349 |
+
super(InfoNCE_Loss, self).__init__()
|
350 |
+
self.temperature = temperature
|
351 |
+
self.cross_entropy = nn.CrossEntropyLoss()
|
352 |
+
|
353 |
+
def forward(self, z_i, z_j):
|
354 |
+
"""
|
355 |
+
Args:
|
356 |
+
z_i (torch.Tensor): Flattened representations from view i, shape (2n, embed_dim)
|
357 |
+
z_j (torch.Tensor): Flattened representations from view j, shape (2n, embed_dim)
|
358 |
+
|
359 |
+
Returns:
|
360 |
+
torch.Tensor: InfoNCE loss
|
361 |
+
"""
|
362 |
+
n = z_i.size(0)
|
363 |
+
z = torch.cat([z_i, z_j], dim=0) # Shape: (2n, embed_dim)
|
364 |
+
|
365 |
+
z = F.normalize(z, dim=1)
|
366 |
+
similarity_matrix = torch.matmul(z, z.T) # Shape: (2n, 2n)
|
367 |
+
|
368 |
+
# Create a mask to exclude self-similarity
|
369 |
+
mask = torch.eye(2 * n, device=z.device, dtype=torch.bool)
|
370 |
+
similarity_matrix = similarity_matrix.masked_fill(mask, -1e4) # Use a manageable negative value
|
371 |
+
|
372 |
+
# Create labels for contrastive learning
|
373 |
+
labels = torch.arange(n, device=z.device)
|
374 |
+
labels = torch.cat([labels + n, labels], dim=0) # Shape: (2n,)
|
375 |
+
|
376 |
+
# Apply temperature scaling
|
377 |
+
similarity_matrix /= self.temperature
|
378 |
+
|
379 |
+
# Compute cross-entropy loss
|
380 |
+
loss = self.cross_entropy(similarity_matrix, labels)
|
381 |
+
return loss
|
382 |
+
|
383 |
+
|
384 |
+
|
385 |
+
class CovarianceRegularization(nn.Module):
|
386 |
+
def __init__(self, lambda_reg=1e-3):
|
387 |
+
super(CovarianceRegularization, self).__init__()
|
388 |
+
self.lambda_reg = lambda_reg
|
389 |
+
|
390 |
+
def forward(self, embeddings):
|
391 |
+
"""
|
392 |
+
Args:
|
393 |
+
embeddings (torch.Tensor): Embedding tensor, shape (batch_size, embed_dim)
|
394 |
+
|
395 |
+
Returns:
|
396 |
+
torch.Tensor: Covariance regularization loss
|
397 |
+
"""
|
398 |
+
batch_size, embed_dim = embeddings.size()
|
399 |
+
mean = embeddings.mean(dim=0)
|
400 |
+
embeddings_centered = embeddings - mean
|
401 |
+
cov = (embeddings_centered.T @ embeddings_centered) / (batch_size - 1)
|
402 |
+
cov_loss = torch.sum(cov ** 2) - torch.sum(torch.diag(cov) ** 2)
|
403 |
+
return self.lambda_reg * cov_loss
|
404 |
+
|
405 |
+
|
406 |
+
class DynamicsPerformanceLoss(nn.Module):
|
407 |
+
def __init__(self, lambda_var=1e-3):
|
408 |
+
super(DynamicsPerformanceLoss, self).__init__()
|
409 |
+
self.lambda_var = lambda_var
|
410 |
+
|
411 |
+
def forward(self, true_next_state, predicted_next_state):
|
412 |
+
"""
|
413 |
+
Args:
|
414 |
+
true_next_state (torch.Tensor): Ground truth next state, shape (batch_size, state_dim)
|
415 |
+
predicted_next_state (torch.Tensor): Predicted next state, shape (batch_size, state_dim)
|
416 |
+
|
417 |
+
Returns:
|
418 |
+
torch.Tensor: Dynamics performance loss
|
419 |
+
"""
|
420 |
+
mse_loss = F.mse_loss(predicted_next_state, true_next_state)
|
421 |
+
variance_loss = torch.var(predicted_next_state, dim=0).mean()
|
422 |
+
return mse_loss + self.lambda_var * variance_loss
|
423 |
+
|
424 |
+
|
425 |
+
class ThoughtConsistencyLoss(nn.Module):
|
426 |
+
def __init__(self):
|
427 |
+
super(ThoughtConsistencyLoss, self).__init__()
|
428 |
+
|
429 |
+
def forward(self, true_next_state, perturbed_next_state):
|
430 |
+
"""
|
431 |
+
Args:
|
432 |
+
true_next_state (torch.Tensor): Ground truth next state, shape (batch_size, state_dim)
|
433 |
+
perturbed_next_state (torch.Tensor): Perturbed next state, shape (batch_size, state_dim)
|
434 |
+
|
435 |
+
Returns:
|
436 |
+
torch.Tensor: Thought-consistency loss
|
437 |
+
"""
|
438 |
+
return F.mse_loss(true_next_state, perturbed_next_state)
|
439 |
+
|
440 |
+
|
441 |
+
class PolicyValueJointLoss(nn.Module):
|
442 |
+
def __init__(self, lambda_value=0.5):
|
443 |
+
super(PolicyValueJointLoss, self).__init__()
|
444 |
+
self.lambda_value = lambda_value
|
445 |
+
self.cross_entropy = nn.CrossEntropyLoss()
|
446 |
+
self.mse_loss = nn.MSELoss()
|
447 |
+
|
448 |
+
def forward(self, policy_logits, true_policy, value_pred, true_value):
|
449 |
+
"""
|
450 |
+
Args:
|
451 |
+
policy_logits (torch.Tensor): Logits from the policy network, shape (batch_size * seq_len, num_actions)
|
452 |
+
true_policy (torch.Tensor): Ground truth policy, shape (batch_size * seq_len, num_actions)
|
453 |
+
value_pred (torch.Tensor): Predicted values, shape (batch_size * seq_len)
|
454 |
+
true_value (torch.Tensor): Ground truth values, shape (batch_size * seq_len)
|
455 |
+
|
456 |
+
Returns:
|
457 |
+
torch.Tensor: Combined policy and value loss
|
458 |
+
"""
|
459 |
+
policy_logits = policy_logits.view(-1, policy_logits.size(-1))
|
460 |
+
true_policy = true_policy.view(-1, true_policy.size(-1))
|
461 |
+
value_pred = value_pred.view(-1)
|
462 |
+
true_value = true_value.view(-1)
|
463 |
+
|
464 |
+
policy_loss = self.cross_entropy(policy_logits, true_policy.argmax(dim=1))
|
465 |
+
value_loss = self.mse_loss(value_pred, true_value)
|
466 |
+
return policy_loss + self.lambda_value * value_loss
|
467 |
+
|
468 |
+
|
469 |
+
|
470 |
+
class ActionDiversityReward(nn.Module):
|
471 |
+
def __init__(self, lambda_div=1e-3):
|
472 |
+
super(ActionDiversityReward, self).__init__()
|
473 |
+
self.lambda_div = lambda_div
|
474 |
+
|
475 |
+
def forward(self, action_embeddings):
|
476 |
+
"""
|
477 |
+
Args:
|
478 |
+
action_embeddings (torch.Tensor): Embeddings of actions, shape (batch_size, embed_dim)
|
479 |
+
|
480 |
+
Returns:
|
481 |
+
torch.Tensor: Action diversity loss
|
482 |
+
"""
|
483 |
+
similarity_matrix = F.cosine_similarity(action_embeddings.unsqueeze(1), action_embeddings.unsqueeze(0), dim=2)
|
484 |
+
# Zero out self-similarity
|
485 |
+
similarity_matrix = similarity_matrix - torch.eye(similarity_matrix.size(0)).to(action_embeddings.device)
|
486 |
+
diversity_loss = torch.sum(similarity_matrix ** 2)
|
487 |
+
return self.lambda_div * diversity_loss
|
488 |
+
|
489 |
+
|
490 |
+
class ExpectedThoughtValueLoss(nn.Module):
|
491 |
+
def __init__(self):
|
492 |
+
super(ExpectedThoughtValueLoss, self).__init__()
|
493 |
+
|
494 |
+
def forward(self, mcts_best_values):
|
495 |
+
"""
|
496 |
+
Args:
|
497 |
+
mcts_best_values (torch.Tensor): Best values from MCTS, shape (batch_size)
|
498 |
+
|
499 |
+
Returns:
|
500 |
+
torch.Tensor: ETV loss
|
501 |
+
"""
|
502 |
+
return -mcts_best_values.mean()
|
503 |
+
|
504 |
+
|
505 |
+
class ExplorationRegularization(nn.Module):
|
506 |
+
def __init__(self, lambda_expl=1e-3):
|
507 |
+
super(ExplorationRegularization, self).__init__()
|
508 |
+
self.lambda_expl = lambda_expl
|
509 |
+
|
510 |
+
def forward(self, visit_counts):
|
511 |
+
"""
|
512 |
+
Args:
|
513 |
+
visit_counts (torch.Tensor): Visit counts for actions, shape (batch_size, num_actions)
|
514 |
+
|
515 |
+
Returns:
|
516 |
+
torch.Tensor: Exploration regularization loss
|
517 |
+
"""
|
518 |
+
reward = torch.sum(1.0 / (visit_counts + 1), dim=-1)
|
519 |
+
return self.lambda_expl * reward.mean()
|
520 |
+
|
521 |
+
|
522 |
+
class KL_DivergenceLoss(nn.Module):
|
523 |
+
def __init__(self):
|
524 |
+
super(KL_DivergenceLoss, self).__init__()
|
525 |
+
|
526 |
+
def forward(self, old_policy, new_policy):
|
527 |
+
"""
|
528 |
+
Args:
|
529 |
+
old_policy (torch.Tensor): Old policy probabilities, shape (batch_size, num_actions)
|
530 |
+
new_policy (torch.Tensor): New policy probabilities, shape (batch_size, num_actions)
|
531 |
+
|
532 |
+
Returns:
|
533 |
+
torch.Tensor: KL divergence loss
|
534 |
+
"""
|
535 |
+
kl_div = F.kl_div(new_policy.log(), old_policy, reduction='batchmean')
|
536 |
+
return kl_div
|
537 |
+
|
538 |
+
# MuZero
|
539 |
+
|
540 |
+
class ActionEncoder(nn.Module):
|
541 |
+
def __init__(self, vocab_size, embed_dim):
|
542 |
+
super(ActionEncoder, self).__init__()
|
543 |
+
self.embedding = nn.Embedding(vocab_size, embed_dim)
|
544 |
+
|
545 |
+
def forward(self, action_sequences):
|
546 |
+
"""
|
547 |
+
Args:
|
548 |
+
action_sequences (torch.Tensor): Tensor of shape (batch_size, seq_len)
|
549 |
+
|
550 |
+
Returns:
|
551 |
+
torch.Tensor: Encoded actions of shape (batch_size, seq_len, embed_dim)
|
552 |
+
"""
|
553 |
+
return self.embedding(action_sequences) #.half() # Convert to half-precision
|
554 |
+
|
555 |
+
class RepresentationNetwork(nn.Module):
|
556 |
+
def __init__(self, vocab_dim, d_model, state_dim):
|
557 |
+
super(RepresentationNetwork, self).__init__()
|
558 |
+
self.proj = nn.Linear(vocab_dim, d_model) # Project from vocab_dim to d_model
|
559 |
+
self.linear = nn.Linear(d_model, state_dim) # Project from d_model to state_dim
|
560 |
+
self.norm = nn.LayerNorm(state_dim)
|
561 |
+
|
562 |
+
def forward(self, transformer_output):
|
563 |
+
"""
|
564 |
+
Args:
|
565 |
+
transformer_output (torch.Tensor): Shape (batch_size, seq_len, vocab_dim)
|
566 |
+
|
567 |
+
Returns:
|
568 |
+
torch.Tensor: Encoded state of shape (batch_size, seq_len, state_dim)
|
569 |
+
"""
|
570 |
+
# First project down from vocab_dim to d_model
|
571 |
+
projected_output = self.proj(transformer_output)
|
572 |
+
# Then project down from d_model to state_dim
|
573 |
+
state = self.linear(projected_output)
|
574 |
+
state = self.norm(state)
|
575 |
+
return state
|
576 |
+
|
577 |
+
|
578 |
+
class DynamicsNetwork(nn.Module):
|
579 |
+
def __init__(self, state_dim, action_dim, hidden_dim):
|
580 |
+
super(DynamicsNetwork, self).__init__()
|
581 |
+
self.rms_norm = nn.LayerNorm(state_dim)
|
582 |
+
self.fc1 = nn.Linear(state_dim + action_dim, hidden_dim)
|
583 |
+
self.activation = nn.GELU()
|
584 |
+
self.fc2 = nn.Linear(hidden_dim, state_dim)
|
585 |
+
|
586 |
+
def forward(self, state, action):
|
587 |
+
"""
|
588 |
+
Args:
|
589 |
+
state (torch.Tensor): Current state, shape (batch_size, seq_len, state_dim)
|
590 |
+
action (torch.Tensor): Action embedding, shape (batch_size, seq_len, action_dim)
|
591 |
+
|
592 |
+
Returns:
|
593 |
+
torch.Tensor: Predicted next state, shape (batch_size, seq_len, state_dim)
|
594 |
+
"""
|
595 |
+
norm_state = self.rms_norm(state)
|
596 |
+
combined = torch.cat([norm_state, action], dim=-1)
|
597 |
+
hidden = self.activation(self.fc1(combined))
|
598 |
+
next_state = self.fc2(hidden)
|
599 |
+
return next_state
|
600 |
+
|
601 |
+
class PredictionNetwork(nn.Module):
|
602 |
+
def __init__(self, state_dim, policy_dim, value_dim):
|
603 |
+
super(PredictionNetwork, self).__init__()
|
604 |
+
self.state_dim = state_dim
|
605 |
+
self.rms_norm = nn.LayerNorm(state_dim)
|
606 |
+
self.policy_head = nn.Linear(state_dim, policy_dim)
|
607 |
+
self.value_head = nn.Linear(state_dim, value_dim)
|
608 |
+
|
609 |
+
def forward(self, state):
|
610 |
+
"""
|
611 |
+
Args:
|
612 |
+
state (torch.Tensor): Predicted state, shape (batch_size, seq_len, state_dim)
|
613 |
+
|
614 |
+
Returns:
|
615 |
+
Tuple[torch.Tensor, torch.Tensor]: Policy logits and value estimates
|
616 |
+
"""
|
617 |
+
norm_state = self.rms_norm(state)
|
618 |
+
policy_logits = self.policy_head(norm_state)
|
619 |
+
value_estimates = self.value_head(norm_state)
|
620 |
+
return policy_logits, value_estimates
|
621 |
+
|
622 |
+
|
623 |
+
class MCTSNode:
|
624 |
+
def __init__(self, state, parent=None, action=None):
|
625 |
+
"""
|
626 |
+
Initialize an MCTS node.
|
627 |
+
|
628 |
+
Args:
|
629 |
+
state (State): The current state representation.
|
630 |
+
parent (MCTSNode, optional): The parent node. Defaults to None.
|
631 |
+
action (int, optional): The action taken to reach this node. Defaults to None.
|
632 |
+
"""
|
633 |
+
self.state = state # Instance of State class
|
634 |
+
self.parent = parent # Parent MCTSNode
|
635 |
+
self.action = action # Action taken to reach this node
|
636 |
+
self.children = {} # Dict mapping actions to MCTSNode
|
637 |
+
self.visit_count = 0
|
638 |
+
self.value_sum = 0.0
|
639 |
+
self.prior = 0.0 # Prior probability from policy network
|
640 |
+
|
641 |
+
def expand(self, actions, priors):
|
642 |
+
"""
|
643 |
+
Expand the node with possible actions and their priors.
|
644 |
+
|
645 |
+
Args:
|
646 |
+
actions (list): List of possible actions (action indices).
|
647 |
+
priors (list): List of prior probabilities corresponding to actions.
|
648 |
+
"""
|
649 |
+
for action, prior in zip(actions, priors):
|
650 |
+
if action not in self.children:
|
651 |
+
child_state = self.state.apply_action(action) # Apply action to get new state
|
652 |
+
child_node = MCTSNode(state=child_state, parent=self, action=action)
|
653 |
+
child_node.prior = float(prior) # Ensure that prior is a float value
|
654 |
+
self.children[action] = child_node
|
655 |
+
|
656 |
+
def is_leaf(self):
|
657 |
+
"""
|
658 |
+
Check if the node is a leaf node (i.e., has no children).
|
659 |
+
|
660 |
+
Returns:
|
661 |
+
bool: True if leaf, False otherwise.
|
662 |
+
"""
|
663 |
+
return len(self.children) == 0
|
664 |
+
|
665 |
+
def ucb_score(self, total_visits, exploration_constant=math.sqrt(2)):
|
666 |
+
"""
|
667 |
+
Calculate the UCB (Upper Confidence Bound) score for the node.
|
668 |
+
|
669 |
+
Args:
|
670 |
+
total_visits (int): Total number of visits to the parent node.
|
671 |
+
exploration_constant (float, optional): Exploration parameter. Defaults to math.sqrt(2).
|
672 |
+
|
673 |
+
Returns:
|
674 |
+
float: The UCB score.
|
675 |
+
"""
|
676 |
+
if self.visit_count == 0:
|
677 |
+
return float('inf')
|
678 |
+
average_value = self.value_sum / self.visit_count
|
679 |
+
exploration_term = exploration_constant * self.prior * math.sqrt(total_visits) / (1 + self.visit_count)
|
680 |
+
return average_value + exploration_term
|
681 |
+
|
682 |
+
class MCTS:
|
683 |
+
def __init__(self, prediction_network, dynamics_network, action_encoder, num_iterations=10, exploration_constant=math.sqrt(2)):
|
684 |
+
"""
|
685 |
+
Initialize the MCTS.
|
686 |
+
|
687 |
+
Args:
|
688 |
+
prediction_network (nn.Module): The Prediction Network.
|
689 |
+
dynamics_network (nn.Module): The Dynamics Network.
|
690 |
+
num_iterations (int): Number of MCTS iterations per search.
|
691 |
+
exploration_constant (float): Exploration parameter for UCB.
|
692 |
+
"""
|
693 |
+
self.action_encoder = action_encoder
|
694 |
+
self.prediction_network = prediction_network
|
695 |
+
self.dynamics_network = dynamics_network
|
696 |
+
self.num_iterations = num_iterations
|
697 |
+
self.exploration_constant = exploration_constant
|
698 |
+
|
699 |
+
def search(self, root_state):
|
700 |
+
"""
|
701 |
+
Perform MCTS starting from the root_state.
|
702 |
+
|
703 |
+
Args:
|
704 |
+
root_state: The initial state from which to start MCTS.
|
705 |
+
|
706 |
+
Returns:
|
707 |
+
The best action determined by MCTS.
|
708 |
+
"""
|
709 |
+
self.root = MCTSNode(state=root_state)
|
710 |
+
|
711 |
+
for _ in range(self.num_iterations):
|
712 |
+
node = self.select(self.root)
|
713 |
+
value = self.evaluate(node)
|
714 |
+
self.backpropagate(node, value)
|
715 |
+
|
716 |
+
return self.best_action()
|
717 |
+
|
718 |
+
def select(self, node):
|
719 |
+
"""
|
720 |
+
Traverse the tree to select a node for evaluation.
|
721 |
+
|
722 |
+
Args:
|
723 |
+
node: The starting node for selection.
|
724 |
+
|
725 |
+
Returns:
|
726 |
+
The node selected for evaluation.
|
727 |
+
"""
|
728 |
+
while not node.is_leaf():
|
729 |
+
best_action, best_node = max(node.children.items(),
|
730 |
+
key=lambda item: item[1].ucb_score(node.visit_count, self.exploration_constant))
|
731 |
+
node = best_node
|
732 |
+
return node
|
733 |
+
|
734 |
+
def evaluate(self, node):
|
735 |
+
"""
|
736 |
+
Evaluate the node by expanding it and predicting its value.
|
737 |
+
|
738 |
+
Args:
|
739 |
+
node: The node to evaluate.
|
740 |
+
|
741 |
+
Returns:
|
742 |
+
The value estimate of the node.
|
743 |
+
"""
|
744 |
+
# Use the prediction network to get policy logits and value estimate
|
745 |
+
policy_logits, value_estimate = self.prediction_network(node.state.representation)
|
746 |
+
|
747 |
+
# Convert logits to probabilities
|
748 |
+
policy = F.softmax(policy_logits, dim=-1).detach().cpu().numpy()
|
749 |
+
|
750 |
+
# Expand the node with possible actions and their priors
|
751 |
+
actions = list(range(policy.shape[-1])) # Assuming actions are indexed from 0 to num_actions-1
|
752 |
+
priors = policy[0].flatten().tolist() # Convert to a 1D list of floats
|
753 |
+
|
754 |
+
node.expand(actions, priors)
|
755 |
+
|
756 |
+
return value_estimate.mean().item()
|
757 |
+
|
758 |
+
|
759 |
+
def backpropagate(self, node, value):
|
760 |
+
"""
|
761 |
+
Backpropagate the value up the tree.
|
762 |
+
|
763 |
+
Args:
|
764 |
+
node: The node to start backpropagation from.
|
765 |
+
value (float): The value to backpropagate.
|
766 |
+
"""
|
767 |
+
while node is not None:
|
768 |
+
node.visit_count += 1
|
769 |
+
node.value_sum += value
|
770 |
+
node = node.parent
|
771 |
+
|
772 |
+
def best_action(self):
|
773 |
+
"""
|
774 |
+
Choose the action with the highest visit count.
|
775 |
+
|
776 |
+
Returns:
|
777 |
+
The best action.
|
778 |
+
"""
|
779 |
+
best_child = max(self.root.children.values(), key=lambda n: n.visit_count)
|
780 |
+
return best_child.action
|
781 |
+
|
782 |
+
class State:
|
783 |
+
def __init__(self, representation, dynamics_network, action_encoder):
|
784 |
+
"""
|
785 |
+
Initialize the State.
|
786 |
+
|
787 |
+
Args:
|
788 |
+
representation (torch.Tensor): Encoded state representation, shape (batch_size, seq_len, state_dim)
|
789 |
+
dynamics_network (nn.Module): The Dynamics Network to predict next states
|
790 |
+
action_encoder (nn.Module): The Action Encoder to encode actions
|
791 |
+
"""
|
792 |
+
self.representation = representation # Shape: (batch_size, seq_len, state_dim)
|
793 |
+
self.dynamics_network = dynamics_network # Reference to Dynamics Network
|
794 |
+
self.action_encoder = action_encoder
|
795 |
+
|
796 |
+
def apply_action(self, action):
|
797 |
+
"""
|
798 |
+
Apply an action to the current state to get a new state.
|
799 |
+
|
800 |
+
Args:
|
801 |
+
action (int): The action to apply (e.g., token index)
|
802 |
+
|
803 |
+
Returns:
|
804 |
+
State: The new state after applying the action
|
805 |
+
"""
|
806 |
+
# Create action sequence filled with action index
|
807 |
+
batch_size, seq_len, _ = self.representation.size()
|
808 |
+
action_sequence = torch.full((batch_size, seq_len), action, dtype=torch.long, device=self.representation.device)
|
809 |
+
# Encode action
|
810 |
+
action_embedding = self.action_encoder(action_sequence)
|
811 |
+
# Predict the next state using the Dynamics Network
|
812 |
+
with torch.no_grad():
|
813 |
+
next_state_representation = self.dynamics_network(self.representation, action_embedding)
|
814 |
+
return State(next_state_representation, self.dynamics_network, self.action_encoder)
|
815 |
+
|
816 |
+
|
817 |
+
|
818 |
+
|
819 |
+
class PPOAgent:
|
820 |
+
def __init__(self, policy_network, optimizer, clip_epsilon=0.2, entropy_coef=0.01, value_coef=0.5):
|
821 |
+
self.policy_network = policy_network
|
822 |
+
self.optimizer = optimizer
|
823 |
+
self.clip_epsilon = clip_epsilon
|
824 |
+
self.entropy_coef = entropy_coef
|
825 |
+
self.value_coef = value_coef
|
826 |
+
|
827 |
+
def compute_loss(self, states, old_log_probs, actions, returns, advantages):
|
828 |
+
# Get policy logits and value estimates
|
829 |
+
policy_logits, value_estimates = self.policy_network(states)
|
830 |
+
batch_size, seq_len, num_actions = policy_logits.size()
|
831 |
+
|
832 |
+
# Flatten tensors
|
833 |
+
policy_logits = policy_logits.view(-1, num_actions) # Shape: (batch_size * seq_len, num_actions)
|
834 |
+
value_estimates = value_estimates.view(-1) # Shape: (batch_size * seq_len)
|
835 |
+
actions = actions.view(-1) # Shape: (batch_size * seq_len)
|
836 |
+
old_log_probs = old_log_probs.view(-1) # Shape: (batch_size * seq_len)
|
837 |
+
returns = returns.view(-1) # Shape: (batch_size * seq_len)
|
838 |
+
advantages = advantages.view(-1) # Shape: (batch_size * seq_len)
|
839 |
+
|
840 |
+
# Compute new log probabilities
|
841 |
+
new_log_probs_all = F.log_softmax(policy_logits, dim=-1) # Shape: (batch_size * seq_len, num_actions)
|
842 |
+
new_log_probs = new_log_probs_all.gather(1, actions.unsqueeze(-1)).squeeze(-1) # Shape: (batch_size * seq_len)
|
843 |
+
|
844 |
+
# Compute ratios
|
845 |
+
ratios = torch.exp(new_log_probs - old_log_probs)
|
846 |
+
|
847 |
+
# PPO surrogate loss
|
848 |
+
surr1 = ratios * advantages
|
849 |
+
surr2 = torch.clamp(ratios, 1 - self.clip_epsilon, 1 + self.clip_epsilon) * advantages
|
850 |
+
policy_loss = -torch.min(surr1, surr2).mean()
|
851 |
+
|
852 |
+
# Value loss
|
853 |
+
value_loss = F.mse_loss(value_estimates, returns)
|
854 |
+
|
855 |
+
# Entropy loss
|
856 |
+
entropy = -(new_log_probs * torch.exp(new_log_probs)).mean()
|
857 |
+
|
858 |
+
# Total loss
|
859 |
+
total_loss = policy_loss + self.value_coef * value_loss - self.entropy_coef * entropy
|
860 |
+
return total_loss
|
861 |
+
|
862 |
+
|
863 |
+
|
864 |
+
|
865 |
+
def compute_loss_world_model(predicted_next_state, true_next_state, policy_logits, true_policy, value_estimates, true_value,
|
866 |
+
alpha, beta, temperature, lambda_reg, lambda_var, lambda_div, lambda_expl):
|
867 |
+
"""
|
868 |
+
Compute the combined loss for the World Model.
|
869 |
+
|
870 |
+
Args:
|
871 |
+
predicted_next_state (torch.Tensor): Predicted next state, shape (batch_size, state_dim)
|
872 |
+
true_next_state (torch.Tensor): Ground truth next state, shape (batch_size, state_dim)
|
873 |
+
policy_logits (torch.Tensor): Policy logits, shape (batch_size, num_actions)
|
874 |
+
true_policy (torch.Tensor): Ground truth policy, shape (batch_size, num_actions)
|
875 |
+
value_estimates (torch.Tensor): Value estimates, shape (batch_size)
|
876 |
+
true_value (torch.Tensor): Ground truth value, shape (batch_size)
|
877 |
+
alpha (float): Entropy regularization weight
|
878 |
+
beta (float): Variance regularization weight
|
879 |
+
temperature (float): Temperature parameter
|
880 |
+
lambda_reg (float): Covariance regularization weight
|
881 |
+
lambda_var (float): Dynamics variance loss weight
|
882 |
+
lambda_div (float): Action diversity reward weight
|
883 |
+
lambda_expl (float): Exploration regularization weight
|
884 |
+
|
885 |
+
Returns:
|
886 |
+
torch.Tensor: Combined loss
|
887 |
+
"""
|
888 |
+
# Cross-entropy loss
|
889 |
+
ce_loss = F.cross_entropy(policy_logits, true_policy.argmax(dim=1))
|
890 |
+
|
891 |
+
# Entropy loss
|
892 |
+
probs = F.softmax(policy_logits / temperature, dim=-1) # (batch_size, num_actions)
|
893 |
+
entropy = -torch.sum(probs * torch.log(probs + 1e-9), dim=-1) # (batch_size)
|
894 |
+
entropy_loss = -alpha * torch.mean(entropy)
|
895 |
+
|
896 |
+
# Variance loss
|
897 |
+
variance = torch.var(probs, dim=-1) # (batch_size)
|
898 |
+
variance_loss = -beta * torch.mean(variance)
|
899 |
+
|
900 |
+
# Covariance Regularization
|
901 |
+
cov_reg = CovarianceRegularization(lambda_reg)(predicted_next_state)
|
902 |
+
|
903 |
+
# Dynamics Performance Loss
|
904 |
+
dynamics_loss = DynamicsPerformanceLoss(lambda_var)(true_next_state, predicted_next_state)
|
905 |
+
|
906 |
+
# Thought-Consistency Loss
|
907 |
+
perturbed_next_state = predicted_next_state + torch.randn_like(predicted_next_state) * 0.01
|
908 |
+
thought_loss = ThoughtConsistencyLoss()(true_next_state, perturbed_next_state)
|
909 |
+
|
910 |
+
# Policy-Value Joint Loss
|
911 |
+
pv_loss = PolicyValueJointLoss()(policy_logits, true_policy, value_estimates, true_value)
|
912 |
+
|
913 |
+
# Action Diversity Reward
|
914 |
+
action_embeddings = predicted_next_state # Assuming actions are derived from state
|
915 |
+
action_diversity = ActionDiversityReward(lambda_div)(action_embeddings)
|
916 |
+
|
917 |
+
# Expected Thought Value (ETV) Loss
|
918 |
+
# Placeholder: Replace with actual MCTS best values
|
919 |
+
mcts_best_values = torch.zeros(value_estimates.size(0)).to(device)
|
920 |
+
etv = ExpectedThoughtValueLoss()(mcts_best_values)
|
921 |
+
|
922 |
+
# Exploration Regularization
|
923 |
+
# Placeholder: Replace with actual visit counts
|
924 |
+
visit_counts = torch.ones(predicted_next_state.size(0), input_dim).to(device)
|
925 |
+
exploration = ExplorationRegularization(lambda_expl)(visit_counts)
|
926 |
+
|
927 |
+
# KL Divergence Regularization
|
928 |
+
# Placeholder: Replace with actual old and new policies
|
929 |
+
old_policy = F.softmax(policy_logits.detach(), dim=-1)
|
930 |
+
new_policy = F.softmax(policy_logits, dim=-1)
|
931 |
+
kl_loss = KL_DivergenceLoss()(old_policy, new_policy)
|
932 |
+
|
933 |
+
# Total Loss
|
934 |
+
total_loss = (
|
935 |
+
ce_loss +
|
936 |
+
entropy_loss +
|
937 |
+
variance_loss +
|
938 |
+
cov_reg +
|
939 |
+
dynamics_loss +
|
940 |
+
thought_loss +
|
941 |
+
pv_loss +
|
942 |
+
action_diversity +
|
943 |
+
etv +
|
944 |
+
exploration +
|
945 |
+
kl_loss
|
946 |
+
)
|
947 |
+
|
948 |
+
return total_loss
|
949 |
+
|
950 |
+
|
951 |
+
def train_epoch_world_model(world_model_components, train_loader, optimizer, scheduler, scaler, args, model_transformer, state_dim, embed_dim, input_dim):
|
952 |
+
representation_network, dynamics_network, prediction_network, action_encoder, ppo_agent = world_model_components
|
953 |
+
representation_network.train()
|
954 |
+
dynamics_network.train()
|
955 |
+
prediction_network.train()
|
956 |
+
action_encoder.train()
|
957 |
+
ppo_agent.policy_network.train()
|
958 |
+
|
959 |
+
mcts = MCTS(prediction_network, dynamics_network, action_encoder, num_iterations=args.mcts_iterations, exploration_constant=args.mcts_exploration_constant)
|
960 |
+
|
961 |
+
total_loss = 0.0
|
962 |
+
optimizer.zero_grad()
|
963 |
+
print(f"Starting World Model training epoch with {len(train_loader)} batches...")
|
964 |
+
|
965 |
+
for i, batch in enumerate(train_loader):
|
966 |
+
print(f"Processing batch {i+1}/{len(train_loader)}...")
|
967 |
+
|
968 |
+
# Ensure batches are on the appropriate device for the Transformer
|
969 |
+
src_batch = batch['input_ids'].to('cpu') # Move to CPU for Transformer model
|
970 |
+
tgt_batch = batch['labels'].to('cpu') # Move to CPU for Transformer model
|
971 |
+
|
972 |
+
with autocast(device_type='cuda'):
|
973 |
+
print("Forward pass through Transformer (frozen)...")
|
974 |
+
with torch.no_grad():
|
975 |
+
transformer_output = model_transformer(src_batch, tgt_batch[:, :-1])
|
976 |
+
|
977 |
+
# Move transformer output to the GPU for further processing
|
978 |
+
transformer_output = transformer_output.to(device)
|
979 |
+
|
980 |
+
# Encode actions directly on the GPU
|
981 |
+
encoded_actions = action_encoder(tgt_batch[:, :-1].to(device)) # Move labels to GPU for encoding
|
982 |
+
|
983 |
+
# World Model - Representation
|
984 |
+
state_representation = representation_network(transformer_output) # On GPU
|
985 |
+
|
986 |
+
batch_size, seq_len, _ = state_representation.size()
|
987 |
+
|
988 |
+
# Initialize list to collect predicted next states for the batch
|
989 |
+
predicted_next_states = []
|
990 |
+
|
991 |
+
# Iterate over each sample in the batch for MCTS
|
992 |
+
for b in range(batch_size):
|
993 |
+
# Create a State instance for the current sample
|
994 |
+
current_state = State(state_representation[b].unsqueeze(0), dynamics_network, action_encoder)
|
995 |
+
|
996 |
+
# Perform MCTS to find the best action
|
997 |
+
best_action = mcts.search(current_state)
|
998 |
+
|
999 |
+
# Create action sequence filled with best_action
|
1000 |
+
action_sequence = torch.full((1, seq_len), best_action, dtype=torch.long, device=device)
|
1001 |
+
|
1002 |
+
# Get action embedding
|
1003 |
+
action_embedding = action_encoder(action_sequence)
|
1004 |
+
|
1005 |
+
# Apply dynamics network
|
1006 |
+
predicted_next_state = dynamics_network(current_state.representation, action_embedding)
|
1007 |
+
|
1008 |
+
predicted_next_states.append(predicted_next_state)
|
1009 |
+
|
1010 |
+
# Concatenate predicted next states to form a batch
|
1011 |
+
predicted_next_state_batch = torch.cat(predicted_next_states, dim=0)
|
1012 |
+
|
1013 |
+
# Prediction Network - Policy logits and value
|
1014 |
+
policy_logits, value_estimates = prediction_network(predicted_next_state_batch)
|
1015 |
+
|
1016 |
+
|
1017 |
+
# Define true_policy and true_value as placeholders on the GPU
|
1018 |
+
true_policy = torch.zeros_like(policy_logits).to(device)
|
1019 |
+
true_value = torch.zeros_like(value_estimates).to(device)
|
1020 |
+
|
1021 |
+
# Compute PPO loss
|
1022 |
+
actions = torch.argmax(policy_logits, dim=-1)
|
1023 |
+
old_log_probs = torch.zeros_like(actions, dtype=torch.float32).to(device)
|
1024 |
+
returns = torch.zeros_like(actions, dtype=torch.float32).to(device)
|
1025 |
+
advantages = torch.zeros_like(actions, dtype=torch.float32).to(device)
|
1026 |
+
|
1027 |
+
# Compute PPO loss using states
|
1028 |
+
ppo_loss = ppo_agent.compute_loss(state_representation, old_log_probs, actions, returns, advantages)
|
1029 |
+
|
1030 |
+
# Compute InfoNCE Loss
|
1031 |
+
z_i = state_representation.view(batch_size * seq_len, state_dim) # Shape: (batch_size * seq_len, state_dim)
|
1032 |
+
z_j = F.dropout(z_i, p=0.1, training=True)
|
1033 |
+
info_nce = InfoNCE_Loss()(z_i, z_j)
|
1034 |
+
|
1035 |
+
# Compute other losses
|
1036 |
+
covariance = CovarianceRegularization()(predicted_next_state_batch.view(-1, predicted_next_state_batch.size(-1)))
|
1037 |
+
dynamics_loss = DynamicsPerformanceLoss()(torch.zeros_like(predicted_next_state_batch).to(device), predicted_next_state_batch)
|
1038 |
+
perturbed_next_state = predicted_next_state_batch + torch.randn_like(predicted_next_state_batch) * 0.01
|
1039 |
+
thought_loss = ThoughtConsistencyLoss()(torch.zeros_like(predicted_next_state_batch).to(device), perturbed_next_state)
|
1040 |
+
pv_loss = PolicyValueJointLoss()(policy_logits, true_policy, value_estimates.squeeze(-1), true_value.squeeze(-1))
|
1041 |
+
action_diversity = ActionDiversityReward()(encoded_actions.view(-1, embed_dim))
|
1042 |
+
mcts_best_values = torch.zeros(actions.size(0)).to(device)
|
1043 |
+
etv = ExpectedThoughtValueLoss()(mcts_best_values)
|
1044 |
+
visit_counts = torch.ones(actions.size(0), policy_logits.size(-1)).to(device)
|
1045 |
+
exploration = ExplorationRegularization()(visit_counts)
|
1046 |
+
old_policy = F.softmax(policy_logits.detach(), dim=-1)
|
1047 |
+
new_policy = F.softmax(policy_logits, dim=-1)
|
1048 |
+
kl_loss = KL_DivergenceLoss()(old_policy, new_policy)
|
1049 |
+
|
1050 |
+
# Total Loss
|
1051 |
+
loss = (
|
1052 |
+
ppo_loss +
|
1053 |
+
info_nce +
|
1054 |
+
covariance +
|
1055 |
+
dynamics_loss +
|
1056 |
+
thought_loss +
|
1057 |
+
pv_loss +
|
1058 |
+
action_diversity +
|
1059 |
+
etv +
|
1060 |
+
exploration +
|
1061 |
+
kl_loss
|
1062 |
+
)
|
1063 |
+
loss = loss / args.accumulation_steps
|
1064 |
+
|
1065 |
+
print("Backward pass...")
|
1066 |
+
scaler.scale(loss).backward()
|
1067 |
+
|
1068 |
+
if (i + 1) % args.accumulation_steps == 0:
|
1069 |
+
print("Gradient clipping...")
|
1070 |
+
scaler.unscale_(optimizer)
|
1071 |
+
torch.nn.utils.clip_grad_norm_(
|
1072 |
+
[param for group in optimizer.param_groups for param in group['params']],
|
1073 |
+
args.max_grad_norm
|
1074 |
+
)
|
1075 |
+
|
1076 |
+
print("Optimizer step...")
|
1077 |
+
scaler.step(optimizer)
|
1078 |
+
scaler.update()
|
1079 |
+
|
1080 |
+
print("Zeroing gradients...")
|
1081 |
+
optimizer.zero_grad()
|
1082 |
+
|
1083 |
+
print("Updating learning rate...")
|
1084 |
+
scheduler.step()
|
1085 |
+
|
1086 |
+
total_loss += loss.item() * args.accumulation_steps
|
1087 |
+
print(f"Batch {i+1} completed. Current loss: {loss.item():.4f}")
|
1088 |
+
|
1089 |
+
avg_loss = total_loss / len(train_loader)
|
1090 |
+
print(f"World Model training epoch completed. Average loss: {avg_loss:.4f}")
|
1091 |
+
return avg_loss
|
1092 |
+
|
1093 |
+
|
1094 |
+
|
1095 |
+
def evaluate_world_model(world_model_components, model_transformer, eval_loader, args):
|
1096 |
+
representation_network, dynamics_network, prediction_network, action_encoder, ppo_agent = world_model_components
|
1097 |
+
representation_network.eval()
|
1098 |
+
dynamics_network.eval()
|
1099 |
+
prediction_network.eval()
|
1100 |
+
action_encoder.eval()
|
1101 |
+
ppo_agent.policy_network.eval()
|
1102 |
+
|
1103 |
+
total_loss = 0.0
|
1104 |
+
with torch.no_grad():
|
1105 |
+
for batch in eval_loader:
|
1106 |
+
src_batch = batch['input_ids'].to(device)
|
1107 |
+
tgt_batch = batch['labels'].to(device)
|
1108 |
+
|
1109 |
+
# Forward pass through Transformer (on CPU)
|
1110 |
+
transformer_output = model_transformer(src_batch, tgt_batch[:, :-1])
|
1111 |
+
|
1112 |
+
# Encode actions
|
1113 |
+
encoded_actions = action_encoder(tgt_batch[:, :-1].to(device)) # Move to GPU if necessary
|
1114 |
+
|
1115 |
+
# World Model - Representation
|
1116 |
+
state = representation_network(transformer_output.to(device))
|
1117 |
+
|
1118 |
+
# Dynamics Network - Predict next state
|
1119 |
+
predicted_next_state = dynamics_network(state, encoded_actions)
|
1120 |
+
|
1121 |
+
# Prediction Network - Policy logits and value
|
1122 |
+
policy_logits, value_estimates = prediction_network(predicted_next_state)
|
1123 |
+
|
1124 |
+
# Placeholder: Define true_policy and true_value
|
1125 |
+
# Replace these with actual targets from your environment or dataset
|
1126 |
+
true_policy = torch.zeros_like(policy_logits).to(device)
|
1127 |
+
true_value = torch.zeros_like(value_estimates).to(device)
|
1128 |
+
|
1129 |
+
# Compute PPO loss
|
1130 |
+
# Placeholder: Replace with actual old_log_probs, actions, returns, and advantages
|
1131 |
+
old_log_probs = torch.zeros_like(policy_logits).to(device)
|
1132 |
+
actions = torch.argmax(policy_logits, dim=-1)
|
1133 |
+
returns = torch.zeros(actions.size(0)).to(device)
|
1134 |
+
advantages = torch.zeros(actions.size(0)).to(device)
|
1135 |
+
|
1136 |
+
ppo_loss = ppo_agent.compute_loss(old_log_probs, actions, returns, advantages)
|
1137 |
+
|
1138 |
+
# Compute other losses
|
1139 |
+
info_nce = InfoNCE_Loss()(state, state) # Placeholder: replace with actual positive pairs
|
1140 |
+
covariance = CovarianceRegularization()(predicted_next_state.view(-1, predicted_next_state.size(-1)))
|
1141 |
+
dynamics_loss = DynamicsPerformanceLoss()(torch.zeros_like(predicted_next_state).to(device), predicted_next_state)
|
1142 |
+
perturbed_next_state = predicted_next_state + torch.randn_like(predicted_next_state) * 0.01
|
1143 |
+
thought_loss = ThoughtConsistencyLoss()(torch.zeros_like(predicted_next_state).to(device), perturbed_next_state)
|
1144 |
+
pv_loss = PolicyValueJointLoss()(policy_logits, true_policy, value_estimates.squeeze(-1), true_value.squeeze(-1))
|
1145 |
+
action_diversity = ActionDiversityReward()(encoded_actions.view(-1, encoded_actions.size(-1)))
|
1146 |
+
mcts_best_values = torch.zeros(actions.size(0)).to(device) # Placeholder: replace with actual MCTS values
|
1147 |
+
etv = ExpectedThoughtValueLoss()(mcts_best_values)
|
1148 |
+
visit_counts = torch.ones(actions.size(0), policy_logits.size(-1)).to(device) # Placeholder: replace with actual visit counts
|
1149 |
+
exploration = ExplorationRegularization()(visit_counts)
|
1150 |
+
old_policy = F.softmax(policy_logits.detach(), dim=-1)
|
1151 |
+
new_policy = F.softmax(policy_logits, dim=-1)
|
1152 |
+
kl_loss = KL_DivergenceLoss()(old_policy, new_policy)
|
1153 |
+
|
1154 |
+
# Total Loss
|
1155 |
+
loss = (
|
1156 |
+
ppo_loss +
|
1157 |
+
info_nce +
|
1158 |
+
covariance +
|
1159 |
+
dynamics_loss +
|
1160 |
+
thought_loss +
|
1161 |
+
pv_loss +
|
1162 |
+
action_diversity +
|
1163 |
+
etv +
|
1164 |
+
exploration +
|
1165 |
+
kl_loss
|
1166 |
+
)
|
1167 |
+
|
1168 |
+
total_loss += loss.item()
|
1169 |
+
|
1170 |
+
avg_loss = total_loss / len(eval_loader)
|
1171 |
+
print(f"World Model evaluation completed. Average loss: {avg_loss:.4f}")
|
1172 |
+
return avg_loss
|
1173 |
+
|
1174 |
+
|
1175 |
+
def main():
|
1176 |
+
args = parse_args()
|
1177 |
+
print("Arguments parsed successfully.")
|
1178 |
+
|
1179 |
+
# Create save directory
|
1180 |
+
if not os.path.exists(args.save_dir):
|
1181 |
+
os.makedirs(args.save_dir)
|
1182 |
+
print(f"Save directory created: {args.save_dir}")
|
1183 |
+
|
1184 |
+
# Load tokenizer
|
1185 |
+
print("Loading tokenizer...")
|
1186 |
+
tokenizer = AutoTokenizer.from_pretrained(args.model_name)
|
1187 |
+
if tokenizer.pad_token is None:
|
1188 |
+
tokenizer.pad_token = tokenizer.eos_token
|
1189 |
+
print("Tokenizer loaded successfully.")
|
1190 |
+
|
1191 |
+
# Define padding_idx and input dimension based on tokenizer
|
1192 |
+
padding_idx = tokenizer.pad_token_id
|
1193 |
+
input_dim = len(tokenizer)
|
1194 |
+
|
1195 |
+
# Load data
|
1196 |
+
print("Loading and preprocessing data...")
|
1197 |
+
train_loader, eval_loader = load_data(args, tokenizer)
|
1198 |
+
print("Data loaded and preprocessed successfully.")
|
1199 |
+
|
1200 |
+
# Define model parameters
|
1201 |
+
d_model = 512 # half to save space
|
1202 |
+
num_heads = 8
|
1203 |
+
num_layers = 6
|
1204 |
+
d_ff = 2048
|
1205 |
+
num_experts = 4
|
1206 |
+
output_dim = input_dim
|
1207 |
+
dropout = 0.1
|
1208 |
+
top_k = 2
|
1209 |
+
state_dim = 128
|
1210 |
+
action_dim = d_model
|
1211 |
+
hidden_dim = 512
|
1212 |
+
vocab_dim = len(tokenizer)
|
1213 |
+
# Initialize and load the Transformer model (on CPU)
|
1214 |
+
print("Initializing and loading Transformer model...")
|
1215 |
+
model_transformer = Transformer(input_dim, d_model, num_heads, num_layers, d_ff, num_experts, output_dim, dropout, top_k)
|
1216 |
+
model_transformer.load_state_dict(torch.load(args.transformer_model_path, map_location='cpu'))
|
1217 |
+
model_transformer.eval()
|
1218 |
+
model_transformer.to('cpu')
|
1219 |
+
print("Transformer model loaded and moved to CPU.")
|
1220 |
+
|
1221 |
+
# Define World Model components
|
1222 |
+
representation_network = RepresentationNetwork(vocab_dim, d_model, state_dim).to(device)
|
1223 |
+
dynamics_network = DynamicsNetwork(state_dim, action_dim, hidden_dim).to(device)
|
1224 |
+
prediction_network = PredictionNetwork(state_dim, input_dim, 1).to(device)
|
1225 |
+
action_encoder = ActionEncoder(input_dim, action_dim).to(device)
|
1226 |
+
|
1227 |
+
# Define Optimizers and Schedulers
|
1228 |
+
optimizer = optim.AdamW(
|
1229 |
+
list(representation_network.parameters()) +
|
1230 |
+
list(dynamics_network.parameters()) +
|
1231 |
+
list(prediction_network.parameters()) +
|
1232 |
+
list(action_encoder.parameters()),
|
1233 |
+
lr=args.learning_rate, weight_decay=args.weight_decay
|
1234 |
+
)
|
1235 |
+
scheduler = CosineAnnealingLR(optimizer, T_max=args.num_epochs)
|
1236 |
+
scaler = GradScaler()
|
1237 |
+
|
1238 |
+
# Initialize PPO Agent
|
1239 |
+
ppo_agent = PPOAgent(
|
1240 |
+
policy_network=prediction_network,
|
1241 |
+
optimizer=optim.AdamW(prediction_network.parameters(), lr=args.learning_rate),
|
1242 |
+
clip_epsilon=0.2,
|
1243 |
+
entropy_coef=0.01,
|
1244 |
+
value_coef=0.5
|
1245 |
+
)
|
1246 |
+
|
1247 |
+
# Bundle World Model components
|
1248 |
+
world_model_components = (representation_network, dynamics_network, prediction_network, action_encoder, ppo_agent)
|
1249 |
+
|
1250 |
+
print("Setup complete. Starting training...")
|
1251 |
+
|
1252 |
+
for epoch in range(args.num_epochs):
|
1253 |
+
print(f"Epoch {epoch + 1}/{args.num_epochs} started.")
|
1254 |
+
|
1255 |
+
# Train World Model
|
1256 |
+
avg_train_loss = train_epoch_world_model(
|
1257 |
+
world_model_components,
|
1258 |
+
train_loader,
|
1259 |
+
optimizer,
|
1260 |
+
scheduler,
|
1261 |
+
scaler,
|
1262 |
+
args,
|
1263 |
+
model_transformer,
|
1264 |
+
state_dim,
|
1265 |
+
d_model, # this is the embedding dimension
|
1266 |
+
input_dim
|
1267 |
+
)
|
1268 |
+
|
1269 |
+
print(f"World Model training epoch {epoch + 1} completed. Average loss: {avg_train_loss:.4f}")
|
1270 |
+
|
1271 |
+
# Evaluate World Model
|
1272 |
+
avg_eval_loss = evaluate_world_model(
|
1273 |
+
world_model_components,
|
1274 |
+
model_transformer,
|
1275 |
+
eval_loader,
|
1276 |
+
args
|
1277 |
+
)
|
1278 |
+
print(f"Evaluation for epoch {epoch + 1} completed. Average loss: {avg_eval_loss:.4f}")
|
1279 |
+
|
1280 |
+
print(f"Epoch {epoch + 1}/{args.num_epochs}, Train Loss: {avg_train_loss:.4f}, Eval Loss: {avg_eval_loss:.4f}")
|
1281 |
+
|
1282 |
+
# Save Models
|
1283 |
+
save_all_models(model_transformer, representation_network, dynamics_network, prediction_network, action_encoder, args.save_dir, epoch + 1)
|
1284 |
+
print(f"Models saved for epoch {epoch + 1}")
|
1285 |
+
|
1286 |
+
print("Training completed.")
|
1287 |
+
|
1288 |
+
|
1289 |
+
if __name__ == '__main__':
|
1290 |
+
main()
|
1291 |
+
|
1292 |
+
|