Training script
Browse files- histology_vit.ipynb +451 -0
histology_vit.ipynb
ADDED
@@ -0,0 +1,451 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cells": [
|
3 |
+
{
|
4 |
+
"cell_type": "code",
|
5 |
+
"execution_count": 1,
|
6 |
+
"metadata": {},
|
7 |
+
"outputs": [
|
8 |
+
{
|
9 |
+
"name": "stdout",
|
10 |
+
"output_type": "stream",
|
11 |
+
"text": [
|
12 |
+
"Data loaded successfully!\n",
|
13 |
+
"Number of classes: 32\n",
|
14 |
+
"Class names: ['Adrenocortical_carcinoma', 'Bladder_Urothelial_Carcinoma', 'Brain_Lower_Grade_Glioma', 'Breast_invasive_carcinoma', 'Cervical_squamous_cell_carcinoma_and_endocervical_adenocarcinoma', 'Cholangiocarcinoma', 'Colon_adenocarcinoma', 'Esophageal_carcinoma', 'Glioblastoma_multiforme', 'Head_and_Neck_squamous_cell_carcinoma', 'Kidney_Chromophobe', 'Kidney_renal_clear_cell_carcinoma', 'Kidney_renal_papillary_cell_carcinoma', 'Liver_hepatocellular_carcinoma', 'Lung_adenocarcinoma', 'Lung_squamous_cell_carcinoma', 'Lymphoid_Neoplasm_Diffuse_Large_B-cell_Lymphoma', 'Mesothelioma', 'Ovarian_serous_cystadenocarcinoma', 'Pancreatic_adenocarcinoma', 'Pheochromocytoma_and_Paraganglioma', 'Prostate_adenocarcinoma', 'Rectum_adenocarcinoma', 'Sarcoma', 'Skin_Cutaneous_Melanoma', 'Stomach_adenocarcinoma', 'Testicular_Germ_Cell_Tumors', 'Thymoma', 'Thyroid_carcinoma', 'Uterine_Carcinosarcoma', 'Uterine_Corpus_Endometrial_Carcinoma', 'Uveal_Melanoma']\n",
|
15 |
+
"ViTForCancerClassification(\n",
|
16 |
+
" (vit): VisionTransformer(\n",
|
17 |
+
" (conv_proj): Conv2d(3, 768, kernel_size=(16, 16), stride=(16, 16))\n",
|
18 |
+
" (encoder): Encoder(\n",
|
19 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
20 |
+
" (layers): Sequential(\n",
|
21 |
+
" (encoder_layer_0): EncoderBlock(\n",
|
22 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
23 |
+
" (self_attention): MultiheadAttention(\n",
|
24 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
25 |
+
" )\n",
|
26 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
27 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
28 |
+
" (mlp): MLPBlock(\n",
|
29 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
30 |
+
" (1): GELU(approximate='none')\n",
|
31 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
32 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
33 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
34 |
+
" )\n",
|
35 |
+
" )\n",
|
36 |
+
" (encoder_layer_1): EncoderBlock(\n",
|
37 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
38 |
+
" (self_attention): MultiheadAttention(\n",
|
39 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
40 |
+
" )\n",
|
41 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
42 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
43 |
+
" (mlp): MLPBlock(\n",
|
44 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
45 |
+
" (1): GELU(approximate='none')\n",
|
46 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
47 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
48 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
49 |
+
" )\n",
|
50 |
+
" )\n",
|
51 |
+
" (encoder_layer_2): EncoderBlock(\n",
|
52 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
53 |
+
" (self_attention): MultiheadAttention(\n",
|
54 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
55 |
+
" )\n",
|
56 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
57 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
58 |
+
" (mlp): MLPBlock(\n",
|
59 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
60 |
+
" (1): GELU(approximate='none')\n",
|
61 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
62 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
63 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
64 |
+
" )\n",
|
65 |
+
" )\n",
|
66 |
+
" (encoder_layer_3): EncoderBlock(\n",
|
67 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
68 |
+
" (self_attention): MultiheadAttention(\n",
|
69 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
70 |
+
" )\n",
|
71 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
72 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
73 |
+
" (mlp): MLPBlock(\n",
|
74 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
75 |
+
" (1): GELU(approximate='none')\n",
|
76 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
77 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
78 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
79 |
+
" )\n",
|
80 |
+
" )\n",
|
81 |
+
" (encoder_layer_4): EncoderBlock(\n",
|
82 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
83 |
+
" (self_attention): MultiheadAttention(\n",
|
84 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
85 |
+
" )\n",
|
86 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
87 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
88 |
+
" (mlp): MLPBlock(\n",
|
89 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
90 |
+
" (1): GELU(approximate='none')\n",
|
91 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
92 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
93 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
94 |
+
" )\n",
|
95 |
+
" )\n",
|
96 |
+
" (encoder_layer_5): EncoderBlock(\n",
|
97 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
98 |
+
" (self_attention): MultiheadAttention(\n",
|
99 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
100 |
+
" )\n",
|
101 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
102 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
103 |
+
" (mlp): MLPBlock(\n",
|
104 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
105 |
+
" (1): GELU(approximate='none')\n",
|
106 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
107 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
108 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
109 |
+
" )\n",
|
110 |
+
" )\n",
|
111 |
+
" (encoder_layer_6): EncoderBlock(\n",
|
112 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
113 |
+
" (self_attention): MultiheadAttention(\n",
|
114 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
115 |
+
" )\n",
|
116 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
117 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
118 |
+
" (mlp): MLPBlock(\n",
|
119 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
120 |
+
" (1): GELU(approximate='none')\n",
|
121 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
122 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
123 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
124 |
+
" )\n",
|
125 |
+
" )\n",
|
126 |
+
" (encoder_layer_7): EncoderBlock(\n",
|
127 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
128 |
+
" (self_attention): MultiheadAttention(\n",
|
129 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
130 |
+
" )\n",
|
131 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
132 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
133 |
+
" (mlp): MLPBlock(\n",
|
134 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
135 |
+
" (1): GELU(approximate='none')\n",
|
136 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
137 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
138 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
139 |
+
" )\n",
|
140 |
+
" )\n",
|
141 |
+
" (encoder_layer_8): EncoderBlock(\n",
|
142 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
143 |
+
" (self_attention): MultiheadAttention(\n",
|
144 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
145 |
+
" )\n",
|
146 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
147 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
148 |
+
" (mlp): MLPBlock(\n",
|
149 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
150 |
+
" (1): GELU(approximate='none')\n",
|
151 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
152 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
153 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
154 |
+
" )\n",
|
155 |
+
" )\n",
|
156 |
+
" (encoder_layer_9): EncoderBlock(\n",
|
157 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
158 |
+
" (self_attention): MultiheadAttention(\n",
|
159 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
160 |
+
" )\n",
|
161 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
162 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
163 |
+
" (mlp): MLPBlock(\n",
|
164 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
165 |
+
" (1): GELU(approximate='none')\n",
|
166 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
167 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
168 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
169 |
+
" )\n",
|
170 |
+
" )\n",
|
171 |
+
" (encoder_layer_10): EncoderBlock(\n",
|
172 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
173 |
+
" (self_attention): MultiheadAttention(\n",
|
174 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
175 |
+
" )\n",
|
176 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
177 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
178 |
+
" (mlp): MLPBlock(\n",
|
179 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
180 |
+
" (1): GELU(approximate='none')\n",
|
181 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
182 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
183 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
184 |
+
" )\n",
|
185 |
+
" )\n",
|
186 |
+
" (encoder_layer_11): EncoderBlock(\n",
|
187 |
+
" (ln_1): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
188 |
+
" (self_attention): MultiheadAttention(\n",
|
189 |
+
" (out_proj): NonDynamicallyQuantizableLinear(in_features=768, out_features=768, bias=True)\n",
|
190 |
+
" )\n",
|
191 |
+
" (dropout): Dropout(p=0.0, inplace=False)\n",
|
192 |
+
" (ln_2): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
193 |
+
" (mlp): MLPBlock(\n",
|
194 |
+
" (0): Linear(in_features=768, out_features=3072, bias=True)\n",
|
195 |
+
" (1): GELU(approximate='none')\n",
|
196 |
+
" (2): Dropout(p=0.0, inplace=False)\n",
|
197 |
+
" (3): Linear(in_features=3072, out_features=768, bias=True)\n",
|
198 |
+
" (4): Dropout(p=0.0, inplace=False)\n",
|
199 |
+
" )\n",
|
200 |
+
" )\n",
|
201 |
+
" )\n",
|
202 |
+
" (ln): LayerNorm((768,), eps=1e-06, elementwise_affine=True)\n",
|
203 |
+
" )\n",
|
204 |
+
" (heads): Sequential(\n",
|
205 |
+
" (head): Linear(in_features=768, out_features=32, bias=True)\n",
|
206 |
+
" )\n",
|
207 |
+
" )\n",
|
208 |
+
")\n"
|
209 |
+
]
|
210 |
+
}
|
211 |
+
],
|
212 |
+
"source": [
|
213 |
+
"import torch\n",
|
214 |
+
"import torch.nn as nn\n",
|
215 |
+
"from torch.utils.data import DataLoader, Dataset, WeightedRandomSampler\n",
|
216 |
+
"import torchvision\n",
|
217 |
+
"from torchvision import datasets, transforms\n",
|
218 |
+
"from torch.utils.data import Subset\n",
|
219 |
+
"import numpy as np\n",
|
220 |
+
"import os\n",
|
221 |
+
"import pickle\n",
|
222 |
+
"from tqdm.auto import tqdm\n",
|
223 |
+
"from pathlib import Path\n",
|
224 |
+
"from torchvision.models import vit_b_16, ViT_B_16_Weights\n",
|
225 |
+
"\n",
|
226 |
+
"os.environ['CUDA_LAUNCH_BLOCKING'] = '1'\n",
|
227 |
+
"\n",
|
228 |
+
"# Paths to save the dataloaders and class information\n",
|
229 |
+
"save_path = \"saved_objects\"\n",
|
230 |
+
"class_info_path = os.path.join(save_path, 'class_info.pkl')\n",
|
231 |
+
"train_dataloader_path = os.path.join(save_path, 'train_dataloader.pkl')\n",
|
232 |
+
"test_dataloader_path = os.path.join(save_path, 'test_dataloader.pkl')\n",
|
233 |
+
"\n",
|
234 |
+
"# Create directory if not exists\n",
|
235 |
+
"os.makedirs(save_path, exist_ok=True)\n",
|
236 |
+
"\n",
|
237 |
+
"# Function to load saved objects\n",
|
238 |
+
"def load_saved_data():\n",
|
239 |
+
" if os.path.exists(class_info_path) and os.path.exists(train_dataloader_path) and os.path.exists(test_dataloader_path):\n",
|
240 |
+
" with open(class_info_path, 'rb') as f:\n",
|
241 |
+
" class_info = pickle.load(f)\n",
|
242 |
+
" total_samples = class_info['total_samples']\n",
|
243 |
+
" class_weights = class_info['class_weights']\n",
|
244 |
+
" sample_weights = class_info['sample_weights']\n",
|
245 |
+
"\n",
|
246 |
+
" with open(train_dataloader_path, 'rb') as f:\n",
|
247 |
+
" train_dataloader = pickle.load(f)\n",
|
248 |
+
"\n",
|
249 |
+
" with open(test_dataloader_path, 'rb') as f:\n",
|
250 |
+
" test_dataloader = pickle.load(f)\n",
|
251 |
+
"\n",
|
252 |
+
" print(\"Data loaded successfully!\")\n",
|
253 |
+
" return total_samples, class_weights, sample_weights, train_dataloader, test_dataloader\n",
|
254 |
+
" else:\n",
|
255 |
+
" return None, None, None, None, None\n",
|
256 |
+
"\n",
|
257 |
+
"# Function to save objects\n",
|
258 |
+
"def save_data(total_samples, class_weights, sample_weights, train_dataloader, test_dataloader):\n",
|
259 |
+
" with open(class_info_path, 'wb') as f:\n",
|
260 |
+
" pickle.dump({\n",
|
261 |
+
" 'total_samples': total_samples,\n",
|
262 |
+
" 'class_weights': class_weights,\n",
|
263 |
+
" 'sample_weights': sample_weights\n",
|
264 |
+
" }, f)\n",
|
265 |
+
"\n",
|
266 |
+
" with open(train_dataloader_path, 'wb') as f:\n",
|
267 |
+
" pickle.dump(train_dataloader, f)\n",
|
268 |
+
"\n",
|
269 |
+
" with open(test_dataloader_path, 'wb') as f:\n",
|
270 |
+
" pickle.dump(test_dataloader, f)\n",
|
271 |
+
"\n",
|
272 |
+
" print(\"Data saved successfully!\")\n",
|
273 |
+
"\n",
|
274 |
+
"# Define the ViT model\n",
|
275 |
+
"class ViTForCancerClassification(nn.Module):\n",
|
276 |
+
" def __init__(self, num_classes):\n",
|
277 |
+
" super(ViTForCancerClassification, self).__init__()\n",
|
278 |
+
" self.vit = vit_b_16(weights=ViT_B_16_Weights.DEFAULT)\n",
|
279 |
+
" \n",
|
280 |
+
" # Get the input features of the classifier\n",
|
281 |
+
" in_features = self.vit.heads.head.in_features # Access the head layer specifically\n",
|
282 |
+
" \n",
|
283 |
+
" # Replace the head with a new classification layer\n",
|
284 |
+
" self.vit.heads.head = nn.Linear(in_features, num_classes)\n",
|
285 |
+
" \n",
|
286 |
+
" def forward(self, x):\n",
|
287 |
+
" return self.vit(x)\n",
|
288 |
+
"\n",
|
289 |
+
"# Function to get attention weights\n",
|
290 |
+
"def get_attention_weights(model, x):\n",
|
291 |
+
" with torch.no_grad():\n",
|
292 |
+
" outputs = model.vit._process_input(x)\n",
|
293 |
+
" outputs = model.vit.encoder(outputs)\n",
|
294 |
+
" return model.vit.encoder.layers[-1].self_attention.attention_weights\n",
|
295 |
+
"\n",
|
296 |
+
"# Try to load saved data\n",
|
297 |
+
"total_samples, class_weights, sample_weights, train_dataloader, test_dataloader = load_saved_data()\n",
|
298 |
+
"\n",
|
299 |
+
"# If the data is not available, run preprocessing\n",
|
300 |
+
"if total_samples is None:\n",
|
301 |
+
" print(\"No saved data found. Running data preprocessing...\")\n",
|
302 |
+
"\n",
|
303 |
+
" # Data loading and preprocessing\n",
|
304 |
+
" data_path = Path('TCGA')\n",
|
305 |
+
" transform = transforms.Compose([\n",
|
306 |
+
" transforms.Resize((224, 224)), # ViT typically expects 224x224 input\n",
|
307 |
+
" transforms.ToTensor(),\n",
|
308 |
+
" transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])\n",
|
309 |
+
" ])\n",
|
310 |
+
"\n",
|
311 |
+
" full_dataset = datasets.ImageFolder(root=data_path, transform=transform)\n",
|
312 |
+
" valid_indices = [i for i, (_, label) in enumerate(full_dataset.samples)]\n",
|
313 |
+
" dataset = Subset(full_dataset, valid_indices)\n",
|
314 |
+
"\n",
|
315 |
+
" class_names = [name for name, idx in full_dataset.class_to_idx.items()]\n",
|
316 |
+
" class_to_idx = {name: idx for name, idx in full_dataset.class_to_idx.items()}\n",
|
317 |
+
" print(class_names, class_to_idx)\n",
|
318 |
+
"\n",
|
319 |
+
" # Calculate class weights\n",
|
320 |
+
" class_counts = [0] * len(class_names)\n",
|
321 |
+
" for _, label in dataset:\n",
|
322 |
+
" class_counts[label] += 1\n",
|
323 |
+
" total_samples = sum(class_counts)\n",
|
324 |
+
" class_weights = [total_samples / (len(class_names) * count) for count in class_counts]\n",
|
325 |
+
" sample_weights = [class_weights[label] for _, label in dataset]\n",
|
326 |
+
"\n",
|
327 |
+
" # Create WeightedRandomSampler\n",
|
328 |
+
" sampler = WeightedRandomSampler(weights=sample_weights, num_samples=len(sample_weights), replacement=True)\n",
|
329 |
+
"\n",
|
330 |
+
" # Create data loaders\n",
|
331 |
+
" BATCH_SIZE = 128\n",
|
332 |
+
" train_dataloader = DataLoader(dataset, batch_size=BATCH_SIZE, sampler=sampler)\n",
|
333 |
+
" test_dataloader = DataLoader(dataset, batch_size=BATCH_SIZE, shuffle=False)\n",
|
334 |
+
"\n",
|
335 |
+
" # Save the processed data for future use\n",
|
336 |
+
" save_data(total_samples, class_weights, sample_weights, train_dataloader, test_dataloader)\n",
|
337 |
+
"\n",
|
338 |
+
"class_names = ['Adrenocortical_carcinoma', 'Bladder_Urothelial_Carcinoma', 'Brain_Lower_Grade_Glioma', 'Breast_invasive_carcinoma', 'Cervical_squamous_cell_carcinoma_and_endocervical_adenocarcinoma', 'Cholangiocarcinoma', 'Colon_adenocarcinoma', 'Esophageal_carcinoma', 'Glioblastoma_multiforme', 'Head_and_Neck_squamous_cell_carcinoma', 'Kidney_Chromophobe', 'Kidney_renal_clear_cell_carcinoma', 'Kidney_renal_papillary_cell_carcinoma', 'Liver_hepatocellular_carcinoma', 'Lung_adenocarcinoma', 'Lung_squamous_cell_carcinoma', 'Lymphoid_Neoplasm_Diffuse_Large_B-cell_Lymphoma', 'Mesothelioma', 'Ovarian_serous_cystadenocarcinoma', 'Pancreatic_adenocarcinoma', 'Pheochromocytoma_and_Paraganglioma', 'Prostate_adenocarcinoma', 'Rectum_adenocarcinoma', 'Sarcoma', 'Skin_Cutaneous_Melanoma', 'Stomach_adenocarcinoma', 'Testicular_Germ_Cell_Tumors', 'Thymoma', 'Thyroid_carcinoma', 'Uterine_Carcinosarcoma', 'Uterine_Corpus_Endometrial_Carcinoma', 'Uveal_Melanoma']\n",
|
339 |
+
"print(f\"Number of classes: {len(class_names)}\")\n",
|
340 |
+
"print(f\"Class names: {class_names}\")\n",
|
341 |
+
"\n",
|
342 |
+
"# Model setup\n",
|
343 |
+
"num_classes = len(class_names)\n",
|
344 |
+
"device = 'cuda' if torch.cuda.is_available() else 'cpu'\n",
|
345 |
+
"model = ViTForCancerClassification(num_classes).to(device)\n",
|
346 |
+
"print(model)\n",
|
347 |
+
"\n",
|
348 |
+
"# Training setup\n",
|
349 |
+
"torch.manual_seed(42)\n",
|
350 |
+
"EPOCHS = 20\n",
|
351 |
+
"class_weights_tensor = torch.FloatTensor(class_weights).to(device)\n",
|
352 |
+
"loss_fn = nn.CrossEntropyLoss(weight=class_weights_tensor)\n",
|
353 |
+
"optimizer = torch.optim.Adam(model.parameters(), lr=0.001)\n",
|
354 |
+
"\n",
|
355 |
+
"results = {\n",
|
356 |
+
" 'train_loss': [], \n",
|
357 |
+
" 'train_acc': [],\n",
|
358 |
+
" 'test_loss': [],\n",
|
359 |
+
" 'test_acc': []\n",
|
360 |
+
"}"
|
361 |
+
]
|
362 |
+
},
|
363 |
+
{
|
364 |
+
"cell_type": "code",
|
365 |
+
"execution_count": null,
|
366 |
+
"metadata": {},
|
367 |
+
"outputs": [],
|
368 |
+
"source": [
|
369 |
+
"import torch\n",
|
370 |
+
"\n",
|
371 |
+
"# Define the checkpoint file (change to the correct path if necessary)\n",
|
372 |
+
"checkpoint_path = 'vit_cancer_model_state_dict_X.pth' # Replace 'X' with the last saved epoch number\n",
|
373 |
+
"\n",
|
374 |
+
"# Load the saved model if it exists\n",
|
375 |
+
"if os.path.exists(checkpoint_path):\n",
|
376 |
+
" print(f\"Loading model from {checkpoint_path}\")\n",
|
377 |
+
" model.load_state_dict(torch.load(checkpoint_path))\n",
|
378 |
+
" start_epoch = int(checkpoint_path.split('_')[-1].split('.')[0]) + 1\n",
|
379 |
+
"else:\n",
|
380 |
+
" print(\"No checkpoint found, starting training from scratch.\")\n",
|
381 |
+
" start_epoch = 0\n",
|
382 |
+
"\n",
|
383 |
+
"# Resume training\n",
|
384 |
+
"for epoch in range(start_epoch, EPOCHS):\n",
|
385 |
+
" print(f\"Epoch {epoch+1}/{EPOCHS}\")\n",
|
386 |
+
" train_loss, train_acc = 0, 0\n",
|
387 |
+
" model.train()\n",
|
388 |
+
" for batch, (X, y) in tqdm(enumerate(train_dataloader), total=len(train_dataloader)):\n",
|
389 |
+
" X, y = X.to(device), y.to(device)\n",
|
390 |
+
" y_logits = model(X)\n",
|
391 |
+
" y_pred_class = torch.argmax(torch.softmax(y_logits, dim=1), dim=1)\n",
|
392 |
+
" loss = loss_fn(y_logits, y)\n",
|
393 |
+
" train_acc += (y_pred_class == y).sum().item() / len(y)\n",
|
394 |
+
" train_loss += loss.item()\n",
|
395 |
+
" \n",
|
396 |
+
" optimizer.zero_grad()\n",
|
397 |
+
" loss.backward()\n",
|
398 |
+
" optimizer.step()\n",
|
399 |
+
" \n",
|
400 |
+
" train_loss /= len(train_dataloader)\n",
|
401 |
+
" train_acc /= len(train_dataloader)\n",
|
402 |
+
" \n",
|
403 |
+
" results['train_loss'].append(train_loss)\n",
|
404 |
+
" results['train_acc'].append(train_acc)\n",
|
405 |
+
" \n",
|
406 |
+
" model.eval()\n",
|
407 |
+
" test_loss, test_acc = 0, 0\n",
|
408 |
+
" with torch.inference_mode():\n",
|
409 |
+
" for batch, (X, y) in tqdm(enumerate(test_dataloader), total=len(test_dataloader)):\n",
|
410 |
+
" X, y = X.to(device), y.to(device)\n",
|
411 |
+
" \n",
|
412 |
+
" test_logits = model(X)\n",
|
413 |
+
" test_pred_labels = test_logits.argmax(dim=1)\n",
|
414 |
+
" loss = loss_fn(test_logits, y)\n",
|
415 |
+
" test_acc += (test_pred_labels == y).sum().item() / len(y)\n",
|
416 |
+
" test_loss += loss.item()\n",
|
417 |
+
" \n",
|
418 |
+
" test_loss /= len(test_dataloader)\n",
|
419 |
+
" test_acc /= len(test_dataloader)\n",
|
420 |
+
" print(f'Training loss: {train_loss:.5f} acc: {train_acc:.5f} | Testing loss: {test_loss:.5f} acc: {test_acc:.5f}')\n",
|
421 |
+
" \n",
|
422 |
+
" results['test_loss'].append(test_loss)\n",
|
423 |
+
" results['test_acc'].append(test_acc)\n",
|
424 |
+
" \n",
|
425 |
+
" # Save the model checkpoint after every epoch\n",
|
426 |
+
" torch.save(model.state_dict(), f'vit_cancer_model_state_dict_{epoch}.pth')"
|
427 |
+
]
|
428 |
+
}
|
429 |
+
],
|
430 |
+
"metadata": {
|
431 |
+
"kernelspec": {
|
432 |
+
"display_name": "Python 3",
|
433 |
+
"language": "python",
|
434 |
+
"name": "python3"
|
435 |
+
},
|
436 |
+
"language_info": {
|
437 |
+
"codemirror_mode": {
|
438 |
+
"name": "ipython",
|
439 |
+
"version": 3
|
440 |
+
},
|
441 |
+
"file_extension": ".py",
|
442 |
+
"mimetype": "text/x-python",
|
443 |
+
"name": "python",
|
444 |
+
"nbconvert_exporter": "python",
|
445 |
+
"pygments_lexer": "ipython3",
|
446 |
+
"version": "3.12.3"
|
447 |
+
}
|
448 |
+
},
|
449 |
+
"nbformat": 4,
|
450 |
+
"nbformat_minor": 2
|
451 |
+
}
|