Create train.py
Browse files
train.py
ADDED
@@ -0,0 +1,263 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# remember to run preprocess.py before training
|
2 |
+
# preprocess while training is not as effecient
|
3 |
+
|
4 |
+
import torch
|
5 |
+
import torch.nn as nn
|
6 |
+
import torch.nn.functional as F
|
7 |
+
from torch.nn import MultiheadAttention
|
8 |
+
import torch.optim as optim
|
9 |
+
from torch.utils.data import Dataset, DataLoader, random_split
|
10 |
+
import json
|
11 |
+
import time
|
12 |
+
import os
|
13 |
+
import h5py
|
14 |
+
import numpy as np
|
15 |
+
from tqdm import tqdm
|
16 |
+
|
17 |
+
class AttentionBlock(nn.Module):
|
18 |
+
def __init__(self, input_dim, num_heads, key_dim, ff_dim, rate=0.1):
|
19 |
+
super(AttentionBlock, self).__init__()
|
20 |
+
self.multihead_attn = MultiheadAttention(embed_dim=input_dim, num_heads=num_heads)
|
21 |
+
self.dropout1 = nn.Dropout(rate)
|
22 |
+
self.layer_norm1 = nn.LayerNorm(input_dim, eps=1e-6)
|
23 |
+
|
24 |
+
self.ffn = nn.Sequential(
|
25 |
+
nn.Linear(input_dim, ff_dim),
|
26 |
+
nn.ReLU(),
|
27 |
+
nn.Dropout(rate),
|
28 |
+
nn.Linear(ff_dim, input_dim),
|
29 |
+
nn.Dropout(rate)
|
30 |
+
)
|
31 |
+
self.layer_norm2 = nn.LayerNorm(input_dim, eps=1e-6)
|
32 |
+
|
33 |
+
def forward(self, x):
|
34 |
+
attn_output, _ = self.multihead_attn(x, x, x)
|
35 |
+
attn_output = self.dropout1(attn_output)
|
36 |
+
out1 = self.layer_norm1(x + attn_output)
|
37 |
+
|
38 |
+
ffn_output = self.ffn(out1)
|
39 |
+
out2 = self.layer_norm2(out1 + ffn_output)
|
40 |
+
|
41 |
+
return out2
|
42 |
+
|
43 |
+
class TextureContrastClassifier(nn.Module):
|
44 |
+
def __init__(self, input_shape, num_heads=4, key_dim=64, ff_dim=256, rate=0.5):
|
45 |
+
super(TextureContrastClassifier, self).__init__()
|
46 |
+
input_dim = input_shape[1] # assuming the input shape is (seq_len, feature_dim)
|
47 |
+
|
48 |
+
self.rich_texture_attention = AttentionBlock(input_dim, num_heads, key_dim, ff_dim, rate)
|
49 |
+
self.poor_texture_attention = AttentionBlock(input_dim, num_heads, key_dim, ff_dim, rate)
|
50 |
+
|
51 |
+
self.rich_texture_dense = nn.Sequential(
|
52 |
+
nn.Linear(input_dim, 128),
|
53 |
+
nn.ReLU(),
|
54 |
+
nn.Dropout(rate)
|
55 |
+
)
|
56 |
+
|
57 |
+
self.poor_texture_dense = nn.Sequential(
|
58 |
+
nn.Linear(input_dim, 128),
|
59 |
+
nn.ReLU(),
|
60 |
+
nn.Dropout(rate)
|
61 |
+
)
|
62 |
+
|
63 |
+
self.fc = nn.Sequential(
|
64 |
+
nn.Flatten(),
|
65 |
+
nn.Linear(input_shape[0] * 128, 256),
|
66 |
+
nn.ReLU(),
|
67 |
+
nn.Dropout(rate),
|
68 |
+
nn.Linear(256, 128),
|
69 |
+
nn.ReLU(),
|
70 |
+
nn.Dropout(rate),
|
71 |
+
nn.Linear(128, 64),
|
72 |
+
nn.ReLU(),
|
73 |
+
nn.Dropout(rate),
|
74 |
+
nn.Linear(64, 32),
|
75 |
+
nn.ReLU(),
|
76 |
+
nn.Dropout(rate),
|
77 |
+
nn.Linear(32, 16),
|
78 |
+
nn.ReLU(),
|
79 |
+
nn.Dropout(rate),
|
80 |
+
nn.Linear(16, 1),
|
81 |
+
nn.Sigmoid()
|
82 |
+
)
|
83 |
+
|
84 |
+
def forward(self, rich_texture, poor_texture):
|
85 |
+
rich_texture = self.rich_texture_attention(rich_texture)
|
86 |
+
rich_texture = self.rich_texture_dense(rich_texture)
|
87 |
+
|
88 |
+
poor_texture = self.poor_texture_attention(poor_texture)
|
89 |
+
poor_texture = self.poor_texture_dense(poor_texture)
|
90 |
+
|
91 |
+
difference = rich_texture - poor_texture
|
92 |
+
output = self.fc(difference)
|
93 |
+
|
94 |
+
return output
|
95 |
+
|
96 |
+
import os
|
97 |
+
import h5py
|
98 |
+
import numpy as np
|
99 |
+
from tqdm import tqdm
|
100 |
+
|
101 |
+
def load_and_split_data(h5_dir, train_ratio=0.8,max_num=40):
|
102 |
+
train_rich, train_poor, train_labels = [], [], []
|
103 |
+
test_rich, test_poor, test_labels = [], [], []
|
104 |
+
|
105 |
+
for file_name in tqdm(os.listdir(h5_dir)[:60]):
|
106 |
+
if file_name.endswith('.h5'):
|
107 |
+
file_path = os.path.join(h5_dir, file_name)
|
108 |
+
try:
|
109 |
+
with h5py.File(file_path, 'r') as h5f:
|
110 |
+
rich = h5f['rich'][:]
|
111 |
+
poor = h5f['poor'][:]
|
112 |
+
labels = h5f['labels'][:]
|
113 |
+
|
114 |
+
dataset_size = len(labels)
|
115 |
+
train_size = int(train_ratio * dataset_size)
|
116 |
+
indices = np.random.permutation(dataset_size)
|
117 |
+
train_indices = indices[:train_size]
|
118 |
+
test_indices = indices[train_size:]
|
119 |
+
|
120 |
+
train_rich.append(rich[train_indices])
|
121 |
+
train_poor.append(poor[train_indices])
|
122 |
+
train_labels.append(labels[train_indices])
|
123 |
+
|
124 |
+
test_rich.append(rich[test_indices])
|
125 |
+
test_poor.append(poor[test_indices])
|
126 |
+
test_labels.append(labels[test_indices])
|
127 |
+
|
128 |
+
except Exception as e:
|
129 |
+
print(f"Error processing {file_name}: {e}")
|
130 |
+
|
131 |
+
train_rich = np.concatenate(train_rich, axis=0)
|
132 |
+
train_poor = np.concatenate(train_poor, axis=0)
|
133 |
+
train_labels = np.concatenate(train_labels, axis=0)
|
134 |
+
|
135 |
+
test_rich = np.concatenate(test_rich, axis=0)
|
136 |
+
test_poor = np.concatenate(test_poor, axis=0)
|
137 |
+
test_labels = np.concatenate(test_labels, axis=0)
|
138 |
+
|
139 |
+
return train_rich, train_poor, train_labels, test_rich, test_poor, test_labels
|
140 |
+
|
141 |
+
class TextureDataset(Dataset):
|
142 |
+
def __init__(self, rich, poor, labels):
|
143 |
+
self.rich = rich
|
144 |
+
self.poor = poor
|
145 |
+
self.labels = labels
|
146 |
+
|
147 |
+
def __len__(self):
|
148 |
+
return len(self.labels)
|
149 |
+
|
150 |
+
def __getitem__(self, idx):
|
151 |
+
rich = torch.tensor(self.rich[idx], dtype=torch.float32)
|
152 |
+
poor = torch.tensor(self.poor[idx], dtype=torch.float32)
|
153 |
+
label = torch.tensor(self.labels[idx], dtype=torch.float32)
|
154 |
+
return rich, poor, label
|
155 |
+
|
156 |
+
def validate(model, test_loader, criterion, device):
|
157 |
+
model.eval()
|
158 |
+
val_loss = 0.0
|
159 |
+
correct = 0
|
160 |
+
total = 0
|
161 |
+
|
162 |
+
with torch.no_grad():
|
163 |
+
for rich, poor, labels in test_loader:
|
164 |
+
rich, poor, labels = rich.to(device), poor.to(device), labels.to(device)
|
165 |
+
|
166 |
+
outputs = model(rich, poor)
|
167 |
+
outputs = outputs.squeeze()
|
168 |
+
|
169 |
+
loss = criterion(outputs, labels)
|
170 |
+
val_loss += loss.item()
|
171 |
+
|
172 |
+
predicted = (outputs > 0.5).float()
|
173 |
+
total += labels.size(0)
|
174 |
+
correct += (predicted == labels).sum().item()
|
175 |
+
|
176 |
+
val_loss /= len(test_loader)
|
177 |
+
val_accuracy = correct / total
|
178 |
+
print(f'Validation Loss: {val_loss:.4f}, Validation Accuracy: {val_accuracy:.4f}')
|
179 |
+
return val_loss, val_accuracy
|
180 |
+
|
181 |
+
|
182 |
+
|
183 |
+
h5_dir = '/content/drive/MyDrive/h5saves'
|
184 |
+
train_rich, train_poor, train_labels, test_rich, test_poor, test_labels = load_and_split_data(h5_dir, train_ratio=0.8)
|
185 |
+
print(f"Training data: {len(train_labels)} samples")
|
186 |
+
print(f"Testing data: {len(test_labels)} samples")
|
187 |
+
train_dataset = TextureDataset(train_rich, train_poor, train_labels)
|
188 |
+
test_dataset = TextureDataset(test_rich, test_poor, test_labels)
|
189 |
+
batch_size = 2048
|
190 |
+
train_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True, num_workers=4)
|
191 |
+
test_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False, num_workers=4)
|
192 |
+
|
193 |
+
input_shape = (128, 256)
|
194 |
+
model = TextureContrastClassifier(input_shape)
|
195 |
+
criterion = nn.BCELoss()
|
196 |
+
optimizer = optim.Adam(model.parameters(), lr=0.0001)
|
197 |
+
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, mode='min', factor=0.1, patience=5, verbose=True)
|
198 |
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
|
199 |
+
model.to(device)
|
200 |
+
|
201 |
+
history = {'train_loss': [], 'val_loss': [], 'train_accuracy':[], 'val_accuracy': []}
|
202 |
+
save_dir = '/content/drive/MyDrive/model_checkpoints'
|
203 |
+
if not os.path.exists(save_dir):
|
204 |
+
os.makedirs(save_dir)
|
205 |
+
num_epochs = 100
|
206 |
+
|
207 |
+
|
208 |
+
|
209 |
+
for epoch in range(num_epochs):
|
210 |
+
model.train()
|
211 |
+
running_loss = 0.0
|
212 |
+
correct = 0
|
213 |
+
total = 0
|
214 |
+
|
215 |
+
batch_loss = 0.0
|
216 |
+
|
217 |
+
for batch_idx, (rich, poor, labels) in enumerate(train_loader):
|
218 |
+
rich, poor, labels = rich.to(device), poor.to(device), labels.to(device)
|
219 |
+
|
220 |
+
optimizer.zero_grad()
|
221 |
+
|
222 |
+
outputs = model(rich, poor)
|
223 |
+
outputs = outputs.squeeze()
|
224 |
+
|
225 |
+
loss = criterion(outputs, labels)
|
226 |
+
loss.backward()
|
227 |
+
optimizer.step()
|
228 |
+
|
229 |
+
running_loss += loss.item()
|
230 |
+
batch_loss += loss.item()
|
231 |
+
|
232 |
+
predicted = (outputs > 0.5).float()
|
233 |
+
total += labels.size(0)
|
234 |
+
correct += (predicted == labels).sum().item()
|
235 |
+
|
236 |
+
if (batch_idx + 1) % 5 == 0:
|
237 |
+
print(f'\rEpoch [{epoch+1}/{num_epochs}], Batch [{batch_idx+1}], Loss: {batch_loss / 5:.4f}, Accuracy: {correct / total:.2f}', end='')
|
238 |
+
batch_loss = 0.0
|
239 |
+
|
240 |
+
avg_train_loss = running_loss / len(train_loader)
|
241 |
+
train_accuracy = correct / total
|
242 |
+
|
243 |
+
val_loss, val_accuracy = validate(model, test_loader, criterion, device)
|
244 |
+
|
245 |
+
history['train_loss'].append(avg_train_loss)
|
246 |
+
history['val_loss'].append(val_loss)
|
247 |
+
history['val_accuracy'].append(val_accuracy)
|
248 |
+
history['train_accuracy'].append(train_accuracy)
|
249 |
+
|
250 |
+
scheduler.step(val_loss)
|
251 |
+
|
252 |
+
checkpoint_path = os.path.join(save_dir, f'model_epoch_{epoch+1}.pth')
|
253 |
+
torch.save(model.state_dict(), checkpoint_path)
|
254 |
+
print(f'\nModel checkpoint saved for epoch {epoch+1}')
|
255 |
+
|
256 |
+
print(f'Epoch [{epoch+1}/{num_epochs:.4f}], Training Loss: {avg_train_loss:.4f}, Training Accuracy: {train_accuracy:.4f} Validation Loss: {val_loss:.4f}, Validation Accuracy: {val_accuracy:.4f}')
|
257 |
+
|
258 |
+
history_path = os.path.join(save_dir, 'training_history.json')
|
259 |
+
with open(history_path, 'w') as f:
|
260 |
+
json.dump(history, f)
|
261 |
+
|
262 |
+
print('Finished Training')
|
263 |
+
print(f'Training history saved at {history_path}')
|