-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathsam_train_losses.py
228 lines (182 loc) · 9.17 KB
/
sam_train_losses.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
from segment_anything import sam_model_registry
import pickle
import matplotlib.pyplot as plt
import os
import torch.nn.functional as F
from tqdm import tqdm
import torch
from torch.optim import Adam
from torch.optim import SGD
from torch.utils.data import DataLoader
import argparse
import torch.nn as nn
from torch.nn.functional import threshold, normalize
from torch.utils import tensorboard
import numpy as np
from dataset import SAMDataset
from utils.common import save_model, load_model
def dice_loss(pred, target, smooth=1.):
intersection = (pred * target).sum()
return 1 - (2. * intersection + smooth) / (pred.sum() + target.sum() + smooth)
def iou_loss(pred, target):
intersection = (pred * target).sum()
union = pred.sum() + target.sum() - intersection
return 1 - (intersection + 1) / (union + 1)
def focal_loss(pred, target, alpha=0.8, gamma=2):
BCE_loss = F.binary_cross_entropy_with_logits(pred, target, reduction='none')
pt = torch.exp(-BCE_loss)
F_loss = alpha * (1-pt)**gamma * BCE_loss
return F_loss.mean()
def train(args):
model_type = 'vit_h'
# checkpoint = 'weights/sam_vit_b_01ec64.pth'
checkpoint = 'weights/sam_vit_h_4b8939.pth'
sam_model = sam_model_registry[model_type](checkpoint=checkpoint)
device = "cuda:2" if torch.cuda.is_available() else "cpu"
sam_model.to(device)
sam_model.train()
train_dataset = SAMDataset(args.data_dir, sam_model.image_encoder.img_size, out_mask_shape=256, split='train')
# val_dataset = SAMDataset(args.data_dir, processor=processor, split='val')
train_dataloader = DataLoader(train_dataset, batch_size=1, shuffle=True, num_workers=4)
# val_dataloader = DataLoader(val_dataset, batch_size=1, shuffle=False)
# Initialize model
# make sure we only compute gradients for mask decoder
for name, param in sam_model.named_parameters():
if name.startswith("vision_encoder") or name.startswith("prompt_encoder"):
param.requires_grad_(False)
# Note: Hyperparameter tuning could improve performance here
optimizer = SGD(sam_model.mask_decoder.parameters(), lr=args.lr, weight_decay=0)
loss_functions = {
'mse': torch.nn.MSELoss(),
'dice': dice_loss,
'iou': iou_loss,
'focal': lambda pred, target: focal_loss(pred, target, gamma=args.gamma),
# Add other losses here
}
if args.loss not in loss_functions:
raise ValueError(f"Loss {args.loss} not recognized. Available options: {list(loss_functions.keys())}")
loss_fn = loss_functions[args.loss]
if args.pretrained_model:
cur_epoch = load_model(sam_model, args.pretrained_model, optimizer)
start_epoch = cur_epoch + 1
else:
start_epoch = 0
os.makedirs(args.save_dir, exist_ok=True)
summary = tensorboard.SummaryWriter(args.save_dir)
print('starting training from epoch', start_epoch)
epoch_loss_history = []
for epoch in range(start_epoch, args.epochs):
epoch_losses = []
pbar = tqdm(train_dataloader, desc=f'epoch {epoch+1}/{args.epochs}; batch: {0}/{len(train_dataloader)}; loss: {0}')
for idx, batch in enumerate(pbar):
try:
batch = next(iter(train_dataloader)) # forward pass
input_images = sam_model.preprocess(batch['image'].to(device))
prompt_boxes = batch['input_boxes'].to(device)
# input_size = batch['input_size'][0]
input_size = tuple(input_images.shape[-2:])
original_image_size = batch['original_image_size'][0].cpu().numpy().tolist()
# No grad here as we don't want to optimise the encoders
with torch.no_grad():
image_embedding = sam_model.image_encoder(input_images)
boxes_torch = torch.as_tensor(prompt_boxes, dtype=torch.float, device=device)
# boxes_torch = boxes_torch[None, :]
sparse_embeddings, dense_embeddings = sam_model.prompt_encoder(
points=None,
boxes=boxes_torch,
masks=None,
)
pred_masks, iou_predictions = sam_model.mask_decoder(
image_embeddings=image_embedding,
image_pe=sam_model.prompt_encoder.get_dense_pe(),
sparse_prompt_embeddings=sparse_embeddings,
dense_prompt_embeddings=dense_embeddings,
multimask_output=False,
)
# upscaled_masks = sam_model.postprocess_masks(low_res_masks, input_size, original_image_size).to(device)
# binary_mask = normalize(threshold(upscaled_masks, 0.0, 0))
ground_truth_masks = batch["ground_truth_mask"].to(device)
# gt_mask_resized = ground_truth_masks.unsqueeze(1).to(device)
gt_binary_mask = torch.as_tensor(ground_truth_masks > 0, dtype=torch.float32).unsqueeze(1)
loss = loss_fn(pred_masks, gt_binary_mask)
# print(loss.item() )
# backward pass (compute gradients of parameters w.r.t. loss)
optimizer.zero_grad()
loss.backward()
# optimize
optimizer.step()
epoch_losses.append(loss.item())
pbar.set_description(f'epoch {epoch+1}/{args.epochs}; batch: {idx+1}/{len(train_dataloader)}; loss: {loss.item():.6f}')
except Exception as e:
print(f"Error encountered in batch {idx}: {e}")
continue
cur_step = epoch * len(train_dataloader) + idx + 1
summary.add_scalar('loss/step', loss.item(), cur_step)
loss = np.mean(epoch_losses)
epoch_loss_history.append(loss)
print(f'EPOCH: {epoch+1}/{args.epochs}, loss: {loss:.6f}')
summary.add_scalar('loss/epoch', loss, epoch+1)
# Save losses after every epoch
with open(os.path.join(args.save_dir, 'epoch_loss_history.pkl'), 'wb') as f:
pickle.dump(epoch_loss_history, f)
save_path = os.path.join(args.save_dir, f"model_with_optim.pt")
# save with optimizer, useful for fine-tuning
save_model(sam_model, save_path, epoch=epoch, optimizer=optimizer, data_parallel=False)
# save model only, can be loaded directly to predict
save_path = os.path.join(args.save_dir, f"model.pt")
save_model(sam_model, save_path, data_parallel=False, make_dict=False)
# save losses
save_path = os.path.join(args.save_dir, 'epoch_loss_history.pkl')
with open(save_path, 'wb') as f:
pickle.dump(epoch_loss_history, f)
print(f"Saved epoch loss history to {save_path}")
return epoch_loss_history
if __name__ == '__main__':
parser = argparse.ArgumentParser('SAM model training')
parser.add_argument('--data_dir', type=str, default='data', help='path to data directory')
parser.add_argument('--save_dir', type=str, default='checkpoints/focal_casia', help='path to save directory')
parser.add_argument('--pretrained_model', type=str, default='', help='path to pretrained model')
parser.add_argument('--gamma', type=float, default=2, help='gamma value for focal loss')
parser.add_argument('--loss', type=str, default='focal', help='Loss function to use. Options: mse, dice, iou, focal.')
parser.add_argument('--epochs', type=int, default=100, help='number of epochs')
# parser.add_argument('--batch_size', type=int, default=1, help='batch size') # doesnot work with batch size > 1
parser.add_argument('--lr', type=float, default=5e-5, help='learning rate')
args = parser.parse_args()
# epoch_loss_history = train(args)
original_save_dir = args.save_dir
loss_histories ={}
# Setting font for plots
font = {'family': 'serif',
'color': 'black',
'weight': 'normal',
'size': 16,
}
for gamma in [2.0]:
args.gamma = gamma
args.save_dir = os.path.join(original_save_dir, f"gamma_{gamma}")
os.makedirs(args.save_dir, exist_ok=True)
# Train with the current gamma value
losses = train(args) # Assuming your train function returns a list of losses
loss_histories[gamma] = losses
# Plot individual gamma
plt.figure(figsize=(10, 6))
plt.plot(losses, linewidth=2)
plt.title(f'Training Loss Curve for Gamma={gamma}', fontdict=font)
plt.xlabel('Epoch', fontdict=font)
plt.ylabel('Loss', fontdict=font)
plt.grid(True, linestyle='--', linewidth=0.5, alpha=0.7)
plt.tight_layout()
plt.savefig(os.path.join(args.save_dir, f'loss_curve_gamma_{gamma}.png'), dpi=300)
plt.show()
# Combined loss plot for all gammas
plt.figure(figsize=(12, 7))
for gamma, losses in loss_histories.items():
plt.plot(losses, label=f'Gamma={gamma}', linewidth=2)
plt.title('Training Loss Curve for Different Gammas', fontdict=font)
plt.xlabel('Epoch', fontdict=font)
plt.ylabel('Loss', fontdict=font)
plt.legend(loc='best', fontsize='medium')
plt.grid(True, linestyle='--', linewidth=0.5, alpha=0.7)
plt.tight_layout()
plt.savefig(os.path.join(original_save_dir, 'combined_loss_curve.png'), dpi=300)
plt.show()