File size: 11,559 Bytes
be2715b |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 |
import numpy as np
import os
join = os.path.join
import gc
from tqdm import tqdm
import torch
import monai, random
from dataloader.sam_transforms import ResizeLongestSide
from segment_anything import sam_model_registry
from dataloader.dataloader import sam_dataloader
from utils.SurfaceDice import compute_dice_coefficient
def fit(cfg,
sam_model,
train_loader,
valid_loader,
optimizer,
criterion,
model_save_path):
"""
Function to fit model
"""
best_valid_dice = 0
device = cfg.base.gpu_id
num_epochs = cfg.train.num_epochs
for epoch in range(num_epochs):
sam_model.train()
epoch_loss = 0
valid_dice = 0
print(f"Epoch #{epoch+1}/{num_epochs}")
for step, batch in enumerate(tqdm(train_loader, desc='Model training', unit='batch', leave=True)):
"""
Load precomputed image embeddings to ease training process
We also load mask labels and bounding boxes directly computed from ground truth masks
"""
image, true_mask, boxes = batch['image'], batch['mask'], batch['bboxes']
sam_model = sam_model.to(f"cuda:{device}")
image = image.to(f"cuda:{device}")
true_mask = true_mask.to(f"cuda:{device}")
"""
We freeze image encoder & prompt encoder, only finetune mask decoder
"""
with torch.no_grad():
"""
Compute image embeddings from a batch of images with SAM's frozen encoder
"""
encoder = torch.nn.DataParallel(sam_model.image_encoder, device_ids=[3, 2, 1, 0], output_device=device)
encoder = encoder.to(f"cuda:{encoder.device_ids[0]}")
sam_model = sam_model.to(f"cuda:{encoder.device_ids[0]}")
image = image.to(f"cuda:{encoder.device_ids[0]}")
image = sam_model.preprocess(image[:, :, :])
image_embedding = encoder(image)
"""
Get bounding boxes to make segmentation prediction
We follow the work by Jun Ma & Bo Wang in Segment Anything in Medical Images (2023)
to get bounding boxes from the masks as the boxes prompt for SAM
"""
box_np = boxes.numpy()
sam_trans = ResizeLongestSide(sam_model.image_encoder.img_size)
box = sam_trans.apply_boxes(box_np, (true_mask.shape[-2], true_mask.shape[-1]))
box_torch = torch.as_tensor(box, dtype=torch.float, device=f"cuda:{device}")
if len(box_torch.shape) == 2:
box_torch = box_torch[:, None, :] # (B, 1, 4)
"""
Encode box prompts information with SAM's frozen prompt encoder
"""
prompt_encoder = torch.nn.DataParallel(sam_model.prompt_encoder, device_ids=[0,1,2,3], output_device=device)
prompt_encoder = prompt_encoder.to(f"cuda:{prompt_encoder.device_ids[0]}")
box_torch = box_torch.to(f"cuda:{prompt_encoder.device_ids[0]}")
sparse_embeddings, dense_embeddings = prompt_encoder(
points=None,
boxes=box_torch,
masks=None,
)
"""
We now finetune mask decoder
"""
sam_model = sam_model.to(f"cuda:{device}")
predicted_mask, iou_predictions = sam_model.mask_decoder(
image_embeddings=image_embedding.to(f"cuda:{device}"), # (B, 256, 64, 64)
image_pe=sam_model.prompt_encoder.get_dense_pe(), # (1, 256, 64, 64)
sparse_prompt_embeddings=sparse_embeddings, # (B, 2, 256)
dense_prompt_embeddings=dense_embeddings, # (B, 256, 64, 64)
multimask_output=False,
) # -> (B, 1, 256, 256)
predicted_mask = predicted_mask.to(f"cuda:{device}")
true_mask = true_mask.to(f"cuda:{device}")
loss = criterion(predicted_mask, true_mask)
"""
Upgrade model's params
"""
optimizer.zero_grad(set_to_none=True)
loss.backward()
clip_value = 1 # Clip gradient
torch.nn.utils.clip_grad_norm_(sam_model.mask_decoder.parameters(), clip_value)
optimizer.step()
epoch_loss += loss.item()
"""
Validation step with Dice as the metric
"""
with torch.no_grad():
valid_dice = eval_dice(sam_model,
valid_loader,
device=device)
epoch_loss /= ((step + 1) * len(train_loader))
print(f'Loss: {epoch_loss}\n---')
"""
Save best model
"""
if best_valid_dice < valid_dice:
best_valid_dice = valid_dice
torch.save(sam_model.state_dict(), join(model_save_path, f'{cfg.base.best_valid_model_checkpoint}{cfg.base.random_seed}.pth'))
print(f"Valid dice: {valid_dice*100}")
print('=======================================')
print(f"Best valid dice: {best_valid_dice*100}")
#%% test
def eval_dice(sam_model,
loader,
device):
"""
Function to evaluate model (for both validation and testing phase)
"""
gc.collect()
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
dice_score = 0.
for _, batch in enumerate(tqdm(loader, leave=False)):
"""
Load precomputed embeddings, mask labels and bounding boxes computed directly from ground truth masks
"""
image, true_mask, boxes = batch['image'], batch['mask'], batch['bboxes']
image = image.to(f"cuda:{device}")
true_mask = true_mask.to(f"cuda:{device}", dtype=torch.float32)
"""
Compute image embeddings
"""
encoder = torch.nn.DataParallel(sam_model.image_encoder, device_ids=[3, 2, 1, 0], output_device=device)
encoder = encoder.to(f"cuda:{encoder.device_ids[0]}")
sam_model = sam_model.to(f"cuda:{encoder.device_ids[0]}")
image = image.to(f"cuda:{encoder.device_ids[0]}")
image = sam_model.preprocess(image[:, :, :])
image_embedding = encoder(image)
"""
Get bboxes
"""
box_np = boxes.numpy()
sam_trans = ResizeLongestSide(sam_model.image_encoder.img_size)
box = sam_trans.apply_boxes(box_np, (image_embedding.shape[0], image_embedding.shape[1]))
box_torch = torch.as_tensor(box, dtype=torch.float32, device=device)
if len(box_torch.shape) == 2:
box_torch = box_torch[:, None, :] # (B, 1, 4)
"""
Prompt encoder component
"""
prompt_encoder = torch.nn.DataParallel(sam_model.prompt_encoder, device_ids=[0,1,2,3], output_device=device)
prompt_encoder = prompt_encoder.to(f"cuda:{prompt_encoder.device_ids[0]}")
box_torch = box_torch.to(f"cuda:{prompt_encoder.device_ids[0]}")
sparse_embeddings, dense_embeddings = prompt_encoder(
points=None,
boxes=box_torch,
masks=None,
)
"""
Mask decoder component
"""
sam_model = sam_model.to(f"cuda:{device}")
mask_segmentation, iou_predictions = sam_model.mask_decoder(
image_embeddings=image_embedding.to(f"cuda:{device}"), # (B, 256, 64, 64)
image_pe=sam_model.prompt_encoder.get_dense_pe(), # (1, 256, 64, 64)
sparse_prompt_embeddings=sparse_embeddings, # (B, 2, 256)
dense_prompt_embeddings=dense_embeddings, # (B, 256, 64, 64)
multimask_output=False,
) # -> (B, 256, 256)
"""
Transform prediction and evaluate
"""
true_mask = true_mask.to("cpu")
medsam_seg_prob = torch.sigmoid(mask_segmentation)
medsam_seg_prob = medsam_seg_prob.detach().cpu().numpy().squeeze()
medsam_seg = (medsam_seg_prob > 0.5).astype(np.uint8) # transform from hard masks to soft masks
dice_score += compute_dice_coefficient(true_mask>0, medsam_seg>0)
return dice_score.cpu().numpy()/len(loader)
def medsam_2d(yml_args, cfg):
"""
Training warm up
"""
torch.multiprocessing.set_start_method('spawn')
random.seed(cfg.base.random_seed)
np.random.seed(cfg.base.random_seed)
torch.manual_seed(cfg.base.random_seed)
torch.cuda.manual_seed(cfg.base.random_seed)
torch.backends.cudnn.deterministic = True
gc.collect()
torch.cuda.empty_cache()
torch.cuda.reset_peak_memory_stats()
"""
General configuration
"""
img_shape = (3, 1024) # hard settings image shape as 3 x 1024 x 1024
model_save_path = join("./work_dir", 'SAM-ViT-B')
os.makedirs(model_save_path, exist_ok=True)
print(f"Fine-tuned SAM in {cfg.base.dataset_name} with {cfg.train.optimizer}, LR = {cfg.train.learning_rate}")
"""
Load SAM with its original checkpoint
"""
sam_model = sam_model_registry["vit_b"](checkpoint=cfg.base.original_checkpoint)
"""
Load precomputed embeddings
"""
train_loader, valid_loader, test_loader, _, _ = sam_dataloader(cfg)
"""
Optimizer & learning rate scheduler config
"""
if cfg.train.optimizer == 'sgd':
optimizer = torch.optim.SGD(sam_model.mask_decoder.parameters(),
lr=float(cfg.train.learning_rate),
momentum=0.9)
elif cfg.train.optimizer == 'adam':
optimizer = torch.optim.Adam(sam_model.mask_decoder.parameters(),
lr=float(cfg.train.learning_rate),
weight_decay=0,
amsgrad=True)
elif cfg.train.optimizer == 'adamw':
optimizer = torch.optim.AdamW(sam_model.mask_decoder.parameters(),
lr=float(cfg.train.learning_rate),
weight_decay=0)
else:
raise NotImplementedError(f"Optimizer {cfg.train.optimizer} is not set up yet")
"""
Loss function
In this work, we use a combination of Dice and Cross Entropy Loss to measure SAM's loss values
"""
criterion = monai.losses.DiceCELoss(sigmoid=True,
squared_pred=True,
reduction='mean')
"""
Train model
"""
if not yml_args.use_test_mode:
fit(cfg,
sam_model=sam_model,
train_loader=train_loader,
valid_loader=valid_loader,
optimizer=optimizer,
criterion=criterion,
model_save_path=model_save_path)
"""
Test model
"""
with torch.no_grad():
sam_model_test_dice = sam_model_registry["vit_b"](checkpoint=join(model_save_path, f'{cfg.base.best_valid_model_checkpoint}{cfg.base.random_seed}.pth'))
sam_model_test_dice.eval()
test_dice_score = eval_dice(sam_model_test_dice,
test_loader,
device=cfg.base.gpu_id)
print(f"Test dice score after training with {cfg.train.optimizer}(lr = {cfg.train.learning_rate}): {test_dice_score*100}") |