File size: 12,930 Bytes
36fdbcf |
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 |
import torch
import torch.nn.functional as F
import random
import numpy as np
from src.utils.util import _bbox_mask
from src.utils import scribble, boundary_selection
from .trainer_basic import Trainer_basic
class Trainer(Trainer_basic):
def __init__(self, args, logger):
super().__init__(args, logger)
def forward(self, sam_model, image, label, iter_nums, train=False, return_each_iter=False):
if return_each_iter:
return_mask_total_iter = torch.zeros([iter_nums, 1, image.size(2), image.size(3), image.size(4)])
image_embedding, feature_list = self.sam.image_encoder(image)
self.click_points = []
self.click_labels = []
return_loss = 0
prev_masks = torch.zeros_like(label, dtype=torch.float).to(label.device)
for iter_num in range(iter_nums):
loss = 0
prev_masks_sigmoid = torch.sigmoid(prev_masks) if iter_num > 0 else prev_masks
points_input, labels_input, box_input = self.get_points(prev_masks_sigmoid, label, train_mode=train)
mask, dice_pred = self.iteration_forward(sam_model, feature_list, image_embedding, prev_masks,
points=[points_input, labels_input], boxes=box_input)
# ========================================================
if self.args.multiple_outputs:
dice_pred_best, max_label_index = torch.max(dice_pred, dim=1)
mask_list = [mask[i, max_label_index[i], :].unsqueeze(0) for i in range(mask.size(0))]
mask_best = torch.stack(mask_list, dim=0)
else:
mask_best = mask
# ========================================================
if train:
if self.args.multiple_outputs:
for i in range(mask.size(1)):
single_mask, single_dice = mask[:, i, :].unsqueeze(1), dice_pred[:, i]
loss += self.calculate_loss(single_mask, prev_masks, single_dice, label, labels_input, iter_num)
else:
loss = self.calculate_loss(mask, prev_masks, dice_pred[:, 0], label, labels_input, iter_num)
# ========================================================
if self.args.refine:
if self.args.no_detach:
mask_refine, error_map = self.sam.mask_decoder.refine(image, mask_best,
[self.click_points, self.click_labels],
mask_best)
else:
mask_refine, error_map = self.sam.mask_decoder.refine(image, mask_best, [self.click_points, self.click_labels], mask_best.detach())
print('dice before refine {} and after {}'.format(
self.get_dice_score(torch.sigmoid(mask_best), label),
self.get_dice_score(torch.sigmoid(mask_refine), label)))
# ========================================================
loss += self.loss_segmentation(mask_refine, label) * 1
mask_best = mask_refine
# ========================================================
else:
if self.args.refine:
if self.args.no_detach:
mask_refine, error_map = self.sam.mask_decoder.refine(image, mask_best,
[self.click_points, self.click_labels],
mask_best)
else:
mask_refine, error_map = self.sam.mask_decoder.refine(image, mask_best,
[self.click_points, self.click_labels],
mask_best.detach())
if iter_num == iter_nums - 1 or iter_num == 0:
self.logger.info('dice before refine {} and after {}, label 0: {}, label 1: {}'.format(
self.get_dice_score(torch.sigmoid(mask_best), label), self.get_dice_score(torch.sigmoid(mask_refine), label),
str(labels_input.numel() - torch.count_nonzero(labels_input)), str(torch.count_nonzero(labels_input)) ) )
mask_best = mask_refine
loss = self.get_dice_score(torch.sigmoid(mask_best), label)
return_loss += loss
prev_masks = mask_best
if return_each_iter:
return_mask_total_iter[iter_num, :] = mask_best
if return_each_iter:
return return_loss / iter_nums, return_mask_total_iter
else:
return return_loss / iter_nums, prev_masks
def get_points(self, prev_masks, label, train_mode=True):
mode = 'train' if train_mode else 'validation'
batch_points, batch_labels = self.get_next_point(prev_masks, label, mode=mode)
points_co = torch.cat(batch_points, dim=0).to(self.args.device) # b x num_clicks x 3
points_la = torch.cat(batch_labels, dim=0).to(self.args.device) # b x num_clicks x 1
self.click_points.append(points_co)
self.click_labels.append(points_la)
points_input = points_co
labels_input = points_la
bbox_coords = _bbox_mask(label[:, 0, :], mode=mode, dynamic=self.args.dynamic_box).to(self.args.device) if self.args.use_box else None
return points_input, labels_input, bbox_coords
def get_next_point(self, prev_seg, label, mode='train'): # prev_seg --> probability
batch_points = []
batch_labels = []
pred_masks = (prev_seg > 0.5)
true_masks = (label > 0)
fn_masks = torch.logical_and(true_masks, torch.logical_not(pred_masks))
fp_masks = torch.logical_and(torch.logical_not(true_masks), pred_masks)
to_point_mask = torch.logical_or(fn_masks, fp_masks)
# do_scribble = random.random()
# sample_method = random.choice(['line', 'center', 'default'])
sample_method = 'center'
scribble_types = {
'line': 'LineScribble',
'center': 'CenterlineScribble',
'default': 'ContourScribble'
}
def create_scribble_mask(scribble_type, data):
scribble_object = getattr(scribble, scribble_type)()
scribble_mask = scribble_object.batch_scribble(data).permute(1, 2, 3, 0)
return scribble_mask > 0
points_list = [len(torch.argwhere(to_point_mask[i])) for i in range(to_point_mask.size(0))]
points_min = min(points_list)
num_clicks = self.args.num_clicks if mode == 'train' else self.args.num_clicks_validation
click_size = points_min if num_clicks > points_min else num_clicks
dynamic_size = random.randint(1, click_size) if self.args.dynamic and mode == 'train' else click_size
print(f"num_clicks {num_clicks} points_length: {points_min} dynamic_size: {dynamic_size}")
for i in range(label.shape[0]):
bp_list, bl_list = [], []
points = torch.argwhere(to_point_mask[i])
point_index = np.random.choice(len(points), size=dynamic_size, replace=False)
points_select = points[point_index] # each row tensor([0, x, y, z]), size --> num_clicks x 4
for click_index in range(dynamic_size):
point = points_select[click_index]
if fn_masks[i, 0, point[1], point[2], point[3]]:
is_positive = True
else:
is_positive = False
bp = point[1:].clone().detach().reshape(1, 1, 3)
bl = torch.tensor([int(is_positive), ]).reshape(1, 1)
bp_list.append(bp)
bl_list.append(bl)
if self.args.use_scribble:
fg, bg_orig = fn_masks[i].permute(3, 0, 1, 2).float(), fp_masks[i].permute(3, 0, 1, 2).float()
# ====== with the purpose of efficiency only for first few epochs ======
bbx = _bbox_mask(label[i, 0, :].unsqueeze(0))
diff_ = 15
i_min, i_max = bbx[:, :, 0], bbx[:, :, 3]
j_min, j_max = bbx[:, :, 1], bbx[:, :, 4]
k_min, k_max = bbx[:, :, 2], bbx[:, :, 5]
if max(0, i_min - diff_) < min(i_max + diff_, 126):
i_min, i_max = max(0, i_min - diff_), min(i_max + diff_, 126)
if max(0, j_min - diff_) < min(j_max + diff_, 126):
j_min, j_max = max(0, j_min - diff_), min(j_max + diff_, 126)
if max(0, k_min - diff_) < min(k_max + diff_, 126):
k_min, k_max = max(0, k_min - diff_), min(k_max + diff_, 126)
bg_mask = torch.zeros_like(bg_orig).permute(1, 2, 3, 0)
bg_mask[:, i_min:i_max, j_min:j_max, k_min:k_max] = 1
bg = bg_orig * bg_mask.permute(3, 0, 1, 2)
print('filter out voxels: {}'.format(torch.count_nonzero(bg_orig) - torch.count_nonzero(bg)))
scribble_type = scribble_types.get(sample_method, scribble_types['default'])
scribble_mask_fg = create_scribble_mask(scribble_type, fg)
limit_num = 500
if torch.count_nonzero(scribble_mask_fg) >= limit_num + 50:
a = torch.argwhere(scribble_mask_fg).size(0) - limit_num
random_number = random.randint(0, a)
fg_coors = torch.argwhere(scribble_mask_fg)[:, 1:].unsqueeze(0)[:, random_number: random_number + limit_num, :] # for computation only
else:
fg_coors = torch.argwhere(scribble_mask_fg)[:, 1:].unsqueeze(0)
fg_coors_label = torch.ones(1, fg_coors.size(1))
bp_list.append(fg_coors)
bl_list.append(fg_coors_label)
scribble_mask_bg = create_scribble_mask(scribble_type, bg)
if torch.count_nonzero(scribble_mask_bg) >= limit_num + 50: # dynamic_size is 50
a = torch.argwhere(scribble_mask_bg).size(0) - limit_num
random_number = random.randint(0, a)
bg_coors = torch.argwhere(scribble_mask_bg)[:, 1:].unsqueeze(0)[:, random_number: random_number + limit_num, :]
else:
bg_coors = torch.argwhere(scribble_mask_bg)[:, 1:].unsqueeze(0)
bg_coors_label = torch.zeros(1, bg_coors.size(1))
bp_list.append(bg_coors)
bl_list.append(bg_coors_label)
batch_points.append(torch.cat(bp_list, dim=1))
batch_labels.append(torch.cat(bl_list, dim=1))
# for scribble
if self.args.use_scribble:
smallest_n = min(tensor.size(1) for tensor in batch_labels)
batch_points = [tensor[:, :smallest_n] if tensor.size(1) > smallest_n else tensor for tensor in batch_points]
batch_labels = [tensor[:, :smallest_n] if tensor.size(1) > smallest_n else tensor for tensor in batch_labels]
# # Check the shapes of the adjusted tensors
# for i, tensor in enumerate(batch_points):
# print(f"Tensor {i + 1} shape: {tensor.shape}")
print('First batch: fn: {:.4f}, fp: {:.4f}, label 0: {}, label 1: {}'.format(
torch.count_nonzero(fn_masks[0]) / torch.count_nonzero(true_masks[0]),
torch.count_nonzero(fp_masks[0]) / torch.count_nonzero(true_masks[0]),
str(batch_labels[0].numel() - torch.count_nonzero(batch_labels[0])),
str(torch.count_nonzero(batch_labels[0]))
)
)
print('--- ===================================== ---')
print('--- above before model, below after model ---')
print('--- ===================================== ---')
return batch_points, batch_labels
def iteration_forward(self, sam_model, features, image_embedding, prev_masks, points=None, boxes=None):
prev_masks = F.interpolate(prev_masks, scale_factor=0.25)
features = [features[i].to(self.args.device) for i in range(0, len(features))]
new_point_embedding, new_image_embedding = sam_model.prompt_encoder(
points=points,
boxes=boxes,
masks=prev_masks,
image_embeddings=image_embedding.to(self.args.device)
)
mask, dice_pred = sam_model.mask_decoder(
prompt_embeddings=new_point_embedding, # (B, 2, 256)
image_embeddings=new_image_embedding, # (B, 256, 64, 64)
feature_list=features,
)
return mask, dice_pred
|