-
Notifications
You must be signed in to change notification settings - Fork 0
/
mytrain.py
355 lines (298 loc) · 15.2 KB
/
mytrain.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
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
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
import collections
import os
import random
import time
from pathlib import Path
import simplejson
import numpy
import numpy as np
import matplotlib.pyplot as plt
import cv2
from statistics import mean
from collections import defaultdict
from Eval import eval
import Frame
from segment_anything.utils.transforms import ResizeLongestSide
import torchvision.utils
from tqdm import tqdm
from torch.nn.functional import threshold, normalize
import torch
import modules
from Frame import load_show
from torch.utils.data import Dataset, DataLoader
from Dataset import own_prepare, MyDataset
from tensorboardX import SummaryWriter
import datetime
# ------------------ 训练 ---------------------------
def Mytrain(# model, optimizer, scheduler=None,
datas, seed, num_epochs, lr, device, wd, loss_name,
model_type='vit_b', checkpoint='checkpoint/sam_vit_b_01ec64.pth',
save_dir=None,
visdual=False,
load_button=False,
train_batchsize=2, val_batchsize=4):
# 导入数据
[train_ids, val_ids, transformed_data, prompts, ground_truth_masks] = datas
data_num = len(train_ids) + len(val_ids)
# ------------------------- 训练准备 -------------------------
# 权重保存目录
nowtime = datetime.datetime.now().strftime("%Y%m%d_%H%M")
if save_dir == None:
save_dir = f'log/epochs{num_epochs}_lr{lr}_wd{wd}_{loss_name}_data{data_num}_{model_type}_seed{seed}_time{nowtime}/'
# ---------------------- 导入模型 -------------------
from segment_anything import SamPredictor, sam_model_registry
sam_model = sam_model_registry[model_type](checkpoint=checkpoint)
sam_model.to(device)
sam_model.train()
# 设置优化器和学习率
import torch.optim.lr_scheduler as scheduler
optimizer = torch.optim.Adam(sam_model.mask_decoder.parameters(), lr=lr, weight_decay=wd)
scheduler = scheduler.ExponentialLR(optimizer, gamma=0.98)
# 设置损失函数
loss_fn = torch.nn.MSELoss()
if loss_name == 'MSE':
loss_fn = torch.nn.MSELoss()
elif loss_name == 'Focal':
loss_fn = modules.B_FocalLoss()
elif loss_name == 'aaadwdFocal':
loss_fn = modules.B_FocalLoss()
elif loss_name == 'BCE':
loss_fn = torch.nn.BCELoss()
# 准备结束工作
from Frame import myframe
solver = myframe()
solver.loads(epoch_nums=num_epochs, model=sam_model, scheduler=scheduler, optimizer=optimizer, lr=lr, save_dir=save_dir)
solver.load_stop_pars()
# 可视化特征图
from visdual import visdualer
visdualer = visdualer(is_visdual=visdual)
visdualer.is_visdual = visdual
# ------------------------------------ 训练 ------------------------------------
for epoch in range(num_epochs):
epoch_train_losses = []
epoch_val_losses = []
epoch_train_accuracy = []
epoch_val_accuracy = []
epoch_train_recall = []
epoch_val_recall = []
torch.cuda.empty_cache()
train_num = 0
# 可视化特征图准备
visdualer.get_img_to_pred_before(transformed_data[train_ids[0]]['image'].shape[1:],
transformed_data[train_ids[0]]['original_image_size'])
start_time_train = time.perf_counter()
for id in train_ids:
# 将其传入gpu,因为dataset类中返回的是long类型的,避免了前面加上batch的维度
# [1,c,h,w]
input_image = transformed_data[id]['image'].to(device)
input_size = transformed_data[id]['input_size']
original_image_size = transformed_data[id]['original_image_size']
ppoints = prompts[id]
ground_truth_mask = ground_truth_masks[id]
# 提取图像embedding
with (torch.no_grad()):
# 我们希望通过将编码器包装在torch.no_grad()上下文管理器中来嵌入图像,
# 否则我们将遇到内存问题,同时我们并不打算对image encoder微调。
# 在该模块下,所有计算得出的tensor的requires_grad都自动设置为False
# 这样所有的节点在反向传播的时候就不会自动求导了,大大节约了显存。
# [1,256,64,64]
image_embedding = sam_model.image_encoder(input_image)
sparse_embeddings, dense_embeddings = sam_model.prompt_encoder(
points=ppoints,
boxes=None,
masks=None,
)
# 最后,我们可以生成mask。
# 请注意,此处我们处于单mask生成模式(与通常输出的3个mask相反)。
# [1,1,256,256]
low_res_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,
)
# 最后一步这里是将掩码放大回原始图像大小,因为它们的分辨率较低。
# 我们可以使用Sam.postprocess_masks来实现这一点。
# 我们还希望从预测的mask中生成二值maks,
# 以便我们可以将它们与我们的ground truths进行比较。
# 使用torch函数非常重要,以免破坏反向传播。
# 以下都是[1,1,512,512]
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))
gt_mask_resized = torch.from_numpy(np.resize(ground_truth_mask, (
1, 1, ground_truth_mask.shape[0], ground_truth_mask.shape[1]))).to(device)
gt_binary_mask = torch.as_tensor(gt_mask_resized > 0, dtype=torch.float32)
per_train_loss = loss_fn(binary_mask, gt_binary_mask)
per_train_accuracy = eval.pixel_accuracy(binary_mask, gt_binary_mask)
per_train_recall = eval.recall(binary_mask, gt_binary_mask)
optimizer.zero_grad()
per_train_loss.backward()
optimizer.step()
epoch_train_losses.append(per_train_loss.item())
epoch_train_accuracy.append(per_train_accuracy)
epoch_train_recall.append(per_train_recall)
# 进度条
load_show(train_num, len(train_ids), start_time_train, legend='train', load_button=load_button)
train_num = train_num + 1
# 可视化特征图
visdualer.get_img_to_pred_load(torch.squeeze(input_image, 0), binary_mask, epoch, 'train')
# --------------------------- 验证集部分 -----------------------------
torch.cuda.empty_cache()
with (torch.no_grad()):
sam_model.eval()
visdualer.get_img_to_pred_before(transformed_data[val_ids[0]]['image'].shape[1:], transformed_data[val_ids[0]]['original_image_size'])
val_num = 0
start_time_val = time.perf_counter()
for id in val_ids:
input_image = transformed_data[id]['image'].to(device)
input_size = transformed_data[id]['input_size']
original_image_size = transformed_data[id]['original_image_size']
ppoints = prompts[id]
ground_truth_mask = ground_truth_masks[id]
# 提取图像embedding
image_embedding = sam_model.image_encoder(input_image)
sparse_embeddings, dense_embeddings = sam_model.prompt_encoder(
points=ppoints,
boxes=None,
masks=None,
)
# 最后,我们可以生成mask。
# 请注意,此处我们处于单mask生成模式(与通常输出的3个mask相反)。
low_res_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,
)
# 最后一步这里是将掩码放大回原始图像大小,因为它们的分辨率较低。
# 我们可以使用Sam.postprocess_masks来实现这一点。
# 我们还希望从预测的mask中生成二值maks,
# 以便我们可以将它们与我们的ground truths进行比较。
# 使用torch函数非常重要,以免破坏反向传播。
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))
gt_mask_resized = torch.from_numpy(np.resize(ground_truth_mask, (
1, 1, ground_truth_mask.shape[0], ground_truth_mask.shape[1]))).to(device)
gt_binary_mask = torch.as_tensor(gt_mask_resized > 0, dtype=torch.float32)
per_loss_val = loss_fn(binary_mask, gt_binary_mask)
per_val_accuracy = eval.pixel_accuracy(binary_mask, gt_binary_mask)
per_val_recall = eval.recall(binary_mask, gt_binary_mask)
epoch_val_losses.append(per_loss_val.item())
epoch_val_accuracy.append(per_val_accuracy)
epoch_val_recall.append(per_val_recall)
# 进度条
load_show(val_num, len(val_ids), start_time_val, legend='val', load_button=load_button)
val_num = val_num + 1
# 可视化特征图
visdualer.get_img_to_pred_load(torch.squeeze(input_image, 0), binary_mask, epoch, 'train')
# 更新
solver.epoch_step(epoch_train_losses, epoch_val_losses, epoch_train_accuracy, epoch_val_accuracy, epoch_train_recall, epoch_val_recall)
if solver.is_stop:
break
# 可视化loss
visdualer.get_loss(mean(epoch_train_losses), mean(epoch_val_losses), epoch)
solver.finsh()
plt = solver.get_lossfig()
solver.get_acc_recall_fig()
solver.get_f1_score()
solver.get_lrfig()
visdualer.close()
return save_dir
# 一旦我们完成了训练并对性能提升感到满意,我们可以使用:
# torch.save(model.state_dict(), PATH)
# 来将调整后的模型的状态字典保存。
# 然后,在我们希望对与我们用于微调模型的数据相似的数据执行推理时,
# 我们可以加载此状态字典。
def test(test_dir, save_dir, class_name, device,
model_type='vit_b', checkpoint='checkpoint/sam_vit_b_01ec64.pth'):
# Load up the model with default weights
from segment_anything import sam_model_registry, SamPredictor
sam_model_orig = sam_model_registry[model_type](checkpoint=checkpoint)
sam_model_orig.to(device)
# Set up predictors for both tuned and original models
tuned_checkpoint = save_dir + 'weight.pth'
sam_model_tuned = sam_model_registry[model_type](checkpoint=checkpoint)
sam_model_tuned.to(device)
Frame.load(sam_model_tuned, tuned_checkpoint)
predictor_original = SamPredictor(sam_model_orig)
predictor_tuned = SamPredictor(sam_model_tuned)
# ----------------------- 测试 --------------------------
for file in sorted(Path(test_dir+'gt/').iterdir()):
k = file.stem[:] # stem代表文件名
# 导入原始图像
image = cv2.imread(test_dir + f'image_prompts/{k}.png')
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
# 读取标签的json文件,并保存points至points_lists
with open(test_dir + f'image_prompts/{k}.json') as f:
label_data = simplejson.load(f)
points_list = []
label_list = []
for shape in label_data['shapes']:
points = shape['points']
points_list.append(points[0])
if shape['label'] == class_name:
label_list.append(1.0)
else:
label_list.append(0.0)
points_torch = torch.as_tensor([points_list], dtype=torch.float32, device=device)
label_torch = torch.as_tensor([label_list], dtype=torch.float32, device=device)
# 读取GT mask
gt_grayscale_org = cv2.imread(test_dir + f'gt/{k}.png', cv2.IMREAD_GRAYSCALE)
gt = (gt_grayscale_org == 255)
torch.cuda.empty_cache()
predictor_tuned.set_image(image)
predictor_original.set_image(image)
input_points = points_torch.cpu().numpy().reshape(-1, 2)
input_label = label_torch.cpu().numpy()[0,:]
masks_tuned, _, _ = predictor_tuned.predict(
point_coords=input_points,
point_labels=input_label,
box=None,
multimask_output=False,
)
masks_orig, _, _ = predictor_original.predict(
point_coords=input_points,
point_labels=input_label,
box=None,
multimask_output=False,
)
# %matplotlib inline
_, axs = plt.subplots(1, 3, figsize=(24, 8))
mngr = plt.get_current_fig_manager() # 获取当前figure manager
mngr.window.wm_geometry("+0+0") # 调整窗口在屏幕上弹出的位置
plt.subplots_adjust(left=0.01, right= 0.99, wspace=0.01)
axs[0].imshow(image)
show_mask(masks_tuned, axs[0])
show_points(input_points, input_label, axs[0])
axs[0].set_title('Mask with Tuned Model', fontsize=12)
axs[0].axis('off')
axs[1].imshow(image)
show_mask(masks_orig, axs[1])
show_points(input_points, input_label, axs[1])
axs[1].set_title('Mask with Untuned Model', fontsize=12)
axs[1].axis('off')
axs[2].imshow(image)
show_mask(gt, axs[2])
show_points(input_points, input_label, axs[2])
axs[2].set_title('GT Mask', fontsize=12)
axs[2].axis('off')
plt.show()
def show_mask(mask, ax, random_color=False):
if random_color:
color = np.concatenate([np.random.random(3), np.array([0.6])], axis=0)
else:
color = np.array([30 / 255, 144 / 255, 255 / 255, 0.6])
h, w = mask.shape[-2:]
mask_image = mask.reshape(h, w, 1) * color.reshape(1, 1, -1)
ax.imshow(mask_image)
def show_box(box, ax):
x0, y0 = box[0], box[1]
w, h = box[2] - box[0], box[3] - box[1]
ax.add_patch(plt.Rectangle((x0, y0), w, h, edgecolor='green', facecolor=(0, 0, 0, 0), lw=2))
def show_points(coords, labels, ax, marker_size=375):
pos_points = coords[labels==1]
neg_points = coords[labels==0]
ax.scatter(pos_points[:, 0], pos_points[:, 1], color='green', marker='*', s=marker_size, edgecolor='white', linewidth=1.25)
ax.scatter(neg_points[:, 0], neg_points[:, 1], color='red', marker='*', s=marker_size, edgecolor='white', linewidth=1.25)