-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathmain_cls.py
376 lines (347 loc) · 16.7 KB
/
main_cls.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
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
import os
import time
import torch
import numpy as np
import torch.nn as nn
import torch.optim as optim
import sklearn.metrics as metrics
from torch.optim.lr_scheduler import CosineAnnealingLR, StepLR, ReduceLROnPlateau
from dataset_modelnet import ModelNet40
from dataset_scanobjectnn import ScanObjectNN
from torch.utils.data import DataLoader
from tqdm import tqdm
from model_cls import MetricLoss, Model
from utils.utils import init_folder, seed_everything, cal_loss, IOStream
from utils.utils_args import process_args
def train_one_epoch(args, model, data_loader, opt, criterion, scheduler, epoch=None):
device = args.device
train_loss = 0.0
count = 0.0
model.train()
train_pred = []
train_true = []
total_time = 0.0
for combined in tqdm(data_loader):
start_time = time.time()
if args.dataset == 'modelnet':
data, normal, label = combined
data, normal = data.to(device).float(), normal.to(device).float()
elif args.dataset == 'scanobject':
data, label = combined
data = data.to(device).float()
else:
raise TypeError("invalid dataset")
label = label.to(device).squeeze().long()
batch_size = data.size()[0]
opt.zero_grad()
if args.use_contrast:
logits, pred_list = model(data[..., :3], normal=None)
loss = criterion[0](logits, label) + criterion[1](pred_list[-1], pred_list[:-1])
else:
logits = model(data[..., :3], normal=None)
loss = criterion[0](logits, label)
loss.backward()
opt.step()
preds = logits.max(dim=1)[1]
count += batch_size
train_loss += loss.item() * batch_size
train_true.append(label.cpu().detach().numpy())
train_pred.append(preds.cpu().detach().numpy())
end_time = time.time()
total_time += (end_time - start_time)
if args.scheduler == 'cos':
scheduler.step()
elif args.scheduler == 'step':
if opt.param_groups[0]['lr'] > 1e-5:
scheduler.step()
if opt.param_groups[0]['lr'] < 1e-5:
for param_group in opt.param_groups:
param_group['lr'] = 1e-5
print('train total time is', total_time)
train_true = np.concatenate(train_true)
train_pred = np.concatenate(train_pred)
outstr = 'Train %d, loss: %.6f, train acc: %.6f, train avg acc: %.6f' % (epoch,
train_loss * 1.0 / count,
metrics.accuracy_score(
train_true, train_pred),
metrics.balanced_accuracy_score(
train_true, train_pred))
io.cprint(outstr)
def test_one_epoch(args, model, data_loader, ce_loss, best_test_acc, epoch=None):
device = args.device
test_loss = 0.0
count = 0.0
test_pred = []
test_true = []
total_time = 0.0
model.eval()
with torch.no_grad():
for combined in data_loader:
start_time = time.time()
if args.dataset == 'modelnet':
data, normal, label = combined
data, normal = data.to(device).float(), normal.to(device).float()
elif args.dataset == 'scanobject':
data, label = combined
data = data.to(device).float()
label = label.to(device).squeeze().long()
batch_size = data.size()[0]
logits = model(data[..., :3], normal=None)
loss = ce_loss(logits, label)
preds = logits.max(dim=1)[1]
count += batch_size
test_loss += loss.item() * batch_size
test_true.append(label.cpu().detach().numpy())
test_pred.append(preds.cpu().detach().numpy())
end_time = time.time()
total_time += (end_time - start_time)
# args.ratio_test.cprint(str(counter * 1.0 / (count * 512)))
test_loss = test_loss * 1.0 / count
print('test total time is', total_time)
test_true = np.concatenate(test_true)
test_pred = np.concatenate(test_pred)
test_acc = metrics.accuracy_score(test_true, test_pred)
avg_per_class_acc = metrics.balanced_accuracy_score(test_true, test_pred)
if args.mode == 'train':
outstr = 'Test %d, loss: %.6f, test acc: %.6f, test avg acc: %.6f' % (epoch,
test_loss,
test_acc,
avg_per_class_acc)
else:
outstr = 'Test loss: %.6f, test acc: %.6f, test avg acc: %.6f' % (test_loss,
test_acc,
avg_per_class_acc)
io.cprint(outstr)
if args.mode == 'train':
if test_acc >= best_test_acc:
best_test_acc = test_acc
torch.save(model.state_dict(), f'{ckpt_path}/model_{epoch}.t7')
if epoch == args.epochs - 1:
io.cprint("Best test acc: %.6f\n" % best_test_acc)
torch.save(model.state_dict(), f'{ckpt_path}/model_best.t7')
return best_test_acc
def main(args):
if args.dataset == 'scanobject':
train_loader = DataLoader(
ScanObjectNN(train=True, num_points=args.num_points, normalize=True, transforms=True, rotate=args.train_rot_mode),
num_workers=8, batch_size=args.train_batch_size, shuffle=True, drop_last=True)
test_loader = DataLoader(
ScanObjectNN(train=False, num_points=args.num_points, normalize=True, transforms=False, rotate=args.test_rot_mode),
num_workers=8, batch_size=args.test_batch_size, shuffle=False, drop_last=False)
elif args.dataset == 'modelnet':
train_loader = DataLoader(
ModelNet40(train=True, use_normal=False, normalize=False, transforms=True, rotate=args.train_rot_mode),
num_workers=8, batch_size=args.train_batch_size, shuffle=True, drop_last=True)
test_loader = DataLoader(
ModelNet40(train=False, use_normal=False, normalize=False, transforms=False, rotate=args.test_rot_mode),
num_workers=8, batch_size=args.test_batch_size, shuffle=False, drop_last=False)
else:
raise ValueError("unknown dataset")
args.device = torch.device("cuda" if args.cuda else "cpu")
device = args.device
if args.dataset == 'scanobject':
num_cls = 15
else:
num_cls = 40
# load the model
model = Model(args, output_channels=num_cls).to(device)
if args.mode == 'test':
best_model = f'{ckpt_path}/model_best.t7'
if os.path.exists(best_model):
model.load_state_dict(torch.load(best_model))
if args.use_sgd:
print("Use SGD")
opt = optim.SGD(model.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.wt_decay)
else:
print("Use Adam")
opt = optim.AdamW(model.parameters(), lr=args.lr, weight_decay=args.wt_decay)
if args.scheduler == 'cos':
print("Use CosineAnnealingLR Scheduler")
scheduler = CosineAnnealingLR(opt, args.epochs, eta_min=args.lr / 100)
elif args.scheduler == 'step':
print("Use StepLR Scheduler")
scheduler = StepLR(opt, step_size=80, gamma=0.1)
elif args.scheduler == 'other':
print("Use ReduceLROnPlateau Scheduler")
scheduler = ReduceLROnPlateau(opt, factor=0.5, patience=10, min_lr=args.lr / 100)
else:
raise ValueError("no such scheduling method")
ce_loss = cal_loss
contrast_loss = MetricLoss()
best_test_acc = 0
if args.mode == 'train':
for epoch in range(args.epochs):
train_one_epoch(args, model, train_loader, opt, [ce_loss, contrast_loss], scheduler, epoch)
best_test_acc = test_one_epoch(args, model, test_loader, ce_loss, best_test_acc, epoch)
else:
test_one_epoch(args, model, test_loader, ce_loss, best_test_acc)
# def test(args):
# test_loader = DataLoader(
# ModelNet40(train=False, use_normal=False, normalize=False, transforms=False, rotate='so3'),
# num_workers=8, batch_size=args.test_batch_size, shuffle=False, drop_last=False)
#
# device = torch.device("cuda" if args.cuda else "cpu")
# num_cls = 40
# save_embeddings = args.save_embeddings
# model = Model(args, save_embeddings=save_embeddings, output_channels=num_cls).to(device)
# model = nn.DataParallel(model)
#
# args.model_path = 'new_checkpoints/' + 'cross_attention_useContrast_v2_' + 'train/models/model.t7'
# model.load_state_dict(torch.load(args.model_path))
# model.eval()
#
# # for i, (data, normal, label) in enumerate(test_loader):
# # data, normal = data.to(device).float(), normal.to(device).float()
# # label = label.to(device).squeeze().long()
# # logits, xyz = model(data[..., :3], normal=normal)
# # preds = logits.max(dim=1)[1]
# # if preds == label:
# # logits[:, preds].backward()
# # # pull the gradients out of the model
# # gradients = model.get_activations_gradient() # (b, C, N)
# #
# # # pool the gradients across the channels
# # pooled_gradients = torch.mean(gradients, dim=[0, 2]) # C
# #
# # # get the activations of the last convolutional layer
# # activations = model.get_activations(data)[0].detach()
# #
# # # weight the channels by corresponding gradients
# # for j in range(1024):
# # activations[:, j] *= pooled_gradients[j]
# #
# # # average the channels of the activations
# # heatmap = torch.mean(activations, dim=1).squeeze()
# # heatmap = F.relu(heatmap)
# #
# # # normalize the heatmap
# # heatmap /= torch.max(heatmap) # (N, 1)
# # combined_vis = torch.cat([xyz[0], heatmap[:, None]], dim=-1) # (N, 4)
# # np.savetxt('./attention_map/saved_cam/data_{0}.txt'.format(i), combined_vis.cpu().numpy())
#
# with torch.no_grad():
# # if args.add_noise:
# # for noise_num in args.noise_point:
# # test_pred = []
# # test_true = []
# # total_time = 0.0
# # for data, label in test_loader:
# # start_time = time.time()
# #
# # data, label = data.to(device), label.to(device).squeeze() # [B, N, 3]
# # batch_size = data.size()[0]
# #
# # # # create noise points
# # # noisy_point = np.random.random((batch_size, noise_num, 3))
# # # noisy_point = torch.from_numpy(normalize_data(noisy_point)).to(device)
# # # data[:, :noise_num, :3] = noisy_point
# # # # data[:, :noise_num, :3] = torch.rand((batch_size, noise_num, 3), device=device)
# #
# # #########################
# # data_transpose = data.transpose(2, 1).contiguous() # [B, 3, N]
# # data_ct = data_transpose - data_transpose.mean(dim=-1, keepdim=True)
# # data_pca = PCA(args.num_points, data_ct, data_transpose) # [B, 3, N]
# # data_pca = data_pca.transpose(2, 1).contiguous() # [B, N, 3]
# # #########################
# #
# # logits = model(data, data_pca)
# # preds = logits.max(dim=1)[1]
# # test_true.append(label.cpu().detach().numpy())
# # test_pred.append(preds.cpu().detach().numpy())
# #
# # end_time = time.time()
# # total_time += (end_time - start_time)
# #
# # print('test total time is', total_time)
# # test_true = np.concatenate(test_true)
# # test_pred = np.concatenate(test_pred)
# # test_acc = metrics.accuracy_score(test_true, test_pred)
# # io.cprint('Test with %d noise points:: test acc: %.6f' % (noise_num, test_acc))
# # if args.add_std:
# # for std in args.std_list:
# # test_pred = []
# # test_true = []
# # total_time = 0.0
# # for data, label in test_loader:
# # start_time = time.time()
# #
# # data, label = data.to(device), label.to(device).squeeze() # [B, N, 3]
# # batch_size = data.size()[0]
# #
# # # create noise
# # noise = torch.normal(mean=0, std=std, size=(batch_size, args.num_points, 3)).to(device)
# #
# # np.random.normal(0, std, size=(batch_size, args.num_points, 3))
# #
# # # data_pca += noise
# # data += noise
# #
# # #########################
# # data_transpose = data.transpose(2, 1).contiguous() # [B, 3, N]
# # data_ct = data_transpose - data_transpose.mean(dim=-1, keepdim=True)
# # data_pca = PCA(args.num_points, data_ct, data_transpose) # [B, 3, N]
# # data_pca = data_pca.transpose(2, 1).contiguous() # [B, N, 3]
# # #########################
# #
# # logits = model(data, data_pca)
# # preds = logits.max(dim=1)[1]
# # test_true.append(label.cpu().detach().numpy())
# # test_pred.append(preds.cpu().detach().numpy())
# #
# # end_time = time.time()
# # total_time += (end_time - start_time)
# #
# # print('test total time is', total_time)
# # test_true = np.concatenate(test_true)
# # test_pred = np.concatenate(test_pred)
# # test_acc = metrics.accuracy_score(test_true, test_pred)
# # io.cprint('Test with %.2f std:: Acc: %.6f' % (std, test_acc))
# test_pred = []
# test_true = []
# total_time = 0.0
# label_list = []
# test_embedding = []
# for i, (data, normal, label) in enumerate(test_loader):
# start_time = time.time()
# data, normal = data.to(device).float(), normal.to(device).float()
# label = label.to(device).squeeze().long()
# label_list.append(label.view(-1, 1))
# batch_size = data.size()[0]
# if save_embeddings:
# logits, embedding_list, relation_list = model(data[..., :3], normal=normal)
# layer1_embedding = relation_list[0].transpose(-1, -2).contiguous() # (B, N, 1)
# layer2_embedding = relation_list[1].transpose(-1, -2).contiguous() # (B, N, 1)
# layer3_embedding = relation_list[2].transpose(-1, -2).contiguous() # (B, N, 1)
# layer4_embedding = relation_list[3].transpose(-1, -2).contiguous() # (B, N, 1)
# layer_pos = relation_list[4] # (B, N, 3)
# attn = torch.cat([layer_pos, layer1_embedding, layer2_embedding, layer3_embedding, layer4_embedding],
# dim=-1) # (B, N, 7)
# for j in range(batch_size):
# np.savetxt('./attention_map/saved_points/data_{0}.txt'.format(i * batch_size + j),
# attn[j].cpu().numpy())
# # test_embedding.append(embedding_list[-1].cpu().numpy())
# else:
# logits = model(data[..., :3], normal=normal)
# preds = logits.max(dim=1)[1]
# test_true.append(label.cpu().detach().numpy())
# test_pred.append(preds.cpu().detach().numpy())
# end_time = time.time()
# total_time += (end_time - start_time)
#
# print('test total time is', total_time)
# # if save_embeddings:
# # test_embedding = np.concatenate(test_embedding)
# # test_true = np.concatenate(test_true)
# # test_pred = np.concatenate(test_pred)
# # test_acc = metrics.accuracy_score(test_true, test_pred)
# # io.cprint('test acc: %.6f' % test_acc)
# #
# # # visualization
# # sne_visual(test_embedding, test_pred)
if __name__ == "__main__":
args = process_args(train=True)
seed_everything(args)
exp_path, ckpt_path = init_folder(args)
io = IOStream(f'{exp_path}/run.log')
io.cprint(str(args))
main(args)