-
Notifications
You must be signed in to change notification settings - Fork 22
Expand file tree
/
Copy pathtrain_cls.py
More file actions
executable file
·480 lines (403 loc) · 17.7 KB
/
train_cls.py
File metadata and controls
executable file
·480 lines (403 loc) · 17.7 KB
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
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
import numpy as np
from tqdm import tqdm
import jittor as jt
from jittor import nn
from jittor.contrib import concat
import torch
jt.flags.use_cuda = 1
from networks.cls.pointnet2 import PointNet2_cls
from networks.cls.pointnet import PointNet as PointNet_cls
from networks.cls.dgcnn import DGCNN
from networks.cls.pointcnn import PointCNNcls
from networks.cls.pointconv import PointConvDensityClsSsg
from networks.cls.kpconv import KPCNN
import math
from os.path import exists, join
from data_utils.modelnet40_loader import ModelNet40
from data_utils.kpconv_loader import KPConvLoader
from misc.utils import LRScheduler
import argparse
from datasets.ModelNet40 import ModelNet40Dataset, ModelNet40Sampler, ModelNet40CustomBatch, Modelnet40Config
import time
import pickle
from jittor_utils import auto_diff
def freeze_random_seed():
np.random.seed(0)
def soft_cross_entropy_loss(output, target, smoothing=True):
''' Calculate cross entropy loss, apply label smoothing if needed. '''
target = target.view(-1)
softmax = nn.Softmax(dim=1)
if smoothing:
eps = 0.2
b, n_class = output.shape
one_hot = jt.zeros(output.shape)
for i in range (b):
one_hot[i, target[i].data] = 1
one_hot = one_hot * (1 - eps) + (1 - one_hot) * eps / (n_class - 1)
# print (one_hot[0].data)
log_prb = jt.log(softmax(output))
loss = -(one_hot * log_prb).sum(dim=1).mean()
else:
loss = nn.cross_entropy_loss(output, target)
return loss
def train(net, optimizer, epoch, dataloader, args):
net.train()
pbar = tqdm(dataloader, desc=f'Epoch {epoch} [TRAIN]')
for pts, normals, labels in pbar:
# #output = net(pts, normals)
if args.model == 'pointnet' or args.model == 'dgcnn' :
pts = pts.transpose(0, 2, 1)
if args.model == 'pointnet2':
output = net(pts, normals)
else :
output = net(pts)
loss = soft_cross_entropy_loss(output, labels)
optimizer.step(loss)
pred = np.argmax(output.data, axis=1)
acc = np.mean(pred == labels.data) * 100
pbar.set_description(f'Epoch {epoch} [TRAIN] loss = {loss.data[0]:.2f}, acc = {acc:.2f}')
def train_kpconv(net, optimizer, epoch, dataloader: KPConvLoader):
net.train()
pbar = tqdm(dataloader, desc=f'Epoch {epoch} [TRAIN]')
jt.sync_all(True)
for input_list in pbar:
L = (len(input_list) - 5) // 4
labels = jt.array(input_list[4 * L + 1]).squeeze(0)
output = net(input_list)
loss = soft_cross_entropy_loss(output, labels)
optimizer.step(loss)
pred = np.argmax(output.data, axis=1)
acc = np.mean(pred == labels.data) * 100
pbar.set_description(f'Epoch {epoch} [TRAIN] loss = {loss.data[0]:.2f}, acc = {acc:.2f}')
# jt.display_memory_info()
def evaluate(net, epoch, dataloader, args):
total_acc = 0
total_num = 0
net.eval()
total_time = 0.0
for pts, normals, labels in tqdm(dataloader, desc=f'Epoch {epoch} [Val]'):
# pts = jt.float32(pts.numpy())
# normals = jt.float32(normals.numpy())
# labels = jt.int32(labels.numpy())
# feature = concat((pts, normals), 2)
if args.model == 'pointnet' or args.model == 'dgcnn' :
pts = pts.transpose(0, 2, 1)
# pts = pts.transpose(0, 2, 1) # for pointnet DGCNN
# output = net(pts, feature)
if args.model == 'pointnet2':
output = net(pts, normals)
else :
output = net(pts)
# output = net()
pred = np.argmax(output.data, axis=1)
acc = np.sum(pred == labels.data)
total_acc += acc
total_num += labels.shape[0]
acc = 0.0
acc = total_acc / total_num
return acc
def evaluate_kpconv(net, epoch, dataloader: KPConvLoader):
total_acc = 0
total_num = 0
net.eval()
total_time = 0.0
for input_list in tqdm(dataloader, desc=f'Epoch {epoch} [Val]'):
L = (len(input_list) - 5) // 4
labels = jt.array(input_list[4 * L + 1]).squeeze(0)
output = net(input_list)
pred = np.argmax(output.data, axis=1)
acc = np.sum(pred == labels.data)
total_acc += acc
total_num += labels.shape[0]
acc = 0.0
acc = total_acc / total_num
return acc
def fast_confusion(true, pred, label_values=None): # used only by kpconv test
"""
Fast confusion matrix (100x faster than Scikit learn). But only works if labels are la
:param true:
:param false:
:param num_classes:
:return:
"""
# Ensure data is in the right format
true = np.squeeze(true)
pred = np.squeeze(pred)
if len(true.shape) != 1:
raise ValueError('Truth values are stored in a {:d}D array instead of 1D array'. format(len(true.shape)))
if len(pred.shape) != 1:
raise ValueError('Prediction values are stored in a {:d}D array instead of 1D array'. format(len(pred.shape)))
if true.dtype not in [np.int32, np.int64]:
raise ValueError('Truth values are {:s} instead of int32 or int64'.format(true.dtype))
if pred.dtype not in [np.int32, np.int64]:
raise ValueError('Prediction values are {:s} instead of int32 or int64'.format(pred.dtype))
true = true.astype(np.int32)
pred = pred.astype(np.int32)
# Get the label values
if label_values is None:
# From data if they are not given
label_values = np.unique(np.hstack((true, pred)))
else:
# Ensure they are good if given
if label_values.dtype not in [np.int32, np.int64]:
raise ValueError('label values are {:s} instead of int32 or int64'.format(label_values.dtype))
if len(np.unique(label_values)) < len(label_values):
raise ValueError('Given labels are not unique')
# Sort labels
label_values = np.sort(label_values)
# Get the number of classes
num_classes = len(label_values)
#print(num_classes)
#print(label_values)
#print(np.max(true))
#print(np.max(pred))
#print(np.max(true * num_classes + pred))
# Start confusion computations
if label_values[0] == 0 and label_values[-1] == num_classes - 1:
# Vectorized confusion
vec_conf = np.bincount(true * num_classes + pred)
# Add possible missing values due to classes not being in pred or true
#print(vec_conf.shape)
if vec_conf.shape[0] < num_classes ** 2:
vec_conf = np.pad(vec_conf, (0, num_classes ** 2 - vec_conf.shape[0]), 'constant')
#print(vec_conf.shape)
# Reshape confusion in a matrix
return vec_conf.reshape((num_classes, num_classes))
else:
# Ensure no negative classes
if label_values[0] < 0:
raise ValueError('Unsupported negative classes')
# Get the data in [0,num_classes[
label_map = np.zeros((label_values[-1] + 1,), dtype=np.int32)
for k, v in enumerate(label_values):
label_map[v] = k
pred = label_map[pred]
true = label_map[true]
# Vectorized confusion
vec_conf = np.bincount(true * num_classes + pred)
# Add possible missing values due to classes not being in pred or true
if vec_conf.shape[0] < num_classes ** 2:
vec_conf = np.pad(vec_conf, (0, num_classes ** 2 - vec_conf.shape[0]), 'constant')
# Reshape confusion in a matrix
return vec_conf.reshape((num_classes, num_classes))
def classification_test(net, test_loader: KPConvLoader, config, num_votes=100): # used only by kpconv test
print("validation size:", config.validation_size, "batch_num:", config.batch_num)
############
# Initialize
############
# Choose test smoothing parameter (0 for no smothing, 0.99 for big smoothing)
softmax = jt.nn.Softmax(1)
# Number of classes including ignored labels
nc_tot = test_loader.num_classes
# Number of classes predicted by the model
nc_model = config.num_classes
# Initiate global prediction over test clouds
test_probs = np.zeros((test_loader.num_models, nc_model))
test_counts = np.zeros((test_loader.num_models, nc_model))
print("probs shape:", test_probs.shape)
t = [time.time()]
mean_dt = np.zeros(1)
last_display = time.time()
while np.min(test_counts) < num_votes:
# Run model on all test examples
# ******************************
# Initiate result containers
probs = []
targets = []
obj_inds = []
idx = 0
# Start validation loop
test_loader.prepare_batch_indices()
for input_list in test_loader:
# print("test", idx)
idx += 1
# batch = ModelNet40CustomBatch([input_list])
# labels, model_inds = batch.labels, batch.model_inds
L = (len(input_list) - 5) // 4
labels = jt.array(input_list[4 * L + 1]).squeeze(0)
model_inds = jt.array(input_list[4 * L + 4]).squeeze(0)
# print(model_inds)
# New time
t = t[-1:]
t += [time.time()]
# Forward pass
outputs = net(input_list)
# Get probs and labels
probs += [softmax(outputs).numpy()]
targets += [labels.numpy()]
obj_inds += [model_inds.numpy()]
# print("probs: ", probs)
# print("targets: ", targets)
# print("obj_inds: ", obj_inds)
# Average timing
t += [time.time()]
mean_dt = 0.95 * mean_dt + 0.05 * (np.array(t[1:]) - np.array(t[:-1]))
# Display
if (t[-1] - last_display) > 5.0:
last_display = t[-1]
message = 'Test vote {:.0f} : {:.1f}% (timings : {:4.2f} {:4.2f})'
print(message.format(np.min(test_counts),
100 * len(obj_inds) / config.validation_size,
1000 * (mean_dt[0]),
1000 * (mean_dt[1])))
# Stack all validation predictions
probs = np.vstack(probs)
targets = np.hstack(targets)
obj_inds = np.hstack(obj_inds)
# print(obj_inds.shape)
if np.any(test_loader.input_labels[obj_inds] != targets):
raise ValueError('wrong object indices')
# Compute incremental average (predictions are always ordered)
test_counts[obj_inds] += 1
# print(test_counts.shape)
# print(test_counts)
test_probs[obj_inds] += (probs - test_probs[obj_inds]) / (test_counts[obj_inds])
# Save/Display temporary results
# ******************************
test_labels = np.array(test_loader.label_values)
# Compute classification results
C1 = fast_confusion(test_loader.input_labels,
np.argmax(test_probs, axis=1),
test_labels)
ACC = 100 * np.sum(np.diag(C1)) / (np.sum(C1) + 1e-6)
print('Test Accuracy = {:.1f}%'.format(ACC), flush=True)
return
def hook():
cfg = Modelnet40Config()
net = KPCNN(cfg)
chkp_path = "/mnt/disk1/chentuo/PointNet/KPConv-PyTorch/results/Log_2022-08-04_15-17-48/checkpoints/current_chkp.tar"
checkpoint = torch.load(chkp_path)
net.load_state_dict(checkpoint['model_state_dict'])
# optimizer不控制
print("Model and training state restored.")
other_params = [v for k, v in net.named_parameters() if 'offset' not in k and 'running' not in k]
optimizer = nn.SGD(other_params, lr = cfg.learning_rate, momentum = cfg.momentum, weight_decay=cfg.weight_decay)
f = open('/mnt/disk1/chentuo/PointNet/PointCloudLib/networks/cls/data.txt', 'rb')
data = pickle.load(f)
f.close()
batch = ModelNet40CustomBatch([data])
hook = auto_diff.Hook("KPCNN")
hook.hook_module(net)
# hook = auto_diff.Hook("KPCNN_optim")
# hook.hook_optimizer(optimizer)
outputs = net(batch)
loss = net.loss(outputs, batch.labels)
acc = net.accuracy(outputs, batch.labels)
# jt.display_memory_info()
# Backward + optimize
# idx = 0
# for k, v in net.named_parameters():
# if 'offset' not in k and 'running' not in k and 'output_loss' not in k:
# print(" [", idx, "] ", k, " ###### ", jt.grad(loss, v))
# idx += 1
optimizer.step(loss)
outputs = net(batch)
# jt.display_memory_info()
exit(0)
if __name__ == '__main__':
# hook()
freeze_random_seed()
parser = argparse.ArgumentParser(description='Point Cloud Recognition')
parser.add_argument('--eval', action='store_true', default=False) # only used by kpconv
parser.add_argument('--model', type=str, default='[pointnet]', metavar='N',
choices=['pointnet', 'pointnet2', 'pointcnn', 'dgcnn', 'pointconv', 'kpconv'],
help='Model to use')
parser.add_argument('--batch_size', type=int, default=32, metavar='batch_size',
help='Size of batch)')
parser.add_argument('--lr', type=float, default=0.02, metavar='LR',
help='learning rate (default: 0.02)')
parser.add_argument('--momentum', type=float, default=0.9, metavar='M',
help='SGD momentum (default: 0.9)')
parser.add_argument('--num_points', type=int, default=1024,
help='num of points to use')
parser.add_argument('--epochs', type=int, default=300, metavar='N',
help='number of episode to train ')
args = parser.parse_args()
if args.model == 'pointnet':
net = PointNet_cls()
elif args.model == 'pointnet2':
net = PointNet2_cls()
elif args.model == 'pointcnn':
net = PointCNNcls()
elif args.model == 'dgcnn':
net = DGCNN()
elif args.model == 'pointconv':
net = PointConvDensityClsSsg()
elif args.model == 'kpconv':
cfg = Modelnet40Config()
net = KPCNN(cfg)
else:
raise Exception("Not implemented")
base_lr = args.lr
if args.model != 'kpconv':
optimizer = nn.SGD(net.parameters(), lr = base_lr, momentum = args.momentum)
else:
other_params = [v for k, v in net.named_parameters() if 'offset' not in k and 'running' not in k]
# idx = 0
# for k, v in net.named_parameters():
# if 'offset' not in k and 'running' not in k:
# print(" [", idx, "] ", k, " ###### ", v.shape)
# idx += 1
# exit(0)
optimizer = nn.SGD(other_params, lr = cfg.learning_rate, momentum = cfg.momentum, weight_decay=cfg.weight_decay)
lr_scheduler = LRScheduler(optimizer, base_lr)
batch_size = args.batch_size
n_points = args.num_points
if args.model != 'kpconv':
train_dataloader = ModelNet40(n_points=n_points, batch_size=batch_size, train=True, shuffle=True)
val_dataloader = ModelNet40(n_points=n_points, batch_size=batch_size, train=False, shuffle=False)
else:
if not args.eval:
train_dataloader = KPConvLoader(cfg, train=True, num_workers=0) # you can change num_workers to speed up
cfg.validation_size = 250
cfg.val_batch_num = 10
val_dataloader = KPConvLoader(cfg, train=False, num_workers=4)
#### load model ####
if args.eval:
chkp_path = "/mnt/disk1/chentuo/PointNet/PointCloudLib/checkpoints/kpconv/best_chkp.tar"
checkpoint = jt.load(chkp_path)
net.load_state_dict(checkpoint['model_state_dict'])
# optimizer.load_state_dict(checkpoint['optimizer_state_dict'])
# optimizer.load_state_dict({"defaults": checkpoint['optimizer_state_dict']['param_groups'][0]})
net.eval()
print("Model and training state restored.")
# print(optimizer.state_dict())
# print("#####")
# print(checkpoint['optimizer_state_dict']['param_groups'][0])
#####
step = 0
best_acc = 0
for epoch in range(args.epochs):
if args.model == 'kpconv':
if not args.eval:
train_kpconv(net, optimizer, epoch, train_dataloader)
acc = evaluate_kpconv(net, epoch, val_dataloader)
train_dataloader.prepare_batch_indices()
val_dataloader.prepare_batch_indices()
if epoch in cfg.lr_decays:
optimizer.lr *= cfg.lr_decays[epoch]
if cfg.saving:
# Get current state dict
checkpoint_directory = 'checkpoints/kpconv'
save_dict = {'epoch': epoch,
'model_state_dict': net.state_dict(),
'optimizer_state_dict': optimizer.state_dict(),
'saving_path': cfg.saving_path}
checkpoint_path = join(checkpoint_directory, 'current_chkp.tar')
if acc > best_acc:
checkpoint_path = join(checkpoint_directory, 'best_chkp.tar')
# Save current state of the network (for restoring purposes)
jt.save(save_dict, checkpoint_path)
# Save checkpoints occasionally
if (epoch + 1) % cfg.checkpoint_gap == 0:
checkpoint_path = join(checkpoint_directory, 'chkp_{:04d}.tar'.format(epoch + 1))
jt.save(save_dict, checkpoint_path)
else: # kpconv eval
classification_test(net, val_dataloader, cfg)
exit(0)
else:
lr_scheduler.step(len(train_dataloader) * batch_size)
train(net, optimizer, epoch, train_dataloader, args)
acc = evaluate(net, epoch, val_dataloader, args)
best_acc = max(best_acc, acc)
print(f'[Epoch {epoch}] val acc={acc:.4f}, best={best_acc:.4f}')