-
Notifications
You must be signed in to change notification settings - Fork 1
/
main.py
458 lines (388 loc) · 19.3 KB
/
main.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
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
from datetime import datetime
import argparse
import imageio
import cv2
import numpy as np
import torch
from functools import partial
import torch.nn as nn
import torch.nn.functional as F
from torch.autograd import Variable
from torch.utils.data import DataLoader
import time
from torchvision import transforms
from model import Net
from losses import L1loss, L2loss, training_loss, robust_training_loss, MultiScale, EPE, EPEp
from dataset import (FlyingChairs, FlyingThings, Sintel, SintelFinal, SintelClean, KITTI, mixup)
import tensorflow as tf
from summary import summary as summary_
from logger import Logger
from pathlib import Path
from flow_utils import (vis_flow, save_flow)
def main():
parser = argparse.ArgumentParser(description='',
formatter_class=argparse.ArgumentDefaultsHelpFormatter)
# mode selection
# ============================================================
modes = parser.add_subparsers(title='modes',
description='valid modes',
help='additional help',
dest='subparser_name')
parser.set_defaults(func=hello_world)
summary_parser = modes.add_parser('summary');
summary_parser.set_defaults(func=summary)
train_parser = modes.add_parser('train');
train_parser.set_defaults(func=train)
pred_parser = modes.add_parser('pred');
pred_parser.set_defaults(func=pred)
test_parser = modes.add_parser('eval');
test_parser.set_defaults(func=test)
# shared args
# ============================================================
parser.add_argument('--device', type=str, default='cuda')
# dataset
parser.add_argument('--num_workers', default=8, type=int, help='num of workers')
# normalization args
parser.add_argument('--input-norm', action='store_true')
parser.add_argument('--rgb_max', type=float, default=255)
parser.add_argument('--batch-norm', action='store_true')
# pyramid args
parser.add_argument('--lv_chs', nargs='+', type=int, default=[3, 16, 32, 64, 96, 128, 192])
parser.add_argument('--output_level', type=int, default=4)
# correlation args
# CostVolumeLayer or cost_volume
parser.add_argument('--corr', type=str, default='cost_volume')
parser.add_argument('--search_range', type=int, default=4)
parser.add_argument('--corr_activation', action='store_true')
# flow estimator
parser.add_argument('--residual', action='store_true')
# args for summary
# ============================================================
summary_parser.add_argument('-i', '--input_shape', type=int, nargs='*', default=(3, 2, 384, 448))
# args for train
# ============================================================
# dataflow
train_parser.add_argument('--crop_type', type=str, default='random')
train_parser.add_argument('--crop_shape', type=int, nargs='+', default=[384, 448])
train_parser.add_argument('--resize_shape', nargs=2, type=int, default=None)
train_parser.add_argument('--resize_scale', type=float, default=None)
train_parser.add_argument('--load', type=str, default=None)
train_parser.add_argument('--batch_size', default=8, type=int, help='mini-batch size')
train_parser.add_argument('--dataset_dir', type=str, required=True)
train_parser.add_argument('--dataset', type=str,
choices=['FlyingChairs', 'FlyingThings', 'SintelFinal', 'SintelClean', 'KITTI'],
required=True)
train_parser.add_argument('--mixup', action='store_true')
train_parser.add_argument('--mixup_alpha', default=0.2, type=float, help='beta parm')
train_parser.add_argument('--mixup_prb', default=0.5, type=float, help='mixup probability')
train_parser.add_argument('--no_transforms', action='store_false')
train_parser.add_argument('--erasing', type=float, default=0.7)
# loss
train_parser.add_argument('--weights', nargs='+', type=float, default=[0.32, 0.08, 0.02, 0.01, 0.005])
train_parser.add_argument('--epsilon', default=0.02)
train_parser.add_argument('--q', type=int, default=0.4)
train_parser.add_argument('--loss', type=str, default='MultiScale', choices=['MultiScale'])
train_parser.add_argument('--optimizer', type=str, default='Adam')
# optimize
train_parser.add_argument('--lr', type=float, default=1e-4)
train_parser.add_argument('--momentum', default=4e-4)
train_parser.add_argument('--beta', default=0.99)
train_parser.add_argument('--weight_decay', type=float, default=4e-4)
train_parser.add_argument('--total_step', type=int, default=200 * 1000)
# summary & log args
train_parser.add_argument('--log_dir', default='train_log/' + datetime.now().strftime('%Y%m%d-%H%M%S'))
train_parser.add_argument('--summary_interval', type=int, default=50)
train_parser.add_argument('--log_interval', type=int, default=50)
train_parser.add_argument('--checkpoint_interval', type=int, default=20)
train_parser.add_argument('--gif_input', type=str, default=None)
train_parser.add_argument('--gif_output', type=str, default='gif')
train_parser.add_argument('--gif_interval', type=int, default=100)
train_parser.add_argument('--max_output', type=int, default=3)
# args for predict
# ============================================================
pred_parser.add_argument('-i', '--input', nargs=2, required=True)
pred_parser.add_argument('-o', '--output', default='output.flo')
pred_parser.add_argument('--load', type=str, required=True)
# args for test
# ============================================================
test_parser.add_argument('--load', type=str, required=True)
test_parser.add_argument('--dataset_dir', type=str, required=True)
test_parser.add_argument('--dataset', type=str,
choices=['FlyingChairs', 'FlyingThings', 'SintelFinal', 'SintelClean', 'KITTI'],
required=True)
args = parser.parse_args()
args.num_levels = len(args.lv_chs)
args.device = torch.device(args.device)
# check args
# ============================================================
if args.subparser_name == 'train':
assert len(args.weights) >= args.output_level + 1
args.func(args)
def hello_world(args):
from functools import reduce
from operator import mul
model = Net(args).to(args.device)
state = model.state_dict()
total_size = 0
for key, value in state.items():
print(f'{key}: {value.size()}')
total_size += reduce(mul, value.size())
print(f'Parameters: {total_size} Size: {total_size * 4 / 1024 / 1024} MB')
def summary(args):
model = Net(args).to(args.device)
summary_(model, args.input_shape)
def train(args):
# Build Model
# ============================================================
model = Net(args).to(args.device)
if args.load is not None:
model.load_state_dict(torch.load(args.load))
total = sum([param.nelement() for param in model.parameters()])
print("Number of parameter: %.2fM" % (total / 1e6))
# Prepare DateTransforms
# Prepare Dataloader
# ============================================================
train_dataset = eval(args.dataset)(args.dataset_dir, 'train', cropper=args.crop_type, crop_shape=args.crop_shape,
resize_shape=args.resize_shape, resize_scale=args.resize_scale,
transforms=args.no_transforms)
# eval_dataset = eval(args.dataset)(args.dataset_dir, 'test', cropper=args.crop_type, crop_shape=args.crop_shape,
# resize_shape=args.resize_shape, resize_scale=args.resize_scale)
# print(len(train_dataset))
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True)
# eval_loader = DataLoader(eval_dataset,
# batch_size=args.batch_size,
# shuffle=True,
# num_workers=args.num_workers,
# pin_memory=True)
# Init logger
logger = Logger(args.log_dir)
p_log = Path(args.log_dir)
forward_time = 0
backward_time = 0
# Start training
# ============================================================
data_iter = iter(train_loader)
iter_per_epoch = len(train_loader)
criterion = eval(args.loss)(args)
# build criterion
optimizer = eval('torch.optim.' + args.optimizer)(model.parameters(), args.lr, weight_decay=args.weight_decay)
total_loss = 0
total_epe = 0
total_loss_levels = [0] * args.num_levels
total_epe_levels = [0] * args.num_levels
# training
# ============================================================
for step in range(1, args.total_step + 1):
# Reset the data_iter
if (step) % iter_per_epoch == 0: data_iter = iter(train_loader)
# Load Data
# ============================================================
# data, target = next(data_iter)
if args.mixup:
data, target = mixup(data_iter, args.mixup_alpha, args.mixup_prb)
else:
data, target = next(data_iter)
# shape: B,3,H,W
squeezer = partial(torch.squeeze, dim=2)
# shape: B,2,H,W
data, target = [d.to(args.device) for d in data], [t.to(args.device) for t in target]
# print(f'datalist={len(data[0])}')
x1_raw = data[0][:, :, 0, :, :]
x2_raw = data[0][:, :, 1, :, :]
if data[0].size(0) != args.batch_size: continue
flow_gt = target[0]
# Forward Pass
# ============================================================
t_forward = time.time()
flows, summaries = model(data[0])
# print(flows.shape)
forward_time += time.time() - t_forward
# Compute Loss
# ============================================================
loss, epe, loss_levels, epe_levels = criterion(flows, flow_gt)
total_loss += loss.item()
total_epe += epe.item()
for l, (loss_, epe_) in enumerate(zip(loss_levels, epe_levels)):
total_loss_levels[l] += loss_.item()
total_epe_levels[l] += epe_.item()
# if args.loss == 'L1':
# loss = L1loss(flow_gt, output_flow)
# elif args.loss == 'PyramidL1':
# loss = robust_training_loss(args, flows, flow_gt_pyramid)
# elif args.loss == 'L2':
# loss = L2loss(flow_gt, output_flow)
# elif args.loss == 'PyramidL2':
# loss = training_loss(args, flows, flow_gt_pyramid)
# backward
# ============================================================
t_backward = time.time()
optimizer.zero_grad()
loss.backward()
optimizer.step()
backward_time += time.time() - t_backward
# Collect Summaries & Output Logs
# ============================================================
if step % args.summary_interval == 0:
# Scalar Summaries
# ============================================================
logger.scalar_summary('lr', optimizer.param_groups[0]['lr'], step)
logger.scalar_summary('loss', total_loss / step, step)
logger.scalar_summary('EPE', total_epe / step, step)
for l, (loss_, epe_) in enumerate(zip(loss_levels, epe_levels)):
logger.scalar_summary(f'loss_lv{l}', total_loss_levels[l] / step, step)
logger.scalar_summary(f'EPE_lv{l}', total_epe_levels[l] / step, step)
# Image Summaries
# ============================================================
B = flows[0].size(0)
vis_batch = []
for b in range(B):
# batch = [np.array(
# F.upsample(flows[l][b].cpu().unsqueeze(0),
# scale_factor=2 ** ((len(flows) - l + 1))).detach().squeeze(
# 0)).transpose(1, 2, 0) for l in range(len(flows) - 1)]
batch = [np.array(
F.interpolate(flows[l][b].unsqueeze(0),
scale_factor=2 ** (len(flows) - l + 1)).detach().squeeze(
0).cpu()).transpose(1, 2, 0) for l in range(len(flows) - 1)]
# for i in batch:
# print(i.shape)
# print(flows[-1][b].detach().cpu().numpy().transpose(1,2,0))
# print(flow_gt[b].detach().cpu().numpy().transpose(1,2,0).shape)
vis = batch + [flows[-1][b].detach().cpu().numpy().transpose(1, 2, 0),
flow_gt[b].detach().cpu().numpy().transpose(1, 2, 0)]
vis = np.concatenate(list(map(vis_flow, vis)), axis=1)
vis_batch.append(vis.transpose(2, 0, 1))
logger.image_summary(f'flow', vis_batch, step)
# for l, x2_warp in enumerate(summaries['x2_warps']):
# out = [i.squeeze(0) for i in np.split(np.array(x2_warp.data).transpose(0,2,3,1), B, axis = 0)]
# for i in out:
# print(i.shape)
# logger.image_summary('tgt_warp', [i.squeeze(0) for i in np.split(np.array(x2_warp.data).transpose(0,2,3,1), B, axis = 0)], step)
# for l, flow in enumerate(flows):
# flow_batchs[0], flow_batchs[1], flow_batchs[2] = [vis_flow(i.squeeze()) for i in np.split(np.array(F.upsample(flow, 2 ** (6-l)).transpose(0,2,3,1)), B, axis = 0)]
# flow_vis = [vis_flow(i.squeeze()) for flow in flows for i in np.split(np.array(flow.data).transpose(0,2,3,1), B, axis = 0)][:min(B, args.max_output)]
# for layer_idx, flow in enumerate(flows):
# flow_vis =
# # flow_gt_vis = [vis_flow(i.squeeze()) for i in np.split(np.array(flow_gt_pyramid[layer_idx].data).transpose(0,2,3,1), B, axis = 0)][:min(B, args.max_output)]
# logger.image_summary(f'flow-lv{layer_idx}', flow_vis, step)
logger.image_summary('src & tgt', [np.concatenate([i.squeeze(0), j.squeeze(0)], axis=1) for i, j in
zip(np.split(np.array(x1_raw.data.cpu()), B,
axis=0),
np.split(np.array(x2_raw.data.cpu()), B,
axis=0))],
step)
# save model
if step % args.checkpoint_interval == 0:
torch.save(model.state_dict(), str(p_log / f'{step}.pkl'))
# print log
if step % args.log_interval == 0:
print(
f'Step [{step}/{args.total_step}], Loss: {total_loss / step:.4f}, EPE: {total_epe / step:.4f}, Forward: {forward_time / step * 1000} ms, Backward: {backward_time / step * 1000} ms')
if step % args.gif_interval == 0:
...
logger.close_summary()
def pred(args):
# Get environment
# Build Model
# ============================================================
print("start pred")
time_back = time.time()
model = Net(args).to(args.device)
model.load_state_dict(torch.load(args.load))
cstime = time.time() - time_back
print("初始化耗时%fs" % cstime)
# Load Data
# ============================================================
x1_raw, x2_raw = map(imageio.imread, args.input)
class StaticCenterCrop(object):
def __init__(self, image_size, crop_size):
self.th, self.tw = crop_size
self.h, self.w = image_size
print(self.th, self.tw, self.h, self.w)
def __call__(self, img):
return img[(self.h - self.th) // 2:(self.h + self.th) // 2, (self.w - self.tw) // 2:(self.w + self.tw) // 2,
:]
x1_raw = np.array(x1_raw)
x2_raw = np.array(x2_raw)
# if args.crop_shape is not None:
# cropper = StaticCenterCrop(x1_raw.shape[:2], args.crop_shape)
# x1_raw = cropper(x1_raw)
# x2_raw = cropper(x2_raw)
# if args.resize_shape is not None:
# resizer = partial(cv2.resize, dsize = (0,0), dst = args.resize_shape)
# x1_raw, x2_raw = map(resizer, [x1_raw, x2_raw])
# elif args.resize_scale is not None:
# resizer = partial(cv2.resize, dsize = (0,0), fx = args.resize_scale, fy = args.resize_scale)
# x1_raw, x2_raw = map(resizer, [x1_raw, x2_raw])
# pad to multiples of 64
H, W = x1_raw.shape[:2]
# print(x1_raw.shape)
x1_raw = np.pad(x1_raw, ((0, (64 - H % 64) if H % 64 else 0), (0, (64 - W % 64) if H % 64 else 0), (0, 0)),
mode='constant')
x2_raw = np.pad(x2_raw, ((0, (64 - H % 64) if H % 64 else 0), (0, (64 - W % 64) if H % 64 else 0), (0, 0)),
mode='constant')
x1_raw = x1_raw[np.newaxis, :, :, :].transpose(0, 3, 1, 2)
x2_raw = x2_raw[np.newaxis, :, :, :].transpose(0, 3, 1, 2)
x = np.stack([x1_raw, x2_raw], axis=2)
x = torch.Tensor(x).to(args.device)
# Forward Pass
# ============================================================
# print(x.shape)
with torch.no_grad():
flows, summaries = model(x)
flow = flows[-1].cpu()
# print(flow.shape)
flow = np.array(flow.data).transpose(0, 2, 3, 1).squeeze(0)
# flow = flow[[[1, 0]]]
# print(flow.shape)
hstime = time.time() - time_back
print("预测耗时%fs" % hstime)
save_flow(args.output, flow)
flow_vis = vis_flow(flow)
imageio.imwrite(args.output.replace('.flo', '.png'), flow_vis)
import matplotlib.pyplot as plt
plt.imshow(flow_vis)
plt.show()
def test(args):
print('load model...')
model = Net(args).to(args.device)
model.load_state_dict(torch.load(args.load))
print('build eval dataset...')
test_dataset = eval(args.dataset)(args.dataset_dir, 'test')
test_loader = DataLoader(test_dataset,
batch_size=1,
shuffle=True,
num_workers=args.num_workers,
pin_memory=True)
total_batches = len(test_loader)
# logs
# ============================================================
time_logs = []
total_epe = 0
for batch_idx, (data, target) in enumerate(test_loader):
# Forward Pass
# ============================================================
t_start = time.time()
data, target = [d.to(args.device) for d in data], [t.to(args.device) for t in target]
with torch.no_grad():
flows, summaries = model(data[0])
time_logs.append(time.time() - t_start)
# Compute EPE
# ============================================================
print(flows.shape)
flow = flows[-1].cpu()
flow = np.array(flow.data).transpose(0, 2, 3, 1).squeeze(0)
print(flow.shape)
targetn = target[0].cpu()
targetn = np.array(targetn).transpose(0, 2, 3, 1).squeeze(0)
print(targetn.shape)
epe = EPEp(flow, targetn, args)
total_epe += epe.item()
# print(f'total_epe={total_epe} batch_idx={batch_idx}')
print(f'[{batch_idx}/{total_batches}] Time: {time_logs[batch_idx]:.2f}s EPE:{total_epe / (batch_idx + 1)}')
if __name__ == '__main__':
main()