-
Notifications
You must be signed in to change notification settings - Fork 54
/
Copy pathtest_on_flt.py
314 lines (252 loc) · 12.2 KB
/
test_on_flt.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
import time
import numpy as np
import timeit
import saverloader
from nets.raftnet import Raftnet
from nets.pips import Pips
import random
from utils.basic import print_, print_stats
import torch
import torch.nn as nn
from torch.utils.data import Dataset, DataLoader
from tensorboardX import SummaryWriter
import torch.nn.functional as F
from flyingthingsdataset import FlyingThingsDataset
import utils.basic
import utils.improc
import utils.test
from fire import Fire
device = 'cuda'
random.seed(125)
np.random.seed(125)
def run_dino(dino, d, sw):
rgbs = d['rgbs'].cuda().float() # B, S, C, H, W
occs = d['occs'].cuda().float() # B, S, 1, H, W
masks = d['masks'].cuda().float() # B, S, 1, H, W
trajs_g = d['trajs'].cuda().float() # B, S, N, 2
vis_g = d['visibles'].cuda().float() # B, S, N
valids = d['valids'].cuda().float() # B, S, N
B, S, C, H, W = rgbs.shape
B, S1, N, D = trajs_g.shape
assert(torch.sum(valids)==B*S*N)
# compute per-sequence visibility labels
vis_g = (torch.sum(vis_g, dim=1, keepdim=True) >= 4).float().repeat(1, S, 1)
_, S, C, H, W = rgbs.shape
trajs_e = utils.test.get_dino_output(dino, rgbs, trajs_g, vis_g)
ate = torch.norm(trajs_e - trajs_g, dim=-1) # B, S, N
ate_all = utils.basic.reduce_masked_mean(ate, valids)
ate_vis = utils.basic.reduce_masked_mean(ate, valids*vis_g)
ate_occ = utils.basic.reduce_masked_mean(ate, valids*(1.0-vis_g))
metrics = {
'ate_all': ate_all.item(),
'ate_vis': ate_vis.item(),
'ate_occ': ate_occ.item(),
}
if sw is not None and sw.save_this:
sw.summ_traj2ds_on_rgbs('inputs_0/orig_trajs_on_rgbs', trajs_g, utils.improc.preprocess_color(rgbs), cmap='winter', linewidth=2)
sw.summ_traj2ds_on_rgbs('outputs/trajs_on_rgbs', trajs_e[0:1], utils.improc.preprocess_color(rgbs[0:1]), cmap='spring', linewidth=2)
gt_rgb = utils.improc.preprocess_color(sw.summ_traj2ds_on_rgb('inputs_0_all/single_trajs_on_rgb', trajs_g[0:1], torch.mean(utils.improc.preprocess_color(rgbs[0:1]), dim=1), cmap='winter', frame_id=metrics['ate_all'], only_return=True, linewidth=2))
gt_black = utils.improc.preprocess_color(sw.summ_traj2ds_on_rgb('inputs_0_all/single_trajs_on_rgb', trajs_g[0:1], torch.ones_like(rgbs[0:1,0])*-0.5, cmap='winter', frame_id=metrics['ate_all'], only_return=True, linewidth=2))
sw.summ_traj2ds_on_rgb('outputs/single_trajs_on_gt_rgb', trajs_e[0:1], gt_rgb[0:1], cmap='spring', linewidth=2)
sw.summ_traj2ds_on_rgb('outputs/single_trajs_on_gt_black', trajs_e[0:1], gt_black[0:1], cmap='spring', linewidth=2)
return metrics
def run_pips(model, d, sw):
rgbs = d['rgbs'].cuda().float() # B, S, C, H, W
occs = d['occs'].cuda().float() # B, S, 1, H, W
masks = d['masks'].cuda().float() # B, S, 1, H, W
trajs_g = d['trajs'].cuda().float() # B, S, N, 2
vis_g = d['visibles'].cuda().float() # B, S, N
valids = d['valids'].cuda().float() # B, S, N
B, S, C, H, W = rgbs.shape
assert(C==3)
B, S, N, D = trajs_g.shape
assert(torch.sum(valids)==B*S*N)
# compute per-sequence visibility labels
vis_g = (torch.sum(vis_g, dim=1, keepdim=True) >= 4).float().repeat(1, S, 1)
_, S, C, H, W = rgbs.shape
preds, preds_anim, vis_e, stats = model(trajs_g[:,0], rgbs, iters=6, trajs_g=trajs_g, vis_g=vis_g, valids=valids, sw=sw)
ate = torch.norm(preds[-1] - trajs_g, dim=-1) # B, S, N
ate_all = utils.basic.reduce_masked_mean(ate, valids)
ate_vis = utils.basic.reduce_masked_mean(ate, valids*vis_g)
ate_occ = utils.basic.reduce_masked_mean(ate, valids*(1.0-vis_g))
metrics = {
'ate_all': ate_all.item(),
'ate_vis': ate_vis.item(),
'ate_occ': ate_occ.item(),
}
trajs_e = preds[-1]
if sw is not None and sw.save_this:
sw.summ_traj2ds_on_rgbs('inputs_0/orig_trajs_on_rgbs', trajs_g, utils.improc.preprocess_color(rgbs), cmap='winter', linewidth=2)
sw.summ_traj2ds_on_rgbs('outputs/trajs_on_rgbs', trajs_e[0:1], utils.improc.preprocess_color(rgbs[0:1]), cmap='spring', linewidth=2)
gt_rgb = utils.improc.preprocess_color(sw.summ_traj2ds_on_rgb('inputs_0_all/single_trajs_on_rgb', trajs_g[0:1], torch.mean(utils.improc.preprocess_color(rgbs[0:1]), dim=1), cmap='winter', frame_id=metrics['ate_all'], only_return=True, linewidth=2))
gt_black = utils.improc.preprocess_color(sw.summ_traj2ds_on_rgb('inputs_0_all/single_trajs_on_rgb', trajs_g[0:1], torch.ones_like(rgbs[0:1,0])*-0.5, cmap='winter', frame_id=metrics['ate_all'], only_return=True, linewidth=2))
sw.summ_traj2ds_on_rgb('outputs/single_trajs_on_gt_rgb', trajs_e[0:1], gt_rgb[0:1], cmap='spring', linewidth=2)
sw.summ_traj2ds_on_rgb('outputs/single_trajs_on_gt_black', trajs_e[0:1], gt_black[0:1], cmap='spring', linewidth=2)
# animate_traj2ds_on_rgbs
rgb_vis = []
black_vis = []
for trajs_e_ in preds_anim:
rgb_vis.append(sw.summ_traj2ds_on_rgb('', trajs_e_[0:1], gt_rgb, only_return=True, cmap='coolwarm'))
black_vis.append(sw.summ_traj2ds_on_rgb('', trajs_e_[0:1], gt_black, only_return=True, cmap='coolwarm'))
sw.summ_rgbs('outputs/animated_trajs_on_black', black_vis)
sw.summ_rgbs('outputs/animated_trajs_on_rgb', rgb_vis)
return metrics
def run_raft(raft, d, sw):
rgbs = d['rgbs'].cuda().float() # B, S, C, H, W
occs = d['occs'].cuda().float() # B, S, 1, H, W
masks = d['masks'].cuda().float() # B, S, 1, H, W
trajs_g = d['trajs'].cuda().float() # B, S, N, 2
vis_g = d['visibles'].cuda().float() # B, S, N
valids = d['valids'].cuda().float() # B, S, N
B, S, C, H, W = rgbs.shape
assert(C==3)
B, S, N, D = trajs_g.shape
assert(torch.sum(valids)==B*S*N)
# compute per-sequence visibility labels
vis_g = (torch.sum(vis_g, dim=1, keepdim=True) >= 4).float().repeat(1, S, 1)
_, S, C, H, W = rgbs.shape
prep_rgbs = utils.improc.preprocess_color(rgbs)
flows_e = []
for s in range(S-1):
rgb0 = prep_rgbs[:,s]
rgb1 = prep_rgbs[:,s+1]
flow, _ = raft(rgb0, rgb1, iters=32)
flows_e.append(flow)
flows_e = torch.stack(flows_e, dim=1) # B, S-1, 2, H, W
coords = []
coord0 = trajs_g[:,0] # B, N, 2
coords.append(coord0)
coord = coord0.clone()
for s in range(S-1):
delta = utils.samp.bilinear_sample2d(
flows_e[:,s], coord[:,:,0], coord[:,:,1]).permute(0,2,1) # B, N, 2, forward flow at the discrete points
coord = coord + delta
coords.append(coord)
trajs_e = torch.stack(coords, dim=1) # B, S, N, 2
ate = torch.norm(trajs_e - trajs_g, dim=-1) # B, S, N
ate_all = utils.basic.reduce_masked_mean(ate, valids)
ate_vis = utils.basic.reduce_masked_mean(ate, valids*vis_g)
ate_occ = utils.basic.reduce_masked_mean(ate, valids*(1.0-vis_g))
metrics = {
'ate_all': ate_all.item(),
'ate_vis': ate_vis.item(),
'ate_occ': ate_occ.item(),
}
if sw is not None and sw.save_this:
sw.summ_traj2ds_on_rgbs('inputs_0/orig_trajs_on_rgbs', trajs_g, utils.improc.preprocess_color(rgbs), cmap='winter', linewidth=2)
sw.summ_traj2ds_on_rgbs('outputs/trajs_on_rgbs', trajs_e[0:1], utils.improc.preprocess_color(rgbs[0:1]), cmap='spring', linewidth=2)
gt_rgb = utils.improc.preprocess_color(sw.summ_traj2ds_on_rgb('inputs_0_all/single_trajs_on_rgb', trajs_g[0:1], torch.mean(utils.improc.preprocess_color(rgbs[0:1]), dim=1), cmap='winter', frame_id=metrics['ate_all'], only_return=True, linewidth=2))
gt_black = utils.improc.preprocess_color(sw.summ_traj2ds_on_rgb('inputs_0_all/single_trajs_on_rgb', trajs_g[0:1], torch.ones_like(rgbs[0:1,0])*-0.5, cmap='winter', frame_id=metrics['ate_all'], only_return=True, linewidth=2))
sw.summ_traj2ds_on_rgb('outputs/single_trajs_on_gt_rgb', trajs_e[0:1], gt_rgb[0:1], cmap='spring', linewidth=2)
sw.summ_traj2ds_on_rgb('outputs/single_trajs_on_gt_black', trajs_e[0:1], gt_black[0:1], cmap='spring', linewidth=2)
return metrics
def main(
exp_name='flt',
B=1,
S=8,
N=16,
modeltype='pips',
init_dir='reference_model',
stride=4,
log_dir='logs_test_on_flt',
dataset_location='/data/flyingthings',
max_iters=0, # auto-select based on dataset
log_freq=100,
shuffle=False,
subset='all',
crop_size=(384,512), # the raw data is 540,960
use_augs=False,
):
# the idea in this file is to evaluate on flyingthings++
assert(modeltype=='pips' or modeltype=='raft' or modeltype=='dino')
## autogen a name
model_name = "%d_%d_%d_%s" % (B, S, N, modeltype)
if use_augs:
model_name += "_A"
model_name += "_%s" % exp_name
import datetime
model_date = datetime.datetime.now().strftime('%H:%M:%S')
model_name = model_name + '_' + model_date
print('model_name', model_name)
writer_t = SummaryWriter(log_dir + '/' + model_name + '/t', max_queue=10, flush_secs=60)
def worker_init_fn(worker_id):
np.random.seed(np.random.get_state()[1][0] + worker_id)
test_dataset = FlyingThingsDataset(
dataset_location=dataset_location,
dset='TEST', subset=subset,
use_augs=use_augs,
N=N, S=S,
crop_size=crop_size)
test_dataloader = DataLoader(
test_dataset,
batch_size=B,
shuffle=shuffle,
num_workers=24,
worker_init_fn=worker_init_fn,
drop_last=True)
test_iterloader = iter(test_dataloader)
global_step = 0
if modeltype=='pips':
model = Pips(S=S, stride=stride).cuda()
_ = saverloader.load(init_dir, model)
model.eval()
elif modeltype=='raft':
model = Raftnet(ckpt_name='../../RAFT/models/raft-things.pth').cuda()
model.eval()
elif modeltype=='dino':
patch_size = 8
model = torch.hub.load('facebookresearch/dino:main', 'dino_vits%d' % patch_size).cuda()
model.eval()
else:
assert(False) # need to choose a valid modeltype
n_pool = 10000
ate_all_pool_t = utils.misc.SimplePool(n_pool, version='np')
ate_vis_pool_t = utils.misc.SimplePool(n_pool, version='np')
ate_occ_pool_t = utils.misc.SimplePool(n_pool, version='np')
if max_iters==0:
max_iters = len(test_dataloader)
print('setting max_iters', max_iters)
while global_step < max_iters:
read_start_time = time.time()
global_step += 1
sw_t = utils.improc.Summ_writer(
writer=writer_t,
global_step=global_step,
log_freq=log_freq,
fps=5,
scalar_freq=int(log_freq/2),
just_gif=True)
gotit = (False,False)
while not all(gotit):
try:
sample, gotit = next(test_iterloader)
except StopIteration:
test_iterloader = iter(test_dataloader)
sample, gotit = next(test_iterloader)
read_time = time.time()-read_start_time
iter_start_time = time.time()
with torch.no_grad():
if modeltype=='pips':
metrics = run_pips(model, sample, sw_t)
elif modeltype=='raft':
metrics = run_raft(model, sample, sw_t)
elif modeltype=='dino':
metrics = run_dino(model, sample, sw_t)
else:
assert(False) # need to choose a valid modeltype
if metrics['ate_all'] > 0:
ate_all_pool_t.update([metrics['ate_all']])
if metrics['ate_vis'] > 0:
ate_vis_pool_t.update([metrics['ate_vis']])
if metrics['ate_occ'] > 0:
ate_occ_pool_t.update([metrics['ate_occ']])
sw_t.summ_scalar('pooled/ate_all', ate_all_pool_t.mean())
sw_t.summ_scalar('pooled/ate_vis', ate_vis_pool_t.mean())
sw_t.summ_scalar('pooled/ate_occ', ate_occ_pool_t.mean())
iter_time = time.time()-iter_start_time
print('%s; step %06d/%d; rtime %.2f; itime %.2f, ate_vis = %.2f, ate_occ = %.2f' % (
model_name, global_step, max_iters, read_time, iter_time,
ate_vis_pool_t.mean(), ate_occ_pool_t.mean()))
writer_t.close()
if __name__ == '__main__':
Fire(main)