-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmultiview_video_alignment.py
352 lines (299 loc) · 12.6 KB
/
multiview_video_alignment.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
import torch
import torch.nn as nn
import torch.functional as F
import numpy as np
import os, time, random, sys, pickle
import torchvision.models as models
import argparse
from torchvision.utils import make_grid
from utils import seed_everything, distance, alignment_error, cycle_error, kendalls_tau
from utils import to_tensor, normalize
import matplotlib.pyplot as plt
import glob, timm
from data_pipeline.trajectory_dataset import TrajectoryDataset, TripletDataset
from torch.utils.data import DataLoader
parser = argparse.ArgumentParser()
parser.add_argument('--seed', type=int, default=42)
# data params
parser.add_argument('--data', type=str, default="panda")
# for unseen, set train_views to (6 for pick), (5 for mc), and (3 for can & lift)
# -1 or 0 means all trained on all views (seen)
parser.add_argument('--train_views', type=int, default=-1)
# -1 means split data to train/val/test by default
parser.add_argument('--train_videos', type=int, default=-1)
parser.add_argument('--batch_size', type=int, default=1)
# model structure params
parser.add_argument('--model', type=str, default="vit_3dtrl_tiny_patch16_224")
parser.add_argument('--pred_depth', type=int, default=1)
parser.add_argument('--pred_campos_from', type=str, default="both-sep")
# parser.add_argument('--embed_dim', type=int, default=768)
parser.add_argument('--pretrained', action='store_true')
#
parser.add_argument('--test', action='store_true')
# misc
parser.add_argument('--note', type=str, default="")
# loss params
parser.add_argument('--lr', type=float, default=1e-6) # main lr
## for distributed
parser.add_argument('--distributed', type=int, default=0)
parser.add_argument('--address', type=str, default="localhost")
parser.add_argument('--port', type=str, default="64751")
parser.add_argument('--gpus', type=int, default=1)
parser.add_argument('--xpar', action='store_true') # close progress bar
args = parser.parse_args()
seed_everything(args.seed)
torch.set_num_threads(8)
from PIL import Image
dataset = args.data
# dataset = "Pandav1_Train_Frame" # "MineRLReachGoalGroundStatic-v0_Frame" "can_mg" "can_mh" "lift_mh" "lift_ph"
if "panda" in dataset.lower():
dataset = "Pandav1_Train_Frame"
train_trs = [f"tr_{x:04d}" for x in range(10)]
eval_trs = [f"tr_{x:04d}" for x in range(10, 15)]
test_trs = [f"tr_{x:04d}" for x in range(15, 20)]
max_views = 9
window = 3
elif "mine" in dataset.lower():
dataset = "MineRLReachGoalGroundStatic-v0_Frame"
train_trs = [f"tr_{x:04d}" for x in range(4)]
eval_trs = [f"tr_{x:04d}" for x in range(4, 6)]
test_trs = [f"tr_{x:04d}" for x in range(6, 8)]
max_views = 8
window = 3
elif "lift_mh" in dataset.lower() or "can_mh" in dataset.lower():
train_trs = [f"tr_{x:04d}" for x in range(200)]
eval_trs = [f"tr_{x:04d}" for x in range(200, 250)]
test_trs = [f"tr_{x:04d}" for x in range(250, 300)]
max_views = 4
window = 2
elif "pouring" in dataset.lower():
data_dir = f"/home/jishang/ftpv_dataset/{dataset}"
train_items = pickle.load(open(os.path.join(data_dir, "train.pkl"), "rb"))[:-10]
val_items = train_items[-10:]
test_items = pickle.load(open(os.path.join(data_dir, "val.pkl"), "rb"))
def get_name(item):
name = item["name"]
name = name.split("_real_")[0]
return name
assert args.train_videos < len(train_items)
if args.train_videos == -1:
train_trs = [get_name(item) for item in train_items]
else:
train_trs = [get_name(item) for item in train_items][:args.train_videos]
eval_trs = [get_name(item) for item in val_items]
test_trs = [get_name(item) for item in test_items]
max_views = 1
window = 3
else:
raise NotImplementedError()
if args.train_views < 0 or args.train_views > max_views:
args.train_views = max_views
train_views = eval_views = range(args.train_views)
if args.train_views < max_views:
test_views = range(args.train_views, max_views)
else:
test_views = range(max_views)
args.note = f"views{len(train_views):02d}-{args.note}"
if args.train_videos > -1:
args.note += f"videos{args.train_videos}"
device = "cuda:0"
# one item is a trajectory with multiple views
# this iterates over trajectory indices
train_dataset = TrajectoryDataset(dataset, train_trs, train_views, window, config=args)
train_loader = DataLoader(train_dataset, batch_size=1, pin_memory=True, num_workers=4, drop_last=False, sampler=None, shuffle=True)
from model.vit import ViT
import model.vit_3dtrl as vit_3dtrl
import model.tnt_3dtrl as tnt_3dtrl
import model.swin_3dtrl as swin_3dtrl
from utils import *
def tc_loss(z1, z2, detach=False):
T, D = z1.size(0), z1.size(1)
dist_mat_ori = cos_distance(z1[0:1].expand_as(z2).reshape(-1, D), z2.reshape(-1, D))
dist_mat = torch.exp(dist_mat_ori*10)
# print("distmat", dist_mat.size(), dist_mat)
return -(torch.log(dist_mat[0]) - torch.log(torch.sum(dist_mat))), dist_mat_ori
def simple_tc_loss(z1, z2):
alpha = 0.3
anc1, pos1, neg1 = z1[..., 0, :], z1[..., 1, :], z1[..., 2, :]
anc2, pos2, neg2 = z2[..., 0, :], z1[..., 1, :], z1[..., 2, :]
match = distance(anc1, anc2) + distance(pos1, pos2)
attract = distance(anc1, pos1) + distance(anc2, pos2) + distance(anc1, pos2) + distance(anc2, pos1)
repulsion = - distance(anc1, neg2) - distance(anc2, neg1) - distance(anc1, neg1) - distance(anc2, neg2)
return torch.clamp(attract + match + repulsion + alpha*100, 0) #distance(z1[0:1].expand_as(z2), z2)
# load multiple views from single trajectory
# this loads actual frames
def load_data(tr, sample_indices, view_range):
base_path = f"/home/jishang/ftpv_dataset/{dataset}/{tr}"
x_f, x_ts = [], [] # l, l*8 flattened
for i in sample_indices:
f = Image.open(f"{base_path}/fpv/{i:06d}.jpg")
f = f.resize((224, 224), Image.BICUBIC)
x_f.append(to_tensor(f))
for view in view_range:
for i in sample_indices:
t = Image.open(f"{base_path}/tpv/view_{view:04d}/{i:06d}.jpg")
t = t.resize((224, 224), Image.BICUBIC)
x_ts.append(to_tensor(t))
x_f = torch.stack([normalize(x) for x in x_f])
x_t = torch.stack([normalize(x) for x in x_ts]).reshape(len(view_range), len(sample_indices), 3, 224, 224)
# print (x_f.size(), x_t.size())
# print (torch.sum(torch.abs(x_t[0,1]-x_t[0,2])), torch.sum(torch.abs(x_t[0,1]-x_t[1,1]))) # check resize correct
x = torch.cat((x_f.unsqueeze(0), x_t))
return x # view, sample (anchor/pos + neg), C, H, W
if not args.test:
if "3dtrl_" in args.model:
if "vit" in args.model:
pkg = vit_3dtrl
elif "swin" in args.model:
pkg = swin_3dtrl
elif "tnt" in args.model:
pkg = tnt_3dtrl
if "swin" in args.model:
m3d = pkg.__dict__[args.model](
global_pool="avg",
npl_depth=args.n_layer,
num_classes=1000)
else:
m3d = pkg.__dict__[args.model](
global_pool=None,
npl_depth=args.n_layer,
num_classes=1000)
else:
m3d = timm.create_model(
args.model,
pretrained=args.pretrained,
num_classes=1000,
global_pool=None)
setattr(m3d, "name", f"{args.model}-{dataset}-{args.note}")
print (m3d.name)
print ("-------------")
print (args)
from torch.optim import Adam
optimizer = Adam(m3d.parameters(), lr=args.lr)
num_negatives = 1
num_positives = 1
m3d.to(device)
best = 0.999
best_ep = -1
from tqdm import tqdm
for ep in range(100):
# train
m3d.train()
losses = []
dist_mats = []
itr = tqdm(train_loader, leave=False) if not args.xpar else train_loader
for tr in itr:
tr = tr.squeeze(0)
frame_dataset = TripletDataset(data=tr, window=window, n_pos=num_positives, n_neg=num_negatives)
frame_dataloader = DataLoader(frame_dataset, batch_size=args.batch_size, pin_memory=True, num_workers=4, drop_last=False, sampler=None, shuffle=True)
for x1, x2 in frame_dataloader:
B, T, C, H, W = x1.size() # T=3
x1 = x1.view(B*T, C, H, W).to(device)
x2 = x2.view(B*T, C, H, W).to(device)
s1 = m3d.forward_features(x1)
s2 = m3d.forward_features(x2)
if "swin" in args.model:
s1, s2 = s1.mean(dim=1).view(B,T,-1), s2.mean(dim=1).view(B,T,-1)
else:
s1, s2 = s1[:, 0].view(B, T, -1), s2[:, 0].view(B, T, -1)
# print (s1.size(), s2.size())
loss = simple_tc_loss(s1, s2).mean()
optimizer.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(m3d.parameters(), max_norm=10.0)
optimizer.step()
losses.append(loss.item())
# eval
m3d.eval()
eval_align_errors = []
eval_cycle_errors = []
eval_taus = []
for tr in eval_trs:
path = f"/home/jishang/ftpv_dataset/{dataset}/{tr}/fpv"
imgs = sorted([ p for p in glob.glob(path + "/*") if '.jpg' in p or '.png' in p])
# print (imgs)
n_frames = len(imgs)
indices = list(range(0, n_frames, 1))
x = load_data(tr, indices, eval_views)
with torch.no_grad():
for i in range(x.size(0)):
for j in range(i+1, x.size(0)):
x1, x2 = x[i].to(device), x[j].to(device)
s1, s2 = m3d.forward_features(x1), m3d.forward_features(x2)
if "swin" in args.model:
s1, s2 = s1.mean(dim=1).cpu(), s2.mean(dim=1).cpu()
else:
s1, s2 = s1[:, 0].cpu(), s2[:, 0].cpu()
eval_align_errors.append(alignment_error(s1, s2))
eval_cycle_errors.append(cycle_error(s1, s2))
eval_taus.append(kendalls_tau(s1, s2))
a_error = np.mean(eval_align_errors)
c_error = np.mean(eval_cycle_errors)
tau = np.mean(eval_taus)
print (f"Ep {ep}, Eval Align Error: {a_error:.4f}, Eval Cycle Error: {c_error:.4f}, Loss {np.mean(losses):.4f}, Eval Kendall's Tau: {tau:.4f}")
if a_error < best:
torch.save(m3d.state_dict(), f"trained_models/{m3d.name}_best.pth")
best = a_error
best_ep = ep
else:
if ep - best_ep >= (10 if dataset == "pouring" else 10):
print (f"Best Ep: {best_ep}")
break
del m3d
# test
if "npl_" in args.model:
if "vit" in args.model:
if "mh" in args.__dict__ and args.mh is True:
pkg = vitnpl_mh
else:
pkg = vitnpl
elif "swin" in args.model:
pkg = swinnpl
elif "tnt" in args.model:
pkg = tntnpl
if "swin" in args.model:
m3d = pkg.__dict__[args.model](
global_pool="avg",
npl_depth=args.n_layer,
num_classes=1000)
else:
m3d = pkg.__dict__[args.model](
global_pool=None,
npl_depth=args.n_layer,
num_classes=1000)
else:
m3d = timm.create_model(
args.model,
pretrained=False,
num_classes=1000,
global_pool=None)
setattr(m3d, "name", f"{args.model}-{dataset}-{args.note}")
m3d.to(device)
m3d.load_state_dict(torch.load(f"trained_models/{m3d.name}_best.pth"))
m3d.eval()
test_align_errors = []
test_cycle_errors = []
test_taus = []
for tr in test_trs:
path = f"/home/jishang/ftpv_dataset/{dataset}/{tr}/fpv"
imgs = sorted([ p for p in glob.glob(path + "/*") if '.jpg' in p or '.png' in p])
# print (imgs)
n_frames = len(imgs)
indices = list(range(0, n_frames, 1))
x = load_data(tr, indices, test_views)
with torch.no_grad():
for i in range(x.size(0)):
for j in range(i+1, x.size(0)):
x1, x2 = x[i].to(device), x[j].to(device)
s1, s2 = m3d.forward_features(x1), m3d.forward_features(x2)
if "swin" in args.model:
s1, s2 = s1.mean(dim=1).cpu(), s2.mean(dim=1).cpu()
else:
s1, s2 = s1[:, 0].cpu(), s2[:, 0].cpu()
test_align_errors.append(alignment_error(s1, s2))
test_cycle_errors.append(cycle_error(s1, s2))
test_taus.append(kendalls_tau(s1, s2))
a_error = np.mean(test_align_errors)
c_error = np.mean(test_cycle_errors)
tau = np.mean(test_taus)
print (f"Test Align Error: {a_error:.4f}, Test Cycle Error: {c_error:.4f}, Test Kendall's Tau: {tau:.4f}")