-
Notifications
You must be signed in to change notification settings - Fork 95
/
Copy pathtrain_git.py
executable file
·181 lines (156 loc) · 7.14 KB
/
train_git.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
"""
It is not yet cleaned up to be runnable in conjunction,
but to provide a reference algorithm.
For easier training, three-stage learning is recommended.
Stage-1 & 2: w.o recurrence
(w_ST, w_LT, w_Flow = 0, 0, 0,
t_stride, sample_duration, sample_frames = 3, 13, 1):
1: temporal aggregation with 1 support and 1 target frame
2: temporal aggregation with 4 support and 1 target frame
Stage-3: with recurrence (w_ST, w_LT, w_Flow = 1, 1, 10,
t_stride, sample_duration, sample_frames = 3, 16, 4):
3: train with recurrence and short- and long-term temporal loss
"""
import torch
import torch.nn.functional as F
from torch.autograd import Variable
import numpy as np
import pytorch_ssim
from pytorch_misc import clip_grad_norm
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
# from networks.resample2d_package.modules.resample2d import Resample2d
from networks.resample2d_package.resample2d import Resample2d
import networks
class Object(object):
pass
""" Flownet """
args = Object()
args.rgb_max = 1.0
args.fp16 = False
FlowNet = networks.FlowNet2(args, requires_grad=False)
model_filename = os.path.join(
"pretrained_models", "FlowNet2_checkpoint.pth.tar")
checkpoint = torch.load(model_filename)
FlowNet.load_state_dict(checkpoint['state_dict'])
FlowNet = FlowNet.cuda()
""" Submodules """
flow_warping = Resample2d().cuda()
downsampler = nn.AvgPool2d((2, 2), stride=2).cuda()
def norm(t):
return torch.sum(t*t, dim=1, keepdim=True)
def repackage_hidden(h):
"""Wraps hidden states in new Variables, detach them from their history."""
if isinstance(h, torch.Tensor):
return h.detach()
else:
return tuple(repackage_hidden(v) for v in h)
def train_lstm_epoch(epoch, data_loader, model, criterion_L1, criterion_ssim,
optimizer, opt):
opt.w_ST, opt.w_LT, opt.w_Flow = 1.0, 1.0, 10.0
model.train()
ts = opt.t_stride
### start epoch
for i, (inputs, masks, _) in enumerate(data_loader):
# inputs: BxCxTxHxW
bs = inputs.size(0)
midx = (inputs.size(2) - 1)//2
inputs = 2.*inputs - 1. # [-1 1]
inverse_masks = 1.-masks
masked_inputs = inputs.clone()*inverse_masks
frame_i, frame_mi, frame_m = [], [], []
for tt in range(opt.sample_frames):
slices = [x*ts+tt for x in range(5)]
frame_i.append(to_var(inputs[:,:,midx*ts+tt,:,:]))
frame_mi.append(to_var(masked_inputs[:,:,slices,:,:]))
frame_m.append(to_var(masks[:,:,slices,:,:]))
optimizer.zero_grad()
lstm_state = None
ST_loss, LT_loss = 0, 0
RECON_loss, HOLE_loss = 0, 0
flow_loss = 0
### forward
prev_mask = frame_m[0][:,:,midx,:,:]
prev_ones = to_var(torch.ones(prev_mask.size()))
prev_feed = torch.cat([frame_mi[0][:,:,midx,:,:], prev_ones,
prev_ones*prev_mask], dim=1)
frame_o1, _, lstm_state, _ ,occs = model(
frame_mi[0], frame_m[0], lstm_state, prev_feed)
# Turned out it still works w.o.LSTM. May work with lstm_state = None
lstm_state = None if opt.no_lstm else repackage_hidden(lstm_state)
frame_o1 = frame_o1.squeeze(2)
RECON_loss += 1*criterion_L1(frame_o1, frame_i[0]) -
criterion_ssim(frame_o1, frame_i[0])
HOLE_loss += 5*criterion_L1(
frame_o1*frame_m[0][:,:,midx,:,:].expand_as(frame_o1),
frame_i[0]*frame_m[0][:,:,midx,:,:].expand_as(frame_o1)
)
frame_o = []
frame_o.append(frame_o1)
### if opt.sample_frames > 1 , Recurrence learning
for tt in range(1, opt.sample_frames):
frame_i1, frame_m1 = frame_i[tt-1], frame_m[tt-1]
frame_mi2 = frame_mi[tt]
frame_i2, frame_m2 = frame_i[tt], frame_m[tt]
frame_o1 = frame_o1.detach() if tt == 1 else frame_o2.detach()
prev_mask = to_var(torch.zeros(frame_m2[:,:,midx,:,:].size()))
prev_ones = to_var(torch.ones(prev_mask.size()))
prev_feed = torch.cat(
[frame_o1,prev_ones, prev_ones*prev_mask], dim=1)
frame_o2, _, lstm_state, _, occs, flow6_256 = model(
frame_mi2, frame_m2, lstm_state, prev_feed, None, 1)
if opt.loss_on_raw:
frame_o2_raw = frame_o2[1].squeeze(2)
frame_o2 = frame_o2[0]
frame_o2 = frame_o2.squeeze(2)
### detach from graph and avoid memory accumulation
lstm_state = None if opt.no_lstm else repackage_hidden(lstm_state)
frame_o.append(frame_o2)
RECON_loss += criterion_L1(frame_o2, frame_i2) -
criterion_ssim(frame_o2, frame_i2)
HOLE_loss += 5*criterion_L1(
frame_o2*frame_m2[:,:,midx,:,:].expand_as(frame_o2),
frame_i2*frame_m2[:,:,midx,:,:].expand_as(frame_i2)
)
if opt.loss_on_raw:
RECON_loss += criterion_L1(frame_o2_raw, frame_i2) -
criterion_ssim(frame_o2_raw, frame_i2)
HOLE_loss += 5*criterion_L1(
frame_o2_raw*frame_m2[:,:,midx,:,:].expand_as(frame_o2_raw),
frame_i2*frame_m2[:,:,midx,:,:].expand_as(frame_i2))
### short-term temporal loss
if opt.w_ST > 0:
flow_i21 = FlowNet(frame_i2, frame_i1)
warp_i1 = flow_warping(frame_i1, flow_i21)
warp_o1 = flow_warping(frame_o1, flow_i21)
noc_mask2 = torch.exp( -50. * torch.sum(
frame_i2 - warp_i1, dim=1).pow(2) ).unsqueeze(1)
ST_loss += criterion_L1(
frame_o2 * noc_mask2, warp_i1 * noc_mask2)
conf = (norm(frame_i2 - warp_i1) < 0.02).float()
flow_loss = criterion_L1(flow6_256 * conf, flow_i21 * conf)
warp_i1_ = flow_warping(frame_i1, flow6_256)
flow_loss += criterion_L1(warp_i1_ * conf, frame_i2 * conf)
warp_o1_ = flow_warping(frame_o1, flow6_256)
flow_loss += criterion_L1(
frame_o2 * conf, warp_o1_.detach() * conf)
if opt.w_LT > 0:
t1 = 0
for t2 in range(t1 + 2, opt.sample_frames):
frame_i1, frame_i2 = frame_i[t1], frame_i[t2]
frame_o1 = frame_o[t1].detach()
frame_o1.requires_grad = False
frame_o2 = frame_o[t2]
flow_i21 = FlowNet(frame_i2, frame_i1)
warp_i1 = flow_warping(frame_i1, flow_i21)
warp_o1 = flow_warping(frame_o1, flow_i21)
noc_mask2 = torch.exp(-50. * torch.sum(
frame_i2 - warp_i1, dim=1).pow(2) ).unsqueeze(1)
LT_loss += criterion_L1(
frame_o2 * noc_mask2, warp_i1 * noc_mask2)
overall_loss = (RECON_loss + GRAD_loss + HOLE_loss + opt.w_ST * ST_loss
+ opt.w_LT * LT_loss + opt.w_FLOW * flow_loss)
overall_loss.backward()
optimizer.step()
return overall_loss.data[0]