-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathutils_all.py
132 lines (122 loc) · 6.73 KB
/
utils_all.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
import os
from mmaction.apis import init_recognizer
import torch
import numpy as np
import argparse
import torch.nn.functional as F
import torch.nn as nn
class ReFiner(nn.Module):
def __init__(self, depth, channels) -> None:
super(ReFiner, self).__init__()
assert len(channels) - 1 == depth
assert channels[0] == 3
assert channels[-1] == 3
layers = []
for i in range(depth):
layers.append(self._make_plain_conv(channels[i], channels[i+1]))
layers.append(nn.BatchNorm3d(channels[i+1]))
layers.append(nn.ReLU(inplace=True))
self.layers = nn.Sequential(*layers)
def _make_plain_conv(self, in_channels, out_channels):
return nn.Conv3d(in_channels, out_channels, kernel_size=(3,3,3), stride=1, padding=(1,1,1))
def forward(self, x):
assert len(x.shape) == 5
return self.layers(x)
def shared_params():
parser = argparse.ArgumentParser()
parser.add_argument('--dataset', type=str, default='hmdb51', help='dataset to distill')
parser.add_argument('--ipc', type=int, default=10)
parser.add_argument('--T', type=int, default=8)
parser.add_argument('--size', type=int, default=112)
parser.add_argument('--inter-mode', type=str, default='duplicate', help='interpretation mode')
parser.add_argument('--gpu-id', type=str, default='0,1')
parser.add_argument('--world-size', default=1, type=int,
help='number of nodes for distributed training')
parser.add_argument('--rank', default=0, type=int,
help='node rank for distributed training')
parser.add_argument('--syn_data_path', type=str,
default='./syn_data', help='where to store synthetic data')
parser.add_argument('--syn_label_path', type=str,
default='./syn_label', help='where to store synthetic label')
parser.add_argument('--dist-backend', default='nccl', type=str,
help='distributed backend')
parser.add_argument('--wandb-project', type=str,
default='Temperature', help='wandb project name')
parser.add_argument('--wandb-api-key', type=str,
default=None, help='wandb api key')
parser.add_argument('--teacher-list', nargs='+', default=[])
parser.add_argument('--pre', default=False, action='store_true', help='pretrained model')
return parser
def adapt_params(args):
root = '/data0/chenyang/ViD/'
args.statistic_path = os.path.join(args.statistic_path, args.dataset)
if args.dataset == 'hmdb51':
args.size = 112
args.num_classes = 51
args.config = {'conv4': '/data0/chenyang/VDC/mmaction2/configs/recognition/conv3/conv3-hmdb51-8x112x112.py'}
args.checkpoint = {'conv4': '/data0/chenyang/VDC/mmaction2/work_dirs/conv3-hmdb51-8x112x112/best_acc_top1_epoch_130.pth'}
elif args.dataset == 'k400':
args.size = 56
args.num_classes = 400
args.config = {'conv4': '/data0/chenyang/VDC/mmaction2/configs/recognition/conv3/conv3-kinetics-8x56x56.py'}
args.checkpoint = {'conv4': '/data0/chenyang/VDC/mmaction2/work_dirs/conv3-kinetics-8x56x56/best_acc_top1_epoch_70.pth'}
elif args.dataset == 'ucf101':
args.size = 112
args.num_classes = 101
args.config = {'conv4': '/data0/chenyang/VDC/mmaction2/configs/recognition/conv3/conv3-ucf101-8x112x112.py', 'slowonly': '/data0/chenyang/VDC/mmaction2/configs/recognition/slowonly/slowonly-r18_8x8_ucf101-frame.py', 'r2plus1d': '/data0/chenyang/VDC/mmaction2/configs/recognition/r2plus1d/r2plus1d_r18_8xb8-8x8x1-180e_ucf101.py', 'i3d': '/data0/chenyang/VDC/mmaction2/configs/recognition/i3d/i3d_k400-pretrained-r18_8x8_ucf101_frame.py'}
args.checkpoint = {'conv4': '/data0/chenyang/VDC/mmaction2/work_dirs/conv3-ucf101-8x112x112/best_acc_top1_epoch_145.pth', 'slowonly': '/data0/chenyang/VDC/mmaction2/work_dirs/slowonly-r18_8x8_ucf101-frame/best_acc_top1_epoch_134.pth', 'r2plus1d': '/data0/chenyang/VDC/mmaction2/work_dirs/r2plus1d_r18_8xb8-8x8x1-180e_ucf101/best_acc_top1_epoch_41.pth', 'i3d': '/data0/chenyang/VDC/mmaction2/work_dirs/i3d_k400-pretrained-r18_8x8_ucf101_frame/best_acc_top1_epoch_48.pth'}
return args
def load_model(config, checkpoint, device='cpu', pretrained=False):
model = init_recognizer(config, checkpoint, device=device)
return model
def Interpolate(inputs, refiner=None, mode='duplicate', start=None):
assert len(inputs.shape) == 6
b, s, c, t, h, w = inputs.shape
# inputs = inputs.cpu()
if mode=='duplicate':
start = torch.randint(0, 9, (b,))
outputs = torch.stack([inputs,inputs],dim=4).reshape(b,s,c,t*2,h,w)
assert outputs[:,:,:,0].equal(outputs[:,:,:,1])
elif mode == 'linear':
# print(inputs.shape)
inputs = inputs.squeeze(1)
outputs = torch.stack([inputs[i] for i in range(b)], dim=0)
outputs = F.interpolate(outputs, size=(8,h,w), mode='trilinear', align_corners=False).unsqueeze(1)
# print(outputs.shape)
elif mode=='sample':
if start is None:
start = torch.randint(0, 9, (b,))
outputs = torch.stack([inputs[i, :, :, start[i]:start[i]+8] for i in range(b)], dim=0)
assert len(outputs.shape) == 6
assert outputs.shape[0] == b
elif mode=='sample_dup':
if start is None:
start = torch.randint(0, t-3, (b,))
sample_len = 4
outputs = torch.stack([inputs[i, :, :, start[i]:start[i]+sample_len] for i in range(b)], dim=0)
outputs = torch.stack([outputs,outputs],dim=4).reshape(b,s,c,2*sample_len,h,w)
# if not outputs[:,:,:,0].equal(outputs[:,:,:,1]):
# indices = torch.where(outputs[:,:,:,0] != outputs[:,:,:,1])
# print(indices)
# print(outputs[:,:,:,0][indices])
# print(outputs[:,:,:,1][indices])
assert torch.allclose(outputs[:,:,:,0], outputs[:,:,:,1])
elif mode=='sample_linear':
if start is None:
start = torch.randint(0, t-3, (b,))
sample_len = 4
inputs = inputs.squeeze(1)
outputs = torch.stack([inputs[i, :, start[i]:start[i]+sample_len] for i in range(b)], dim=0)
outputs = F.interpolate(outputs, size=(8,h,w), mode='trilinear', align_corners=False).unsqueeze(1)
elif mode == 'refine':
if start is None:
start = torch.randint(0, t-3, (b,))
sample_len = 4
outputs = torch.stack([inputs[i, :, :, start[i]:start[i]+sample_len] for i in range(b)], dim=0)
outputs = refiner(torch.stack([outputs,outputs],dim=4).reshape(b,s,c,2*sample_len,h,w))
elif mode=='none':
start = torch.randint(0, 9, (b,))
outputs = inputs
else:
assert 0, 'not Impletmented'
return outputs