-
Notifications
You must be signed in to change notification settings - Fork 2
/
sia_extract_param.py
230 lines (158 loc) · 7.51 KB
/
sia_extract_param.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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Tue Apr 23 15:15:39 2019
@author: luoyao
"""
#!/usr/bin/env python3
# coding: utf-8
import torchvision
import torch
import torch.nn as nn
import torch.utils.data as data
import torchvision.transforms as transforms
import torch.backends.cudnn as cudnn
import time
import numpy as np
from siamese_utils import Normalize, SiaTestDataset, reconstruct_vertex, reconstruct_vertex_shp
import scipy.io as sio
import os.path as osp
import os
from io_utils import _load, _dump, mkdir
class sia_net(nn.Module):
def __init__(self , model):
super(sia_net, self).__init__()
#取掉model的后两层
self.fc1 = nn.Sequential(
nn.Sequential(*list(model.children())[:-2]),
nn.AdaptiveAvgPool2d(1))
# self.relu = nn.ReLU(inplace=True)
self.fc1_0 = nn.Sequential(
nn.Linear(2048, 1024),
nn.Linear(1024, 512))
self.fc1_1 = nn.Sequential(
nn.Linear(2048, 62))
def forward_once(self, x):
x = self.fc1(x)
x = x.view(x.size()[0], -1)
feature = self.fc1_0(x) #feature
# feature = self.relu(feature)
param = self.fc1_1(x)
return feature, param
def forward(self, input_l, input_r):
feature_l, param_l = self.forward_once(input_l)
feature_r, param_r = self.forward_once(input_r)
return feature_l, feature_r, param_l, param_r
def load_resnet50():
resnet = torchvision.models.resnet50()
model = sia_net(resnet)
return model
def sia_extract_param(checkpoint_fp, root = '', filelists = None, device_ids = [0],
batch_size = 128, num_workers = 8):
map_location = {f'cuda:{i}': 'cuda:0' for i in range(8)}
checkpoint = torch.load(checkpoint_fp, map_location=map_location)['state_dict'] ## 把张量从GPU 0~7 移动到 GPU 0, get paramerm's weight
torch.cuda.set_device(device_ids[0]) #bing cong zi dian zhong na chu key=='state_dict' de nei rong
model = load_resnet50() #get a model explain or document
model = nn.DataParallel(model, device_ids=device_ids).cuda()
model.load_state_dict(checkpoint)
dataset = SiaTestDataset(filelists=filelists, root=root,
transform=transforms.Compose([transforms.ToTensor(), Normalize(mean=127.5, std=128)]))
data_loader = data.DataLoader(dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers)
cudnn.benchmark = True #总的来说，大部分情况下，设置这个 flag 可以让内置的 cuDNN 的 auto-tuner 自动寻找最适合当前配置的高效算法，来达到优化运行效率的问题。
model.eval() #jiang model bian cheng test pattern
end = time.time()
param_62d = []
feature_512d = []
with torch.no_grad(): #bu ji suan gradient
for _, inputs in enumerate(data_loader):
inputs = inputs.cuda() #fang dao gpu shang jin xing yun suan
feature, _, param, _ = model(inputs, inputs)
for i in range(param.shape[0]): #output.shape[0] = 128 == batch_size
param_gen = param[i].cpu().numpy().flatten()
feature_gen = feature[i].cpu().numpy().flatten()
param_62d.append(param_gen)
feature_512d.append(feature_gen)
param_62d = np.array(param_62d, dtype=np.float32) # from list convert to array
feature_512d = np.array(feature_512d, dtype=np.float32)
print(f'Extracting params take {time.time() - end: .3f}s')
return feature_512d, param_62d
#def sia_extract_param_lfw(checkpoint_fp, filelists = None, device_ids = [0],
# batch_size = 128, num_workers = 8):
# map_location = {f'cuda:{i}': 'cuda:0' for i in range(8)}
# checkpoint = torch.load(checkpoint_fp, map_location=map_location)['state_dict'] ## 把张量从GPU 0~7 移动到 GPU 0, get paramerm's weight
# torch.cuda.set_device(device_ids[0]) #bing cong zi dian zhong na chu key=='state_dict' de nei rong
# model = load_resnet50() #get a model explain or document
# model = nn.DataParallel(model, device_ids=device_ids).cuda()
# model.load_state_dict(checkpoint)
# dataset = LFW_Dataset(filelists=filelists,
# transform=transforms.Compose([transforms.ToTensor(), Normalize(mean=127.5, std=128)]))
# data_loader = data.DataLoader(dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers)
# cudnn.benchmark = True #总的来说，大部分情况下，设置这个 flag 可以让内置的 cuDNN 的 auto-tuner 自动寻找最适合当前配置的高效算法，来达到优化运行效率的问题。
# model.eval() #jiang model bian cheng test pattern
#
# end = time.time()
# param_62d = []
# feature_512d = []
#
# with torch.no_grad(): #bu ji suan gradient
# for _, inputs in enumerate(data_loader):
# inputs = inputs.cuda() #fang dao gpu shang jin xing yun suan
# feature, _, param, _ = model(inputs, inputs)
#
#
# for i in range(param.shape[0]): #output.shape[0] = 128 == batch_size
# param_gen = param[i].cpu().numpy().flatten()
# feature_gen = feature[i].cpu().numpy().flatten()
#
# param_62d.append(param_gen)
# feature_512d.append(feature_gen)
#
# param_62d = np.array(param_62d, dtype=np.float32) # from list convert to array
# feature_512d = np.array(feature_512d, dtype=np.float32)
#
# print(f'Extracting params take {time.time() - end: .3f}s')
#
# return feature_512d, param_62d
def extract_3DMM(data_info):
_, param_62d = sia_extract_param(data_info['checkpoint_fp'], data_info['root'], data_info['filelists_test'])
return param_62d
def benchmark_3d_vertex_shp(params):
outputs = []
for i in range(params.shape[0]):
lm = reconstruct_vertex_shp(params[i])
outputs.append(lm)
return outputs
def benchmark_3d_vertex(params, dense = True):
outputs = []
for i in range(params.shape[0]):
lm = reconstruct_vertex(params[i], dense)
outputs.append(lm)
return outputs
def benchmark_3d_vertex_save(params, img_names_list, method='', dense = True):
save_path = 'result/'+method+'/'
mkdir(save_path)
for i in range(params.shape[0]):
lm = reconstruct_vertex(params[i], dense = True)
fn = img_names_list[i]
wfp = osp.join(save_path, fn.replace('.jpg', '.mat'))
print(wfp)
sio.savemat(wfp, {'vertex': lm})
def benchmark_3d_vertex_shp_save(params, img_names_list, method='', dense = True):
save_path = 'result-no-pose/'+method+'/'
mkdir(save_path)
for i in range(params.shape[0]):
lm = reconstruct_vertex_shp(params[i], dense = True)
fn = img_names_list[i]
wfp = osp.join(save_path, fn.replace('.jpg', '.mat'))
print(wfp)
sio.savemat(wfp, {'vertex': lm})
def feature_512d_save(feature_512d, img_name_list):
for i in range(feature_512d.shape[0]):
feature = feature_512d[i]
fn = img_name_list[i]
dirname, basename = os.path.split(fn)
dirname.replace('align_lfw', 'feature_lfw')
if not os.path.exists(dirname):
os.makedirs(dirname)
wfp = osp.join(dirname, basename.replace('.png', '.npy'))
_dump(wfp, feature)