-
Notifications
You must be signed in to change notification settings - Fork 0
/
fid.py
226 lines (143 loc) · 6.83 KB
/
fid.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
from dataset import BasicDataset
from skimage.io import imsave
import os
from shutil import rmtree
import numpy as np
import torch
import hashlib
import argparse
import sys
sys.path.append("./pytorch_fid/")
from utils import get_models, nullcontext
from pytorch_fid import fid_score
from pytorch_fid.inception import InceptionV3
TEMP_DIR = 'dump'
STATISTICS_DIR = 'fid_stats'
BATCH_SIZE = 20
DIMS = 768
def get_tag(string,dims):
tag = hashlib.md5(string.encode()).hexdigest()
return f'{dims}_{tag}'
def get_data_statistics(data_path,model,dims=DIMS,cuda=True,batch_size=BATCH_SIZE):
# Create temporary dump directory
tag = get_tag(data_path,dims)
fname = os.path.join(STATISTICS_DIR,tag+'.npz')
# Check if statistics exist
if os.path.exists(fname):
stats = np.load(fname)
m, s, num = stats['m'], stats['s'], stats['num_samples']
else:
temp_path = os.path.join(TEMP_DIR,tag) + '/'
if not os.path.exists(temp_path):
os.mkdir(temp_path)
# Load all images and convert to save as 0-255 .png or .jpg (normalize per image)
dataset = BasicDataset(model_dir=data_path)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=batch_size, num_workers=4, shuffle=False)
num = 0
for batch in dataloader:
batch = batch.detach().cpu().numpy()
batch = ((batch+1)/2*255).astype(np.uint8)
for i in range(batch.shape[0]):
img = np.tile(batch[i].transpose(1,2,0),(1,1,3))
imsave(os.path.join(temp_path,f'img_{num}.png'),img)
num += 1
else:
num = len(os.listdir(temp_path))
if cuda:
model.cuda()
m,s = fid_score._compute_statistics_of_path(temp_path, model, batch_size=batch_size, cuda=cuda, dims=dims)
# Remove dump folder
rmtree(temp_path)
np.savez(fname, m = m, s = s, num_samples = num)
return m,s,num
def get_gen_statistics(gen,name,num_samples,model,dims=DIMS,cuda=True,batch_size=BATCH_SIZE):
tag = get_tag(name,dims)
temp_path = os.path.join(TEMP_DIR,tag) + '/'
# Generate a folder full of generated images
if not os.path.exists(temp_path):
os.mkdir(temp_path)
num = 0
n_batches = int(num_samples / batch_size)+1
for k in range(n_batches):
batch = gen.sample(batch_size = batch_size).detach().cpu().numpy()
batch = ((batch+1)/2*255).astype(np.uint8)
for i in range(batch_size):
img = np.tile(batch[i].transpose(1,2,0),(1,1,3))
imsave(os.path.join(temp_path,f'img_{num}.png'),img)
num += 1
if cuda:
model.cuda()
m,s = fid_score._compute_statistics_of_path(temp_path, model, batch_size=batch_size, cuda=cuda, dims=dims)
# Remove dump dir
rmtree(temp_path)
return m,s
def get_fid_score(gen,gen_name,data_path,model,num_samples=-1,**kwargs):
# Load precalculated statistics
m1,s1,num_data = get_data_statistics(data_path,model,**kwargs)
if num_samples < 0:
num_samples = num_data
# Calculate generator statistics
m2,s2 = get_gen_statistics(gen,gen_name,num_samples,model,**kwargs)
# Calculate FID Score
fid_value = fid_score.calculate_frechet_distance(m1, s1, m2, s2)
return fid_value
class fid_scorer():
def __init__(self,
device,
dims=DIMS,
num_samples=-1,
verbose=True,
):
self.dims = dims
self.data_path = None
self.gen_name = None
self.num_samples = num_samples
self.cuda = device.type == 'cuda'
if verbose:
print('Loading inception model')
block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[dims]
self.inception = InceptionV3([block_idx]).to(device)
#self.inception = torch.nn.Module()
if verbose:
print('Finished loading inception')
def set_params(self,data_path,gen_name,writer):
self.writer = writer
self.data_path = data_path
self.gen_name = gen_name
def get_score(self,gen,step):
assert self.data_path and self.gen_name, "run set_params()"
fid_score = get_fid_score(gen,self.gen_name,self.data_path,self.inception,
num_samples=self.num_samples,cuda=self.cuda)
if self.writer:
self.writer.add_scalar('FID_score',fid_score,step)
return fid_score
if __name__ == '__main__':
print('Loading hyperparameters.')
# Get location of hparams.txt
parser = argparse.ArgumentParser(description='Train a GAN!')
parser.add_argument('-H', '--hparams', metavar='filename', type=str, default='hparams.txt',
help='File containing hyperparamters', dest='hparams')
parser.add_argument('-g', '--gpu', metavar='number', type=int, default='0',
help='Number of GPU to use', dest='gpu')
parser.add_argument('-d', '--dims', metavar='number', type=int, default=DIMS,
help='Number of dimensions of inception layer', dest='dims',
choices=[64,192,768,2048])
parser.add_argument('-s', '--samples', metavar='number', type=int, default=-1,
help='Number of samples from generator', dest='samples')
args = parser.parse_args()
# Set device
cuda = torch.cuda.is_available()
device = torch.device(f"cuda:{args.gpu}" if cuda else "cpu")
print(f'Using device: {device}')
print('Loading inception model')
block_idx = InceptionV3.BLOCK_INDEX_BY_DIM[args.dims]
inception = InceptionV3([block_idx]).to(device)
print('Finished loading inception')
with torch.cuda.device(device) if cuda else nullcontext:
# Load params from text file
models = get_models(args.hparams,device,load_discr=False)
print('Entering Hyperparameter Loop')
for h,gen,_ in models:
fid = get_fid_score(gen,h.name,h.dataroot,inception,
num_samples=args.samples,cuda=cuda)
print(fid)