-
Notifications
You must be signed in to change notification settings - Fork 1
/
Utils.py
261 lines (217 loc) · 10.9 KB
/
Utils.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
import json
import torch
import argparse
import librosa
import os
import cv2
import numpy as np
from torchvision.utils import save_image
def getArgs():
parser = argparse.ArgumentParser()
# Json path
parser.add_argument("--train_music_json", type=str, default='./Jsons/portrait/train_music.json', help="Path to train.json.")
parser.add_argument("--val_music_json", type=str, default='./Jsons/portrait/val_music.json', help="Path to val.json.")
parser.add_argument("--test_music_json", type=str, default='./Jsons/portraittest_music.json', help="Path to test.json.")
parser.add_argument("--sample_train_music_json", type=str, default='./Jsons/portrait/sample_train_music.json', help="Path to train.json.")
parser.add_argument("--sample_val_music_json", type=str, default='./Jsons/portrait/sample_val_music.json', help="Path to val.json.")
parser.add_argument("--train_paint_json", type=str, default='./Jsons/portrait/train_paint.json', help="Path to train.json.")
parser.add_argument("--val_paint_json", type=str, default='./Jsons/portrait/val_paint.json', help="Path to val.json.")
parser.add_argument("--test_paint_json", type=str, default='./Jsons/portrait/test_paint.json', help="Path to test.json.")
parser.add_argument("--sample_train_paint_json", type=str, default='./Jsons/portriat/sample_train_paint.json', help="Path to train.json.")
parser.add_argument("--sample_val_paint_json", type=str, default='./Jsons/portrait/sample_val_paint.json', help="Path to val.json.")
# Image
parser.add_argument("--paint_resize_min_edge", type=int, default=300)
parser.add_argument("--paint_crop_H", type=int, default=256)
parser.add_argument("--paint_crop_W", type=int, default=256)
# Audio
parser.add_argument("--sr", type=int, default=22050)
parser.add_argument("--n_mels", type=int, default=128)
parser.add_argument("--n_fft", type=int, default=1024)
# parser.add_argument("--hop_length", type=int, nargs="+", default=[256, 426, 512])
# parser.add_argument("--audio_length", type=float, nargs="+", default=[3.0, 5.0, 6.0])
# parser.add_argument("--win_sizes", type=int, nargs="+", default=[512, 1024, 2048])
parser.add_argument("--win_sizes", type=int, nargs="+", default=[1024])
parser.add_argument("--hop_length", type=int, default=256)
parser.add_argument("--audio_length", type=float, default=2.97)
# Year
parser.add_argument("--year_base", type=int, default=1730)
parser.add_argument("--year_interval", type=int, default=10)
# Save model path
# parser.add_argument("--save_model_path", type=str, default="./Result/", help="Path to save the models.")
parser.add_argument("--save_output_path", type=str, default="./Result/", help="Path to save the whole outputs.")
parser.add_argument("--load_model_path", type=str, default="", help="Path to load the models.")
#parser.add_argument("--load_model_epoch", type=int, default="-1", help="Path to load the models.")
# blahhhhh
parser.add_argument("--epochs", type=int, default=1000)
parser.add_argument("--sample_epoch", type=int, default=5)
parser.add_argument("--save_model_epoch", type=int, default=5)
parser.add_argument("--tfboard_log_epoch", type=int, default=5)
parser.add_argument("--batch_size", type=int, default=1)
parser.add_argument("--num_workers", type=int, default=1)
parser.add_argument("--gpu_ids", type=int, nargs="+", default=[0, 1, 2], help="Which gpu to use.")
parser.add_argument("--z_dim", type=int, default=32)
parser.add_argument("--npy", action='store_true')
# Experiment strategy
parser.add_argument("--regression", action = 'store_true', help = 'If use regression in ACGAN')
parser.add_argument("--mean", type = float, default = 1830, help = 'The year mean of normalization')
parser.add_argument("--std", type = float, default = 60.553, help = 'The year standard deviation of normalization')
parser.add_argument("--triplet", action = 'store_true', help = 'If use regression in ACGAN')
parser.add_argument("--split_num", type = int, default = 0, help = 'How many fragment you want to split along time axis. ')
# Evaluation
parser.add_argument("--eva_base", type=int, default=1)
parser.add_argument("--eva_count", type=int, default=1)
args = parser.parse_args()
presentParameters(vars(args))
return args
def presentParameters(args_dict):
"""
Print the parameters setting line by line
Arg: args_dict - The dict object which is transferred from argparse Namespace object
"""
Log("========== Parameters ==========")
for key in sorted(args_dict.keys()):
Log("{:>25} : {}".format(key, args_dict[key]))
Log("===============================")
def mkdir(path):
if not os.path.exists(path):
os.makedirs(path)
def gridTranspose(img_list):
"""
Transpose the tensor in grid view
Arg: img_list (List) - Contains tensors, and the shape is [(BCHW), (BCHW), ... for k times]
Ret: The transposed tensor, and the shape is (B*k, C, H, W)
"""
# Check if the whole tensor has the same shape
for i, tensor in enumerate(img_list):
if i > 0:
if tensor.size() != img_list[i-1].size():
raise Exception("Size {} and {}. You should make sure the tensor you want to transpose has the same shape!".format(
tensor.size(), img_list[i-1].size()
))
# Transpose
result = []
num_tensor = len(img_list) # The number of different music samples
num_obj = img_list[0].size(0) # The number of object in each music sample
for i in range(num_obj):
for j in range(num_tensor):
result.append(img_list[j][i])
result = torch.stack(result, 0)
assert len(result.size()) == 4
return result
def dumpWav(mel, sr, n_fft, hop_length):
return librosa.feature.inverse.mel_to_audio(mel, sr=sr, n_fft=n_fft, hop_length=hop_length)
def label2year(year_tag, args):
y_base = args.year_base
y_int = args.year_interval
return year_tag * y_int + y_base
def dumpPaint(paint, path, epoch, batch, itr=0, name=None):
std = 0.5
mean = 0.5
paint = paint*std+mean
paint *= 255.0
paint = np.transpose(paint, (1, 2, 0))
if name == None:
filename = os.path.join(path, "paint_{}_{}_{}.jpg".format(epoch, batch, itr))
else:
filename = os.path.join(path, name)
cv2.imwrite(filename, paint.astype(np.uint8))
def dumpMal(music, path, epoch, batch, itr=0):
filename = os.path.join(path, "mel_{}_{}_{}".format(epoch, batch, itr))
np.save(filename, music)
def dumpYear(year_label, args, path, epoch, batch, itr=0):
dump = {"year":label2year(year_label, args).tolist()}
filename = os.path.join(path, "year_{}_{}_{}.json".format(epoch, batch, itr))
with open(filename, "w") as fp:
json.dump(dump, fp, indent=4)
def dumpMusicLatent(m_latent, path, epoch, batch, itr=0, name=None):
if name == None:
filename = os.path.join(path, "lt_{}_{}_{}".format(epoch, batch, itr))
else:
filename = os.path.join(path, name)
np.save(filename, m_latent)
def sampleGt(dataloader, args):
"""
Sample ground truth with dataloader with batch_size=1
"""
for b, meta in enumerate(dataloader):
music = meta["music"]
paint = meta["paint"]
year = meta["year"]
music = np.squeeze(music.cpu().data.numpy())
paint = np.squeeze(paint.cpu().data.numpy())
year = np.squeeze(year.cpu().data.numpy())
path = os.path.join(args.save_output_path, "{:02d}".format(b), "GT")
mkdir(path)
# Sample
Log ("Sample to {}".format(path))
dumpMal(music, path, b, b)
dumpPaint(paint, path, b, b)
dumpYear(year, args, path, b, b)
def sample(model, dataloader, epoch, args):
"""
Sample output with dataloader with batch_size=1
"""
with torch.no_grad():
grid_result_list = []
gt_list = []
for b, meta in enumerate(dataloader):
music = meta["music"]
paint = meta["paint"]
year = meta["year"]
music = music.float().cuda()
# Forward
music = torch.cat([music[:, :, :, :music.size(3)//3], music[:, :, :, music.size(3)//3:2*music.size(3)//3], music[:, :, :, 2*music.size(3)//3:]], 1)
pred_music, pred_paint, pred_year, m_latent = model(music)
pred_music = np.squeeze(pred_music.cpu().data.numpy())
pred_paint = np.squeeze(pred_paint.cpu().data.numpy())
pred_year = np.squeeze(pred_year.cpu().data.numpy())
m_latent = np.squeeze(m_latent.cpu().data.numpy())
path = os.path.join(args.save_output_path, "{:02d}".format(b))
# Sample
Log ("Epoch:[{}]. Sample to {}".format(epoch, path))
dumpMal(pred_music, path, epoch, b)
dumpPaint(pred_paint, path, epoch, b)
dumpYear(pred_year, args, path, epoch, b)
dumpMusicLatent(m_latent, path, epoch, b)
# Sample for several times
if b % 5 == 0:
gt_list.append(paint[0].cpu())
_, pred_paint, _, _ = model(torch.cat([music] * 9, 0))
grid_result_list.append(pred_paint.cpu())
# Save grid
# grid_result_list = [torch.stack(gt_list, 0)] + grid_result_list
grid_result_list = [torch.stack(gt_list, 0)] + [gridTranspose(grid_result_list)]
grid_result = torch.cat(grid_result_list, 0)
filename = os.path.join(args.save_output_path, "Grid", "{}.jpg".format(epoch))
save_image(grid_result, filename, normalize=True, nrow=len(gt_list))
###################################################################################3
# Define the function to compute style
###################################################################################3
def gram_matrix(y):
(b, ch, h, w) = y.size()
features = y.view(b, ch, w * h)
features_t = features.transpose(1, 2)
gram = features.bmm(features_t) / (ch * h * w)
return gram
def normalize_batch(batch):
# normalize using imagenet mean and std
mean = batch.new_tensor([0.485, 0.456, 0.406]).view(-1, 1, 1)
std = batch.new_tensor([0.229, 0.224, 0.225]).view(-1, 1, 1)
batch = batch.div_(255.0)
return (batch - mean) / std
def get_gram(img, vgg):
assert torch.min(img) >= 0.0 and torch.max(img) <= 255.0
img = img.repeat(1, 1, 1, 1).cuda()
features_style = vgg(normalize_batch(img))
gram_style = [gram_matrix(y) for y in features_style]
return gram_style
def Log(string):
print("[MbST] {}".format(string))
if __name__ == "__main__":
sr, n_fft, hop_length = 22050, 1024, 512
y, sr = librosa.core.load("mezame.wav", sr=sr)
print ("Mel spec")
mel = librosa.feature.melspectrogram(y=y, sr=sr, n_fft=n_fft, hop_length=hop_length)
re_y = dumpWav(mel, sr, n_fft, hop_length)
print ("Dump file")
librosa.output.write_wav("re_mezame.wav", y=y, sr=sr)