Delete utils.py
Browse files
utils.py
DELETED
@@ -1,224 +0,0 @@
|
|
1 |
-
# coding: utf-8
|
2 |
-
__author__ = 'Roman Solovyev (ZFTurbo): https://github.com/ZFTurbo/'
|
3 |
-
|
4 |
-
import time
|
5 |
-
import numpy as np
|
6 |
-
import torch
|
7 |
-
import torch.nn as nn
|
8 |
-
import yaml
|
9 |
-
from ml_collections import ConfigDict
|
10 |
-
from omegaconf import OmegaConf
|
11 |
-
from tqdm import tqdm
|
12 |
-
|
13 |
-
def get_model_from_config(model_type, config_path):
|
14 |
-
with open(config_path) as f:
|
15 |
-
if model_type == 'htdemucs':
|
16 |
-
config = OmegaConf.load(config_path)
|
17 |
-
else:
|
18 |
-
config = ConfigDict(yaml.load(f, Loader=yaml.FullLoader))
|
19 |
-
|
20 |
-
if model_type == 'mdx23c':
|
21 |
-
from models.mdx23c_tfc_tdf_v3 import TFC_TDF_net
|
22 |
-
model = TFC_TDF_net(config)
|
23 |
-
elif model_type == 'htdemucs':
|
24 |
-
from models.demucs4ht import get_model
|
25 |
-
model = get_model(config)
|
26 |
-
elif model_type == 'segm_models':
|
27 |
-
from models.segm_models import Segm_Models_Net
|
28 |
-
model = Segm_Models_Net(config)
|
29 |
-
elif model_type == 'torchseg':
|
30 |
-
from models.torchseg_models import Torchseg_Net
|
31 |
-
model = Torchseg_Net(config)
|
32 |
-
elif model_type == 'mel_band_roformer':
|
33 |
-
from models.bs_roformer import MelBandRoformer
|
34 |
-
model = MelBandRoformer(
|
35 |
-
**dict(config.model)
|
36 |
-
)
|
37 |
-
elif model_type == 'bs_roformer':
|
38 |
-
from models.bs_roformer import BSRoformer
|
39 |
-
model = BSRoformer(
|
40 |
-
**dict(config.model)
|
41 |
-
)
|
42 |
-
elif model_type == 'swin_upernet':
|
43 |
-
from models.upernet_swin_transformers import Swin_UperNet_Model
|
44 |
-
model = Swin_UperNet_Model(config)
|
45 |
-
elif model_type == 'bandit':
|
46 |
-
from models.bandit.core.model import MultiMaskMultiSourceBandSplitRNNSimple
|
47 |
-
model = MultiMaskMultiSourceBandSplitRNNSimple(
|
48 |
-
**config.model
|
49 |
-
)
|
50 |
-
elif model_type == 'bandit_v2':
|
51 |
-
from models.bandit_v2.bandit import Bandit
|
52 |
-
model = Bandit(
|
53 |
-
**config.kwargs
|
54 |
-
)
|
55 |
-
elif model_type == 'scnet_unofficial':
|
56 |
-
from models.scnet_unofficial import SCNet
|
57 |
-
model = SCNet(
|
58 |
-
**config.model
|
59 |
-
)
|
60 |
-
elif model_type == 'scnet':
|
61 |
-
from models.scnet import SCNet
|
62 |
-
model = SCNet(
|
63 |
-
**config.model
|
64 |
-
)
|
65 |
-
else:
|
66 |
-
print('Unknown model: {}'.format(model_type))
|
67 |
-
model = None
|
68 |
-
|
69 |
-
return model, config
|
70 |
-
|
71 |
-
def _getWindowingArray(window_size, fade_size):
|
72 |
-
fadein = torch.linspace(0, 1, fade_size)
|
73 |
-
fadeout = torch.linspace(1, 0, fade_size)
|
74 |
-
window = torch.ones(window_size)
|
75 |
-
window[-fade_size:] *= fadeout
|
76 |
-
window[:fade_size] *= fadein
|
77 |
-
return window
|
78 |
-
|
79 |
-
|
80 |
-
def demix_track(config, model, mix, device, pbar=False):
|
81 |
-
C = config.audio.chunk_size
|
82 |
-
N = config.inference.num_overlap
|
83 |
-
fade_size = C // 10
|
84 |
-
step = int(C // N)
|
85 |
-
border = C - step
|
86 |
-
batch_size = config.inference.batch_size
|
87 |
-
|
88 |
-
length_init = mix.shape[-1]
|
89 |
-
|
90 |
-
# Do pad from the beginning and end to account floating window results better
|
91 |
-
if length_init > 2 * border and (border > 0):
|
92 |
-
mix = nn.functional.pad(mix, (border, border), mode='reflect')
|
93 |
-
|
94 |
-
# windowingArray crossfades at segment boundaries to mitigate clicking artifacts
|
95 |
-
windowingArray = _getWindowingArray(C, fade_size)
|
96 |
-
|
97 |
-
with torch.cuda.amp.autocast(enabled=config.training.use_amp):
|
98 |
-
with torch.inference_mode():
|
99 |
-
if config.training.target_instrument is not None:
|
100 |
-
req_shape = (1, ) + tuple(mix.shape)
|
101 |
-
else:
|
102 |
-
req_shape = (len(config.training.instruments),) + tuple(mix.shape)
|
103 |
-
|
104 |
-
result = torch.zeros(req_shape, dtype=torch.float32)
|
105 |
-
counter = torch.zeros(req_shape, dtype=torch.float32)
|
106 |
-
i = 0
|
107 |
-
batch_data = []
|
108 |
-
batch_locations = []
|
109 |
-
progress_bar = tqdm(total=mix.shape[1], desc="Processing audio chunks", leave=False) if pbar else None
|
110 |
-
|
111 |
-
while i < mix.shape[1]:
|
112 |
-
# print(i, i + C, mix.shape[1])
|
113 |
-
part = mix[:, i:i + C].to(device)
|
114 |
-
length = part.shape[-1]
|
115 |
-
if length < C:
|
116 |
-
if length > C // 2 + 1:
|
117 |
-
part = nn.functional.pad(input=part, pad=(0, C - length), mode='reflect')
|
118 |
-
else:
|
119 |
-
part = nn.functional.pad(input=part, pad=(0, C - length, 0, 0), mode='constant', value=0)
|
120 |
-
batch_data.append(part)
|
121 |
-
batch_locations.append((i, length))
|
122 |
-
i += step
|
123 |
-
|
124 |
-
if len(batch_data) >= batch_size or (i >= mix.shape[1]):
|
125 |
-
arr = torch.stack(batch_data, dim=0)
|
126 |
-
x = model(arr)
|
127 |
-
|
128 |
-
window = windowingArray
|
129 |
-
if i - step == 0: # First audio chunk, no fadein
|
130 |
-
window[:fade_size] = 1
|
131 |
-
elif i >= mix.shape[1]: # Last audio chunk, no fadeout
|
132 |
-
window[-fade_size:] = 1
|
133 |
-
|
134 |
-
for j in range(len(batch_locations)):
|
135 |
-
start, l = batch_locations[j]
|
136 |
-
result[..., start:start+l] += x[j][..., :l].cpu() * window[..., :l]
|
137 |
-
counter[..., start:start+l] += window[..., :l]
|
138 |
-
|
139 |
-
batch_data = []
|
140 |
-
batch_locations = []
|
141 |
-
|
142 |
-
if progress_bar:
|
143 |
-
progress_bar.update(step)
|
144 |
-
|
145 |
-
if progress_bar:
|
146 |
-
progress_bar.close()
|
147 |
-
|
148 |
-
estimated_sources = result / counter
|
149 |
-
estimated_sources = estimated_sources.cpu().numpy()
|
150 |
-
np.nan_to_num(estimated_sources, copy=False, nan=0.0)
|
151 |
-
|
152 |
-
if length_init > 2 * border and (border > 0):
|
153 |
-
# Remove pad
|
154 |
-
estimated_sources = estimated_sources[..., border:-border]
|
155 |
-
|
156 |
-
if config.training.target_instrument is None:
|
157 |
-
return {k: v for k, v in zip(config.training.instruments, estimated_sources)}
|
158 |
-
else:
|
159 |
-
return {k: v for k, v in zip([config.training.target_instrument], estimated_sources)}
|
160 |
-
|
161 |
-
|
162 |
-
def demix_track_demucs(config, model, mix, device, pbar=False):
|
163 |
-
S = len(config.training.instruments)
|
164 |
-
C = config.training.samplerate * config.training.segment
|
165 |
-
N = config.inference.num_overlap
|
166 |
-
batch_size = config.inference.batch_size
|
167 |
-
step = C // N
|
168 |
-
# print(S, C, N, step, mix.shape, mix.device)
|
169 |
-
|
170 |
-
with torch.cuda.amp.autocast(enabled=config.training.use_amp):
|
171 |
-
with torch.inference_mode():
|
172 |
-
req_shape = (S, ) + tuple(mix.shape)
|
173 |
-
result = torch.zeros(req_shape, dtype=torch.float32)
|
174 |
-
counter = torch.zeros(req_shape, dtype=torch.float32)
|
175 |
-
i = 0
|
176 |
-
batch_data = []
|
177 |
-
batch_locations = []
|
178 |
-
progress_bar = tqdm(total=mix.shape[1], desc="Processing audio chunks", leave=False) if pbar else None
|
179 |
-
|
180 |
-
while i < mix.shape[1]:
|
181 |
-
# print(i, i + C, mix.shape[1])
|
182 |
-
part = mix[:, i:i + C].to(device)
|
183 |
-
length = part.shape[-1]
|
184 |
-
if length < C:
|
185 |
-
part = nn.functional.pad(input=part, pad=(0, C - length, 0, 0), mode='constant', value=0)
|
186 |
-
batch_data.append(part)
|
187 |
-
batch_locations.append((i, length))
|
188 |
-
i += step
|
189 |
-
|
190 |
-
|
191 |
-
if len(batch_data) >= batch_size or (i >= mix.shape[1]):
|
192 |
-
arr = torch.stack(batch_data, dim=0)
|
193 |
-
x = model(arr)
|
194 |
-
for j in range(len(batch_locations)):
|
195 |
-
start, l = batch_locations[j]
|
196 |
-
result[..., start:start+l] += x[j][..., :l].cpu()
|
197 |
-
counter[..., start:start+l] += 1.
|
198 |
-
batch_data = []
|
199 |
-
batch_locations = []
|
200 |
-
|
201 |
-
if progress_bar:
|
202 |
-
progress_bar.update(step)
|
203 |
-
|
204 |
-
if progress_bar:
|
205 |
-
progress_bar.close()
|
206 |
-
|
207 |
-
estimated_sources = result / counter
|
208 |
-
estimated_sources = estimated_sources.cpu().numpy()
|
209 |
-
np.nan_to_num(estimated_sources, copy=False, nan=0.0)
|
210 |
-
|
211 |
-
if S > 1:
|
212 |
-
return {k: v for k, v in zip(config.training.instruments, estimated_sources)}
|
213 |
-
else:
|
214 |
-
return estimated_sources
|
215 |
-
|
216 |
-
|
217 |
-
def sdr(references, estimates):
|
218 |
-
# compute SDR for one song
|
219 |
-
delta = 1e-7 # avoid numerical errors
|
220 |
-
num = np.sum(np.square(references), axis=(1, 2))
|
221 |
-
den = np.sum(np.square(references - estimates), axis=(1, 2))
|
222 |
-
num += delta
|
223 |
-
den += delta
|
224 |
-
return 10 * np.log10(num / den)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|