Spaces:
Running
on
Zero
Running
on
Zero
Nithya
commited on
Commit
·
98eb218
1
Parent(s):
3752793
updated parent repo and restructured things
Browse files- .gitattributes +0 -35
- .gitignore +0 -1
- app.py +65 -150
- models/diffusion_pitch/config.gin +36 -35
- models/pitch_to_audio/config.gin +39 -36
- requirements.txt +1 -19
- src/dataset.py +0 -312
- src/generate_utils.py +0 -88
- src/model.py +0 -1130
- src/pitch_to_audio_utils.py +0 -121
- src/preprocess_utils.py +0 -127
- src/process_encodec.py +0 -22
- src/utils.py +0 -65
.gitattributes
DELETED
@@ -1,35 +0,0 @@
|
|
1 |
-
*.7z filter=lfs diff=lfs merge=lfs -text
|
2 |
-
*.arrow filter=lfs diff=lfs merge=lfs -text
|
3 |
-
*.bin filter=lfs diff=lfs merge=lfs -text
|
4 |
-
*.bz2 filter=lfs diff=lfs merge=lfs -text
|
5 |
-
*.ckpt filter=lfs diff=lfs merge=lfs -text
|
6 |
-
*.ftz filter=lfs diff=lfs merge=lfs -text
|
7 |
-
*.gz filter=lfs diff=lfs merge=lfs -text
|
8 |
-
*.h5 filter=lfs diff=lfs merge=lfs -text
|
9 |
-
*.joblib filter=lfs diff=lfs merge=lfs -text
|
10 |
-
*.lfs.* filter=lfs diff=lfs merge=lfs -text
|
11 |
-
*.mlmodel filter=lfs diff=lfs merge=lfs -text
|
12 |
-
*.model filter=lfs diff=lfs merge=lfs -text
|
13 |
-
*.msgpack filter=lfs diff=lfs merge=lfs -text
|
14 |
-
*.npy filter=lfs diff=lfs merge=lfs -text
|
15 |
-
*.npz filter=lfs diff=lfs merge=lfs -text
|
16 |
-
*.onnx filter=lfs diff=lfs merge=lfs -text
|
17 |
-
*.ot filter=lfs diff=lfs merge=lfs -text
|
18 |
-
*.parquet filter=lfs diff=lfs merge=lfs -text
|
19 |
-
*.pb filter=lfs diff=lfs merge=lfs -text
|
20 |
-
*.pickle filter=lfs diff=lfs merge=lfs -text
|
21 |
-
*.pkl filter=lfs diff=lfs merge=lfs -text
|
22 |
-
*.pt filter=lfs diff=lfs merge=lfs -text
|
23 |
-
*.pth filter=lfs diff=lfs merge=lfs -text
|
24 |
-
*.rar filter=lfs diff=lfs merge=lfs -text
|
25 |
-
*.safetensors filter=lfs diff=lfs merge=lfs -text
|
26 |
-
saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
27 |
-
*.tar.* filter=lfs diff=lfs merge=lfs -text
|
28 |
-
*.tar filter=lfs diff=lfs merge=lfs -text
|
29 |
-
*.tflite filter=lfs diff=lfs merge=lfs -text
|
30 |
-
*.tgz filter=lfs diff=lfs merge=lfs -text
|
31 |
-
*.wasm filter=lfs diff=lfs merge=lfs -text
|
32 |
-
*.xz filter=lfs diff=lfs merge=lfs -text
|
33 |
-
*.zip filter=lfs diff=lfs merge=lfs -text
|
34 |
-
*.zst filter=lfs diff=lfs merge=lfs -text
|
35 |
-
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
.gitignore
DELETED
@@ -1 +0,0 @@
|
|
1 |
-
src/__pycache__/
|
|
|
|
app.py
CHANGED
@@ -1,91 +1,28 @@
|
|
1 |
import spaces
|
2 |
-
from gradio import Interface, Audio
|
3 |
import gradio as gr
|
4 |
import numpy as np
|
5 |
import torch
|
6 |
-
import subprocess
|
7 |
import librosa
|
8 |
import matplotlib.pyplot as plt
|
9 |
import pandas as pd
|
10 |
import os
|
11 |
from functools import partial
|
12 |
import gin
|
13 |
-
import
|
14 |
-
|
15 |
-
from
|
16 |
-
import src.pitch_to_audio_utils as p2a
|
17 |
import torchaudio
|
18 |
from absl import app
|
19 |
from torch.nn.functional import interpolate
|
20 |
-
import pdb
|
21 |
import logging
|
22 |
import crepe
|
23 |
from hmmlearn import hmm
|
24 |
-
import time
|
25 |
import soundfile as sf
|
|
|
26 |
|
27 |
pitch_path = 'models/diffusion_pitch/'
|
28 |
-
# pitch_path = '/network/scratch/n/nithya.shikarpur/checkpoints/pitch-diffusion/corrected-attention-v3/4833583'
|
29 |
audio_path = 'models/pitch_to_audio/'
|
30 |
-
|
31 |
-
# db_path_audio = '/home/mila/n/nithya.shikarpur/scratch/pitch-diffusion/data/merged_data-finalest/cached-audio-pitch-16k'
|
32 |
-
|
33 |
-
device = 'cuda'
|
34 |
-
|
35 |
-
global_ind = -1
|
36 |
-
global_audios = np.array([0.0])
|
37 |
-
global_pitches = np.array([0])
|
38 |
-
singer = 3
|
39 |
-
audio_components = []
|
40 |
-
preprocessed_primes = []
|
41 |
-
selected_prime = None
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
def make_prime_npz(prime):
|
46 |
-
np.savez('./temp/prime.npz', concatenated_array=[[prime]])
|
47 |
-
|
48 |
-
def load_pitch_fns():
|
49 |
-
pitch_model, pitch_qt, _, pitch_task_fn = load_pitch_model(
|
50 |
-
os.path.join(pitch_path, 'config.gin'),
|
51 |
-
os.path.join(pitch_path, 'last.ckpt'),
|
52 |
-
os.path.join(pitch_path, 'qt.joblib'),
|
53 |
-
device=device
|
54 |
-
)
|
55 |
-
invert_pitch_fn = partial(
|
56 |
-
invert_pitch_read,
|
57 |
-
min_norm_pitch=gin.query_parameter('dataset.pitch_read_w_downsample.min_norm_pitch'),
|
58 |
-
time_downsample=gin.query_parameter('dataset.pitch_read_w_downsample.time_downsample'),
|
59 |
-
pitch_downsample=gin.query_parameter('dataset.pitch_read_w_downsample.pitch_downsample'),
|
60 |
-
qt_transform=pitch_qt,
|
61 |
-
min_clip=gin.query_parameter('dataset.pitch_read_w_downsample.min_clip'),
|
62 |
-
max_clip=gin.query_parameter('dataset.pitch_read_w_downsample.max_clip')
|
63 |
-
)
|
64 |
-
return pitch_model, pitch_qt, pitch_task_fn, invert_pitch_fn
|
65 |
-
|
66 |
-
def interpolate_pitch(pitch, audio_seq_len):
|
67 |
-
pitch = interpolate(pitch, size=audio_seq_len, mode='linear')
|
68 |
-
# plt.plot(pitch[0].squeeze(0).detach().cpu().numpy())
|
69 |
-
# plt.savefig(f"./temp/interpolated_pitch.png")
|
70 |
-
# plt.close()
|
71 |
-
return pitch
|
72 |
-
|
73 |
-
def load_audio_fns():
|
74 |
-
ckpt = os.path.join(audio_path, 'last.ckpt')
|
75 |
-
config = os.path.join(audio_path, 'config.gin')
|
76 |
-
qt = os.path.join(audio_path, 'qt.joblib')
|
77 |
-
# qt = '/home/mila/n/nithya.shikarpur/scratch/pitch-diffusion/data/merged_data-finalest/cached-audio-pitch-16k/qt.joblib'
|
78 |
-
|
79 |
-
audio_model, audio_qt = load_audio_model(config, ckpt, qt, device=device)
|
80 |
-
audio_seq_len = gin.query_parameter('%AUDIO_SEQ_LEN')
|
81 |
-
|
82 |
-
invert_audio_fn = partial(
|
83 |
-
p2a.normalized_mels_to_audio,
|
84 |
-
qt=audio_qt,
|
85 |
-
n_iter=200
|
86 |
-
)
|
87 |
-
|
88 |
-
return audio_model, audio_qt, audio_seq_len, invert_audio_fn
|
89 |
|
90 |
def predict_voicing(confidence):
|
91 |
# https://github.com/marl/crepe/pull/26
|
@@ -136,73 +73,67 @@ def extract_pitch(audio, unvoice=True, sr=16000, frame_shift_ms=10, log=True):
|
|
136 |
|
137 |
return time, f0, confidence
|
138 |
|
139 |
-
def
|
140 |
-
|
141 |
-
|
|
|
|
|
142 |
samples = pitch_model.sample_sdedit(noisy_pitch, num_samples, num_steps)
|
143 |
-
inverted_pitches = [invert_pitch_fn(samples.detach().cpu().numpy()[0])[0]]
|
144 |
|
145 |
-
if outfolder is not None:
|
146 |
-
os.makedirs(outfolder, exist_ok=True)
|
147 |
-
# pdb.set_trace()
|
148 |
-
for i, pitch in enumerate(inverted_pitches):
|
149 |
-
flattened_pitch = pitch.flatten()
|
150 |
-
pd.DataFrame({'f0': flattened_pitch}).to_csv(f"{outfolder}/{i}.csv", index=False)
|
151 |
-
plt.plot(np.where(flattened_pitch == 0, np.nan, flattened_pitch))
|
152 |
-
plt.savefig(f"{outfolder}/{i}.png")
|
153 |
-
plt.close()
|
154 |
return samples, inverted_pitches
|
155 |
|
156 |
-
def generate_audio(audio_model, f0s, invert_audio_fn,
|
|
|
157 |
singer_tensor = torch.tensor(np.repeat(singers, repeats=f0s.shape[0])).to(audio_model.device)
|
158 |
samples, _, singers = audio_model.sample_cfg(f0s.shape[0], f0=f0s, num_steps=num_steps, singer=singer_tensor, strength=3)
|
159 |
audio = invert_audio_fn(samples)
|
160 |
-
|
161 |
-
if outfolder is not None:
|
162 |
-
os.makedirs(outfolder, exist_ok=True)
|
163 |
-
for i, a in enumerate(audio):
|
164 |
-
logging.log(logging.INFO, f"Saving audio {i}")
|
165 |
-
torchaudio.save(f"{outfolder}/{i}.wav", torch.tensor(a).detach().unsqueeze(0).cpu(), 16000)
|
166 |
return audio
|
167 |
|
168 |
@spaces.GPU(duration=120)
|
169 |
-
def generate(pitch, num_samples=
|
170 |
-
|
171 |
-
global preprocessed_primes
|
172 |
-
# pdb.set_trace()
|
173 |
logging.log(logging.INFO, 'Generate function')
|
174 |
-
pitch, inverted_pitch =
|
175 |
if pitch_qt is not None:
|
|
|
176 |
def undo_qt(x, min_clip=200):
|
177 |
pitch= pitch_qt.inverse_transform(x.reshape(-1, 1)).reshape(1, -1)
|
178 |
pitch = np.around(pitch) # round to nearest integer, done in preprocessing of pitch contour fed into model
|
179 |
pitch[pitch < 200] = np.nan
|
180 |
return pitch
|
181 |
pitch = torch.tensor(np.array([undo_qt(x) for x in pitch.detach().cpu().numpy()])).to(pitch_model.device)
|
182 |
-
interpolated_pitch = interpolate_pitch(pitch=pitch, audio_seq_len=audio_seq_len)
|
183 |
-
interpolated_pitch = torch.nan_to_num(interpolated_pitch, nan=196)
|
184 |
interpolated_pitch = interpolated_pitch.squeeze(1) # to match input size by removing the extra dimension
|
185 |
-
audio = generate_audio(audio_model, interpolated_pitch, invert_audio_fn, singers=singers, num_steps=100
|
186 |
-
|
187 |
-
audio = audio.detach().cpu().numpy()[:, :]
|
188 |
pitch = pitch.detach().cpu().numpy()
|
189 |
-
# state = [(16000, audio[0]), (16000, audio[1])]
|
190 |
-
# pdb.set_trace()
|
191 |
pitch_vals = np.where(pitch[0][:, 0] == 0, np.nan, pitch[0].flatten())
|
192 |
-
fig1 = plt.figure()
|
193 |
-
# plt.plot(np.arange(0, 400), pitch_vals[:400], figure=fig1, label='User Input')
|
194 |
-
plt.plot(pitch_vals, figure=fig1, label='Pitch')
|
195 |
-
# plt.legend(fig1)
|
196 |
-
# state.append(fig1)
|
197 |
-
plt.close(fig1)
|
198 |
-
return (16000, audio[0]), fig1, pitch_vals
|
199 |
|
200 |
-
|
201 |
-
|
202 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
203 |
|
204 |
-
@spaces.GPU(duration=
|
205 |
-
def
|
206 |
global selected_prime, pitch_task_fn
|
207 |
|
208 |
if audio is None:
|
@@ -215,40 +146,32 @@ def set_prime_and_generate(audio, full_pitch, full_audio, full_user):
|
|
215 |
audio /= np.max(np.abs(audio))
|
216 |
audio = librosa.resample(audio, orig_sr=sr, target_sr=16000) # convert only last 4 s
|
217 |
mic_audio = audio.copy()
|
218 |
-
audio = audio[-12*16000:]
|
219 |
_, f0, _ = extract_pitch(audio)
|
220 |
-
mic_f0 = f0.copy()
|
221 |
-
f0 = pitch_task_fn({
|
222 |
-
'
|
223 |
-
'
|
224 |
-
|
225 |
-
|
226 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
227 |
f0 = f0.reshape(1, 1, -1)
|
228 |
f0 = torch.tensor(f0).to(pitch_model.device).float()
|
229 |
-
audio, pitch,
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
-
|
234 |
-
|
235 |
-
|
236 |
-
plt.plot(np.arange(0, len(mic_f0)), mic_f0, label='User Input', figure=fig)
|
237 |
-
plt.close(fig)
|
238 |
-
return audio, full_pitch, full_audio, full_user, pitch
|
239 |
-
|
240 |
-
def save_session(full_pitch, full_audio, full_user):
|
241 |
-
pass
|
242 |
-
# os.makedirs(output_folder, exist_ok=True)
|
243 |
-
# filename = f'session-{time.time()}'
|
244 |
-
# logging.log(logging.INFO, f"Saving session to {filename}")
|
245 |
-
# pd.DataFrame({'pitch': full_pitch, 'time': np.arange(0, len(full_pitch)/100, 0.01), 'user': full_user}).to_csv(os.path.join(output_folder, filename + '.csv'), index=False)
|
246 |
-
# sf.write(os.path.join(output_folder, filename + '.wav'), full_audio[1], 16000)
|
247 |
|
248 |
with gr.Blocks() as demo:
|
249 |
-
full_audio = gr.State((16000, np.array([])))
|
250 |
-
full_pitch = gr.State(np.array([]))
|
251 |
-
full_user = gr.State(np.array([]))
|
252 |
with gr.Row():
|
253 |
with gr.Column():
|
254 |
audio = gr.Audio(label="Input")
|
@@ -257,17 +180,9 @@ with gr.Blocks() as demo:
|
|
257 |
with gr.Column():
|
258 |
generated_audio = gr.Audio(label="Generated Audio")
|
259 |
generated_pitch = gr.Plot(label="Generated Pitch")
|
260 |
-
sbmt.click(
|
261 |
-
save = gr.Button("Save Session")
|
262 |
-
save.click(save_session, inputs=[full_pitch, full_audio, full_user])
|
263 |
-
|
264 |
-
|
265 |
|
266 |
def main(argv):
|
267 |
-
# audio = np.random.randint(0, high=128, size=(44100*5), dtype=np.int16)
|
268 |
-
# sr = 44100
|
269 |
-
# pdb.set_trace()
|
270 |
-
# p, a = set_prime_and_generate((sr, audio))
|
271 |
|
272 |
demo.launch(share=True)
|
273 |
|
|
|
1 |
import spaces
|
|
|
2 |
import gradio as gr
|
3 |
import numpy as np
|
4 |
import torch
|
|
|
5 |
import librosa
|
6 |
import matplotlib.pyplot as plt
|
7 |
import pandas as pd
|
8 |
import os
|
9 |
from functools import partial
|
10 |
import gin
|
11 |
+
from gamadhani.utils.generate_utils import load_pitch_fns, load_audio_fns
|
12 |
+
import gamadhani.utils.pitch_to_audio_utils as p2a
|
13 |
+
from gamadhani.utils.utils import get_device
|
|
|
14 |
import torchaudio
|
15 |
from absl import app
|
16 |
from torch.nn.functional import interpolate
|
|
|
17 |
import logging
|
18 |
import crepe
|
19 |
from hmmlearn import hmm
|
|
|
20 |
import soundfile as sf
|
21 |
+
import pdb
|
22 |
|
23 |
pitch_path = 'models/diffusion_pitch/'
|
|
|
24 |
audio_path = 'models/pitch_to_audio/'
|
25 |
+
device = get_device()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
|
27 |
def predict_voicing(confidence):
|
28 |
# https://github.com/marl/crepe/pull/26
|
|
|
73 |
|
74 |
return time, f0, confidence
|
75 |
|
76 |
+
def generate_pitch_reinterp(pitch, pitch_model, invert_pitch_fn, num_samples, num_steps, noise_std=0.4):
|
77 |
+
'''Generate pitch values for the melodic reinterpretation task'''
|
78 |
+
# hardcoding the amount of noise to be added
|
79 |
+
noisy_pitch = torch.Tensor(pitch[:, :, -1200:]).to(pitch_model.device) + (torch.normal(mean=0.0, std=noise_std*torch.ones((1200)))).to(pitch_model.device)
|
80 |
+
noisy_pitch = torch.clamp(noisy_pitch, -5.19, 5.19) # clipping the pitch values to be within the range of the model
|
81 |
samples = pitch_model.sample_sdedit(noisy_pitch, num_samples, num_steps)
|
82 |
+
inverted_pitches = [invert_pitch_fn(f0=samples.detach().cpu().numpy()[0])[0]] # pitch values in Hz
|
83 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
84 |
return samples, inverted_pitches
|
85 |
|
86 |
+
def generate_audio(audio_model, f0s, invert_audio_fn, singers=[3], num_steps=100):
|
87 |
+
'''Generate audio given pitch values'''
|
88 |
singer_tensor = torch.tensor(np.repeat(singers, repeats=f0s.shape[0])).to(audio_model.device)
|
89 |
samples, _, singers = audio_model.sample_cfg(f0s.shape[0], f0=f0s, num_steps=num_steps, singer=singer_tensor, strength=3)
|
90 |
audio = invert_audio_fn(samples)
|
91 |
+
|
|
|
|
|
|
|
|
|
|
|
92 |
return audio
|
93 |
|
94 |
@spaces.GPU(duration=120)
|
95 |
+
def generate(pitch, num_samples=1, num_steps=100, singers=[3], outfolder='temp', audio_seq_len=750, pitch_qt=None ):
|
96 |
+
|
|
|
|
|
97 |
logging.log(logging.INFO, 'Generate function')
|
98 |
+
pitch, inverted_pitch = generate_pitch_reinterp(pitch, pitch_model, invert_pitch_fn, num_samples=num_samples, num_steps=100)
|
99 |
if pitch_qt is not None:
|
100 |
+
# if there is not pitch quantile transformer, undo the default quantile transformation that occurs
|
101 |
def undo_qt(x, min_clip=200):
|
102 |
pitch= pitch_qt.inverse_transform(x.reshape(-1, 1)).reshape(1, -1)
|
103 |
pitch = np.around(pitch) # round to nearest integer, done in preprocessing of pitch contour fed into model
|
104 |
pitch[pitch < 200] = np.nan
|
105 |
return pitch
|
106 |
pitch = torch.tensor(np.array([undo_qt(x) for x in pitch.detach().cpu().numpy()])).to(pitch_model.device)
|
107 |
+
interpolated_pitch = p2a.interpolate_pitch(pitch=pitch, audio_seq_len=audio_seq_len) # interpolate pitch values to match the audio model's input size
|
108 |
+
interpolated_pitch = torch.nan_to_num(interpolated_pitch, nan=196) # replace nan values with silent token
|
109 |
interpolated_pitch = interpolated_pitch.squeeze(1) # to match input size by removing the extra dimension
|
110 |
+
audio = generate_audio(audio_model, interpolated_pitch, invert_audio_fn, singers=singers, num_steps=100)
|
111 |
+
audio = audio.detach().cpu().numpy()
|
|
|
112 |
pitch = pitch.detach().cpu().numpy()
|
|
|
|
|
113 |
pitch_vals = np.where(pitch[0][:, 0] == 0, np.nan, pitch[0].flatten())
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
114 |
|
115 |
+
# generate plot of model output to display on interface
|
116 |
+
model_output_plot = plt.figure()
|
117 |
+
plt.plot(pitch_vals, figure=model_output_plot, label='Model Output')
|
118 |
+
plt.close(model_output_plot)
|
119 |
+
return (16000, audio[0]), model_output_plot, pitch_vals
|
120 |
+
|
121 |
+
# pdb.set_trace()
|
122 |
+
pitch_model, pitch_qt, pitch_task_fn, invert_pitch_fn, _ = load_pitch_fns(
|
123 |
+
os.path.join(pitch_path, 'last.ckpt'), \
|
124 |
+
model_type = 'diffusion', \
|
125 |
+
config_path = os.path.join(pitch_path, 'config.gin'), \
|
126 |
+
qt_path = os.path.join(pitch_path, 'qt.joblib'), \
|
127 |
+
)
|
128 |
+
audio_model, audio_qt, audio_seq_len, invert_audio_fn = load_audio_fns(
|
129 |
+
os.path.join(audio_path, 'last.ckpt'),
|
130 |
+
qt_path = os.path.join(audio_path, 'qt.joblib'),
|
131 |
+
config_path = os.path.join(audio_path, 'config.gin')
|
132 |
+
)
|
133 |
+
partial_generate = partial(generate, num_samples=1, num_steps=100, singers=[3], outfolder=None, pitch_qt=pitch_qt) # generate function with default arguments
|
134 |
|
135 |
+
@spaces.GPU(duration=120)
|
136 |
+
def set_guide_and_generate(audio):
|
137 |
global selected_prime, pitch_task_fn
|
138 |
|
139 |
if audio is None:
|
|
|
146 |
audio /= np.max(np.abs(audio))
|
147 |
audio = librosa.resample(audio, orig_sr=sr, target_sr=16000) # convert only last 4 s
|
148 |
mic_audio = audio.copy()
|
149 |
+
audio = audio[-12*16000:] # consider only last 12 s
|
150 |
_, f0, _ = extract_pitch(audio)
|
151 |
+
mic_f0 = f0.copy() # save the user input pitch values
|
152 |
+
f0 = pitch_task_fn(**{
|
153 |
+
'inputs': {
|
154 |
+
'pitch': {
|
155 |
+
'data': torch.Tensor(f0), # task function expects a tensor
|
156 |
+
'sampling_rate': 100
|
157 |
+
}
|
158 |
+
},
|
159 |
+
'qt_transform': pitch_qt,
|
160 |
+
'time_downsample': 1, # pitch will be extracted at 100 Hz, thus no downsampling
|
161 |
+
'seq_len': None,
|
162 |
+
})['sampled_sequence']
|
163 |
+
# pdb.set_trace()
|
164 |
f0 = f0.reshape(1, 1, -1)
|
165 |
f0 = torch.tensor(f0).to(pitch_model.device).float()
|
166 |
+
audio, pitch, _ = partial_generate(f0)
|
167 |
+
mic_f0 = np.where(mic_f0 == 0, np.nan, mic_f0)
|
168 |
+
# plot user input
|
169 |
+
user_input_plot = plt.figure()
|
170 |
+
plt.plot(np.arange(0, len(mic_f0)), mic_f0, label='User Input', figure=user_input_plot)
|
171 |
+
plt.close(user_input_plot)
|
172 |
+
return audio, user_input_plot, pitch
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
173 |
|
174 |
with gr.Blocks() as demo:
|
|
|
|
|
|
|
175 |
with gr.Row():
|
176 |
with gr.Column():
|
177 |
audio = gr.Audio(label="Input")
|
|
|
180 |
with gr.Column():
|
181 |
generated_audio = gr.Audio(label="Generated Audio")
|
182 |
generated_pitch = gr.Plot(label="Generated Pitch")
|
183 |
+
sbmt.click(set_guide_and_generate, inputs=[audio], outputs=[generated_audio, user_input, generated_pitch])
|
|
|
|
|
|
|
|
|
184 |
|
185 |
def main(argv):
|
|
|
|
|
|
|
|
|
186 |
|
187 |
demo.launch(share=True)
|
188 |
|
models/diffusion_pitch/config.gin
CHANGED
@@ -1,7 +1,9 @@
|
|
1 |
from __gin__ import dynamic_registration
|
2 |
-
from
|
3 |
-
from src import
|
4 |
-
from src import
|
|
|
|
|
5 |
import torch
|
6 |
|
7 |
# Macros:
|
@@ -23,47 +25,46 @@ utils.build_warmed_exponential_lr_scheduler.eta_min = 0.1
|
|
23 |
utils.build_warmed_exponential_lr_scheduler.peak_iteration = 10000
|
24 |
utils.build_warmed_exponential_lr_scheduler.start_factor = 0.01
|
25 |
|
26 |
-
# Parameters for
|
27 |
# ==============================================================================
|
28 |
-
|
29 |
-
|
30 |
@utils.build_warmed_exponential_lr_scheduler
|
31 |
|
32 |
-
# Parameters for dataset.
|
33 |
# ==============================================================================
|
34 |
-
dataset.
|
35 |
-
|
36 |
-
|
37 |
-
|
38 |
-
|
39 |
-
|
40 |
-
|
41 |
-
|
|
|
42 |
|
43 |
# Parameters for train/dataset.pitch_read_w_downsample:
|
44 |
# ==============================================================================
|
45 |
-
train/dataset.
|
46 |
|
47 |
-
# Parameters for train/dataset.
|
48 |
# ==============================================================================
|
49 |
-
|
|
|
50 |
|
51 |
-
# Parameters for val/dataset.SequenceDataset:
|
52 |
-
# ==============================================================================
|
53 |
-
val/dataset.SequenceDataset.task_fn = @dataset.pitch_read_w_downsample
|
54 |
|
55 |
-
# Parameters for
|
56 |
# ==============================================================================
|
57 |
-
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
68 |
-
|
69 |
-
|
|
|
1 |
from __gin__ import dynamic_registration
|
2 |
+
from gamadhani import src
|
3 |
+
from gamadhani.src import dataset
|
4 |
+
from gamadhani.src import model_diffusion
|
5 |
+
from gamadhani.src import task_functions
|
6 |
+
from gamadhani.utils import utils
|
7 |
import torch
|
8 |
|
9 |
# Macros:
|
|
|
25 |
utils.build_warmed_exponential_lr_scheduler.peak_iteration = 10000
|
26 |
utils.build_warmed_exponential_lr_scheduler.start_factor = 0.01
|
27 |
|
28 |
+
# Parameters for model_diffusion.UNetBase.configure_optimizers:
|
29 |
# ==============================================================================
|
30 |
+
model_diffusion.UNetBase.configure_optimizers.optimizer_cls = @torch.optim.AdamW
|
31 |
+
model_diffusion.UNetBase.configure_optimizers.scheduler_cls = \
|
32 |
@utils.build_warmed_exponential_lr_scheduler
|
33 |
|
34 |
+
# Parameters for dataset.Task:
|
35 |
# ==============================================================================
|
36 |
+
src.dataset.Task.kwargs = {
|
37 |
+
"decoder_key" : 'pitch',
|
38 |
+
"max_clip" : 600,
|
39 |
+
"min_clip" : 200,
|
40 |
+
"min_norm_pitch" : -4915,
|
41 |
+
"pitch_downsample" : 10,
|
42 |
+
"seq_len" : %SEQ_LEN,
|
43 |
+
"time_downsample" : 2}
|
44 |
+
|
45 |
|
46 |
# Parameters for train/dataset.pitch_read_w_downsample:
|
47 |
# ==============================================================================
|
48 |
+
# train/dataset.Task.kwargs = {"transpose_pitch": %TRANSPOSE_VALUE}
|
49 |
|
50 |
+
# Parameters for train/dataset.Task:
|
51 |
# ==============================================================================
|
52 |
+
src.dataset.Task.read_fn = @src.task_functions.pitch_read_downsample_diff
|
53 |
+
src.dataset.Task.invert_fn = @src.task_functions.invert_pitch_read_downsample_diff
|
54 |
|
|
|
|
|
|
|
55 |
|
56 |
+
# Parameters for model_diffusion.UNet:
|
57 |
# ==============================================================================
|
58 |
+
model_diffusion.UNet.dropout = 0.3
|
59 |
+
model_diffusion.UNet.features = [512, 640, 1024]
|
60 |
+
model_diffusion.UNet.inp_dim = 1
|
61 |
+
model_diffusion.UNet.kernel_size = 5
|
62 |
+
model_diffusion.UNet.nonlinearity = 'mish'
|
63 |
+
model_diffusion.UNet.norm = True
|
64 |
+
model_diffusion.UNet.num_attns = 4
|
65 |
+
model_diffusion.UNet.num_convs = 4
|
66 |
+
model_diffusion.UNet.num_heads = 8
|
67 |
+
model_diffusion.UNet.project_dim = 256
|
68 |
+
model_diffusion.UNet.seq_len = %SEQ_LEN
|
69 |
+
model_diffusion.UNet.strides = [4, 2, 2]
|
70 |
+
model_diffusion.UNet.time_dim = 128
|
models/pitch_to_audio/config.gin
CHANGED
@@ -1,8 +1,9 @@
|
|
1 |
from __gin__ import dynamic_registration
|
2 |
-
from
|
3 |
-
from src import
|
4 |
-
from src import
|
5 |
-
from
|
|
|
6 |
import torch
|
7 |
|
8 |
# Macros:
|
@@ -27,10 +28,10 @@ utils.build_warmed_exponential_lr_scheduler.eta_min = 0.1
|
|
27 |
utils.build_warmed_exponential_lr_scheduler.peak_iteration = 10000
|
28 |
utils.build_warmed_exponential_lr_scheduler.start_factor = 0.01
|
29 |
|
30 |
-
# Parameters for
|
31 |
# ==============================================================================
|
32 |
-
|
33 |
-
|
34 |
@utils.build_warmed_exponential_lr_scheduler
|
35 |
|
36 |
# Parameters for pitch_to_audio_utils.from_mels:
|
@@ -39,11 +40,6 @@ pitch_to_audio_utils.from_mels.nfft = %NFFT
|
|
39 |
pitch_to_audio_utils.from_mels.num_mels = %NUM_MELS
|
40 |
pitch_to_audio_utils.from_mels.sr = %SR
|
41 |
|
42 |
-
# Parameters for dataset.load_cached_dataset:
|
43 |
-
# ==============================================================================
|
44 |
-
dataset.load_cached_dataset.audio_len = %AUDIO_SEQ_LEN
|
45 |
-
dataset.load_cached_dataset.return_singer = %SINGER_CONDITIONING
|
46 |
-
|
47 |
# Parameters for pitch_to_audio_utils.normalized_mels_to_audio:
|
48 |
# ==============================================================================
|
49 |
pitch_to_audio_utils.normalized_mels_to_audio.n_iter = 100
|
@@ -53,7 +49,13 @@ pitch_to_audio_utils.normalized_mels_to_audio.sr = %SR
|
|
53 |
|
54 |
# Parameters for dataset.SequenceDataset:
|
55 |
# ==============================================================================
|
56 |
-
dataset.SequenceDataset.
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
|
58 |
# Parameters for pitch_to_audio_utils.torch_gl:
|
59 |
# ==============================================================================
|
@@ -65,27 +67,28 @@ pitch_to_audio_utils.torch_gl.sr = %SR
|
|
65 |
# ==============================================================================
|
66 |
pitch_to_audio_utils.torch_istft.nfft = %NFFT
|
67 |
|
68 |
-
# Parameters for
|
69 |
# ==============================================================================
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
78 |
-
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
|
84 |
-
|
85 |
-
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
|
|
|
1 |
from __gin__ import dynamic_registration
|
2 |
+
from gamadhani import src
|
3 |
+
from gamadhani.src import dataset
|
4 |
+
from gamadhani.src import model_diffusion
|
5 |
+
from gamadhani.utils import pitch_to_audio_utils
|
6 |
+
from gamadhani.utils import utils
|
7 |
import torch
|
8 |
|
9 |
# Macros:
|
|
|
28 |
utils.build_warmed_exponential_lr_scheduler.peak_iteration = 10000
|
29 |
utils.build_warmed_exponential_lr_scheduler.start_factor = 0.01
|
30 |
|
31 |
+
# Parameters for model_diffusion.UNetPitchConditioned.configure_optimizers:
|
32 |
# ==============================================================================
|
33 |
+
model_diffusion.UNetPitchConditioned.configure_optimizers.optimizer_cls = @torch.optim.AdamW
|
34 |
+
model_diffusion.UNetPitchConditioned.configure_optimizers.scheduler_cls = \
|
35 |
@utils.build_warmed_exponential_lr_scheduler
|
36 |
|
37 |
# Parameters for pitch_to_audio_utils.from_mels:
|
|
|
40 |
pitch_to_audio_utils.from_mels.num_mels = %NUM_MELS
|
41 |
pitch_to_audio_utils.from_mels.sr = %SR
|
42 |
|
|
|
|
|
|
|
|
|
|
|
43 |
# Parameters for pitch_to_audio_utils.normalized_mels_to_audio:
|
44 |
# ==============================================================================
|
45 |
pitch_to_audio_utils.normalized_mels_to_audio.n_iter = 100
|
|
|
49 |
|
50 |
# Parameters for dataset.SequenceDataset:
|
51 |
# ==============================================================================
|
52 |
+
dataset.SequenceDataset.task = @dataset.Task()
|
53 |
+
|
54 |
+
# Parameters for dataset.Task:
|
55 |
+
# ==============================================================================
|
56 |
+
dataset.Task.read_fn = @dataset.load_cached_dataset
|
57 |
+
dataset.Task.kwargs = {"audio_len": %AUDIO_SEQ_LEN,
|
58 |
+
"return_singer": %SINGER_CONDITIONING}
|
59 |
|
60 |
# Parameters for pitch_to_audio_utils.torch_gl:
|
61 |
# ==============================================================================
|
|
|
67 |
# ==============================================================================
|
68 |
pitch_to_audio_utils.torch_istft.nfft = %NFFT
|
69 |
|
70 |
+
# Parameters for model_diffusion.UNetPitchConditioned:
|
71 |
# ==============================================================================
|
72 |
+
model_diffusion.UNetPitchConditioned.audio_seq_len = %AUDIO_SEQ_LEN
|
73 |
+
model_diffusion.UNetPitchConditioned.cfg = True
|
74 |
+
model_diffusion.UNetPitchConditioned.cond_drop_prob = 0.2
|
75 |
+
model_diffusion.UNetPitchConditioned.dropout = 0.3
|
76 |
+
model_diffusion.UNetPitchConditioned.f0_dim = 128
|
77 |
+
model_diffusion.UNetPitchConditioned.features = [512, 640, 1024]
|
78 |
+
model_diffusion.UNetPitchConditioned.inp_dim = %NUM_MELS
|
79 |
+
model_diffusion.UNetPitchConditioned.kernel_size = 5
|
80 |
+
model_diffusion.UNetPitchConditioned.log_samples_every = 10
|
81 |
+
model_diffusion.UNetPitchConditioned.log_wandb_samples_every = 50
|
82 |
+
model_diffusion.UNetPitchConditioned.loss_w_padding = True
|
83 |
+
model_diffusion.UNetPitchConditioned.nonlinearity = 'mish'
|
84 |
+
model_diffusion.UNetPitchConditioned.norm = False
|
85 |
+
model_diffusion.UNetPitchConditioned.num_attns = 4
|
86 |
+
model_diffusion.UNetPitchConditioned.num_convs = 4
|
87 |
+
model_diffusion.UNetPitchConditioned.num_heads = 8
|
88 |
+
model_diffusion.UNetPitchConditioned.project_dim = 256
|
89 |
+
model_diffusion.UNetPitchConditioned.singer_conditioning = %SINGER_CONDITIONING
|
90 |
+
model_diffusion.UNetPitchConditioned.singer_dim = 128
|
91 |
+
model_diffusion.UNetPitchConditioned.singer_vocab = 55
|
92 |
+
model_diffusion.UNetPitchConditioned.sr = %SR
|
93 |
+
model_diffusion.UNetPitchConditioned.strides = [4, 2, 2]
|
94 |
+
model_diffusion.UNetPitchConditioned.time_dim = 128
|
requirements.txt
CHANGED
@@ -1,22 +1,4 @@
|
|
1 |
-
absl_py==1.4.0
|
2 |
-
einops==0.8.0
|
3 |
-
gin_config==0.5.0
|
4 |
-
joblib==1.2.0
|
5 |
-
librosa==0.10.0
|
6 |
-
lmdb==1.4.1
|
7 |
-
matplotlib==3.9.2
|
8 |
-
numpy==1.24.4
|
9 |
-
pandas==2.0.3
|
10 |
-
protobuf==3.20.3
|
11 |
-
pytorch_lightning==1.9.0
|
12 |
-
scikit_learn==1.2.0
|
13 |
-
setuptools==67.8.0
|
14 |
-
torch==2.4.0
|
15 |
-
torchaudio==2.4.0
|
16 |
-
tqdm==4.65.0
|
17 |
-
wandb==0.15.4
|
18 |
-
x_transformers==1.30.2
|
19 |
crepe==0.0.15
|
20 |
hmmlearn==0.3.2
|
21 |
tensorflow==2.17.0
|
22 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
crepe==0.0.15
|
2 |
hmmlearn==0.3.2
|
3 |
tensorflow==2.17.0
|
4 |
+
GaMaDHaNi @ git+https://github.com/snnithya/GaMaDHaNi.git@782dde8f48ff15a50394bcc7506df1ece0e0310e
|
src/dataset.py
DELETED
@@ -1,312 +0,0 @@
|
|
1 |
-
from typing import Callable, Dict, Optional, Tuple
|
2 |
-
import lmdb
|
3 |
-
import torch
|
4 |
-
import pdb
|
5 |
-
import numpy as np
|
6 |
-
from torch.utils.data import Dataset
|
7 |
-
from random import randint
|
8 |
-
from sklearn.preprocessing import QuantileTransformer
|
9 |
-
# from protobuf.data_example import AudioExample
|
10 |
-
import gin
|
11 |
-
import sys
|
12 |
-
import src.pitch_to_audio_utils as p2a
|
13 |
-
|
14 |
-
TensorDict = Dict[str, torch.Tensor]
|
15 |
-
|
16 |
-
@gin.configurable
|
17 |
-
class SequenceDataset(Dataset):
|
18 |
-
|
19 |
-
def __init__(
|
20 |
-
self,
|
21 |
-
db_path: str,
|
22 |
-
task_fn: Optional[Callable[[TensorDict], TensorDict]] = None,
|
23 |
-
device: Optional[torch.device] = None
|
24 |
-
) -> None:
|
25 |
-
super().__init__()
|
26 |
-
self._env = None
|
27 |
-
self._keys = None
|
28 |
-
self._db_path = db_path
|
29 |
-
self.task_fn = task_fn
|
30 |
-
self.device = device
|
31 |
-
|
32 |
-
def __len__(self):
|
33 |
-
return len(self.keys)
|
34 |
-
|
35 |
-
def __getitem__(self, index):
|
36 |
-
# pdb.set_trace()
|
37 |
-
with self.env.begin() as txn:
|
38 |
-
ae = AudioExample(txn.get(self.keys[index]))
|
39 |
-
ae = ae.as_dict()
|
40 |
-
if self.task_fn is not None:
|
41 |
-
ae = self.task_fn(ae)
|
42 |
-
if self.device is not None:
|
43 |
-
ae = {k: torch.tensor(v, device=self.device) for k, v in ae.items()}
|
44 |
-
return ae
|
45 |
-
|
46 |
-
@property
|
47 |
-
def env(self):
|
48 |
-
if self._env is None:
|
49 |
-
self._env = lmdb.open(
|
50 |
-
self._db_path,
|
51 |
-
lock=False,
|
52 |
-
readahead=False,
|
53 |
-
)
|
54 |
-
return self._env
|
55 |
-
|
56 |
-
@property
|
57 |
-
def keys(self):
|
58 |
-
if self._keys is None:
|
59 |
-
with self.env.begin(write=False) as txn:
|
60 |
-
self._keys = list(txn.cursor().iternext(values=False))
|
61 |
-
self._keys = self._keys
|
62 |
-
return self._keys
|
63 |
-
|
64 |
-
class MelPitchDataLoader(torch.utils.data.DataLoader):
|
65 |
-
def __init__(self, *args, **kwargs):
|
66 |
-
super().__init__(*args, **kwargs)
|
67 |
-
|
68 |
-
def __iter__(self):
|
69 |
-
for batch in super().__iter__():
|
70 |
-
# Apply online transform to each sample in the batch
|
71 |
-
audio, f0 = batch
|
72 |
-
|
73 |
-
# generate mel spectrogram
|
74 |
-
mel = p2a.audio_to_normalized_mels(audio) # doing mel conversion here since it is done in a batch and thus presumably faster
|
75 |
-
|
76 |
-
yield zip(mel, f0)
|
77 |
-
|
78 |
-
@gin.configurable
|
79 |
-
def pitch_read_w_downsample(
|
80 |
-
inputs: TensorDict,
|
81 |
-
seq_len: int,
|
82 |
-
decoder_key: str,
|
83 |
-
min_norm_pitch: int,
|
84 |
-
transpose_pitch: Optional[int] = None,
|
85 |
-
time_downsample: int = 1,
|
86 |
-
pitch_downsample: int = 1,
|
87 |
-
qt_transform: Optional[QuantileTransformer] = None,
|
88 |
-
min_clip: int = 200,
|
89 |
-
max_clip: int = 600,
|
90 |
-
add_noise_to_silence: bool = False
|
91 |
-
):
|
92 |
-
# pdb.set_trace()
|
93 |
-
# print(min_norm_pitch, seq_len, transpose_pitch, qt_transform)
|
94 |
-
data = inputs[decoder_key]["data"]
|
95 |
-
if seq_len is not None:
|
96 |
-
start = randint(0, data.shape[0] - seq_len*time_downsample - 1)
|
97 |
-
end = start + seq_len*time_downsample
|
98 |
-
f0 = inputs[decoder_key]['data'][start:end:time_downsample].copy()
|
99 |
-
else:
|
100 |
-
f0 = data.copy()
|
101 |
-
|
102 |
-
# normalize pitch
|
103 |
-
f0[f0 == 0] = np.nan
|
104 |
-
norm_f0 = f0.copy()
|
105 |
-
norm_f0[~np.isnan(norm_f0)] = (1200) * np.log2(norm_f0[~np.isnan(norm_f0)] / 440)
|
106 |
-
del f0
|
107 |
-
|
108 |
-
# descretize pitch
|
109 |
-
norm_f0[~np.isnan(norm_f0)] = np.around(norm_f0[~np.isnan(norm_f0)])
|
110 |
-
norm_f0[~np.isnan(norm_f0)] = norm_f0[~np.isnan(norm_f0)] - (min_norm_pitch)
|
111 |
-
|
112 |
-
norm_f0[~np.isnan(norm_f0)] = norm_f0[~np.isnan(norm_f0)] // pitch_downsample + 1 # reserve 0 for silence
|
113 |
-
# data augmentation
|
114 |
-
if transpose_pitch:
|
115 |
-
transpose_amt = randint(-transpose_pitch, transpose_pitch) # in cents
|
116 |
-
transposed_values = norm_f0[~np.isnan(norm_f0)] + (transpose_amt//pitch_downsample)
|
117 |
-
norm_f0[~np.isnan(norm_f0)] = transposed_values
|
118 |
-
|
119 |
-
# clip values HACK to change
|
120 |
-
norm_f0[~np.isnan(norm_f0)] = np.clip(norm_f0[~np.isnan(norm_f0)], min_clip, max_clip)
|
121 |
-
|
122 |
-
# add silence token of min_clip - 4
|
123 |
-
if add_noise_to_silence:
|
124 |
-
norm_f0[np.isnan(norm_f0)] = min_clip - 4 + np.clip(np.random.normal(size=norm_f0[np.isnan(norm_f0)].shape), -3, 3) # making sure noise is between -3 and 3 and thus won't spill into pitched values
|
125 |
-
else:
|
126 |
-
norm_f0[np.isnan(norm_f0)] = min_clip - 4
|
127 |
-
|
128 |
-
if qt_transform:
|
129 |
-
qt_inp = norm_f0.reshape(-1, 1)
|
130 |
-
norm_f0 = qt_transform.transform(qt_inp).reshape(-1)
|
131 |
-
|
132 |
-
return norm_f0.reshape(1, -1)
|
133 |
-
|
134 |
-
def hz_to_cents(f0, ref=440, min_norm_pitch=0, pitch_downsample=1, min_clip=200, max_clip=600, silence_token=None):
|
135 |
-
# pdb.set_trace()
|
136 |
-
f0[f0 == 0] = np.nan
|
137 |
-
norm_f0 = f0.copy()
|
138 |
-
norm_f0[~np.isnan(norm_f0)] = (1200) * np.log2(norm_f0[~np.isnan(norm_f0)] / ref)
|
139 |
-
# descretize pitch
|
140 |
-
norm_f0[~np.isnan(norm_f0)] = np.around(norm_f0[~np.isnan(norm_f0)])
|
141 |
-
norm_f0[~np.isnan(norm_f0)] = norm_f0[~np.isnan(norm_f0)] - (min_norm_pitch)
|
142 |
-
norm_f0[~np.isnan(norm_f0)] = norm_f0[~np.isnan(norm_f0)] // pitch_downsample + 1 # reserve 0 for silence
|
143 |
-
norm_f0[~np.isnan(norm_f0)] = np.clip(norm_f0[~np.isnan(norm_f0)], min_clip, max_clip) #HACK
|
144 |
-
if silence_token is not None:
|
145 |
-
norm_f0[np.isnan(norm_f0)] = silence_token
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
return norm_f0
|
150 |
-
|
151 |
-
@gin.configurable
|
152 |
-
def mel_pitch(
|
153 |
-
inputs: TensorDict,
|
154 |
-
min_norm_pitch: int,
|
155 |
-
audio_seq_len: int=None,
|
156 |
-
pitch_downsample: int = 1,
|
157 |
-
qt_transform: Optional[QuantileTransformer] = None,
|
158 |
-
min_clip: int = 200,
|
159 |
-
max_clip: int = 600,
|
160 |
-
nfft: int = 2048,
|
161 |
-
convert_audio_to_mel: bool = False
|
162 |
-
):
|
163 |
-
hop_size = nfft // 4
|
164 |
-
audio_data = inputs['audio']['data']
|
165 |
-
audio_sr = inputs['audio']['sampling_rate']
|
166 |
-
pitch_data = inputs['pitch']['data']
|
167 |
-
pitch_sr = inputs['pitch']['sampling_rate']
|
168 |
-
# pdb.set_trace()
|
169 |
-
if audio_seq_len is not None:
|
170 |
-
# if audio_seq_len is given, cuts audio/pitch else returns the entire chunk
|
171 |
-
pitch_seq_len = np.around((audio_seq_len/audio_sr) * pitch_sr ).astype(int)
|
172 |
-
pitch_start = randint(0, pitch_data.shape[0] - pitch_seq_len - 1)
|
173 |
-
pitch_end = pitch_start + pitch_seq_len
|
174 |
-
pitch_data = pitch_data[pitch_start:pitch_end]
|
175 |
-
audio_start = np.around(pitch_start * audio_sr // pitch_sr).astype(int)
|
176 |
-
audio_end = np.around(audio_start + audio_seq_len).astype(int)
|
177 |
-
# pdb.set_trace()
|
178 |
-
audio_data = audio_data[audio_start:audio_end]
|
179 |
-
else:
|
180 |
-
pitch_seq_len = np.around((audio_data.shape[0]/audio_sr) * pitch_sr ).astype(int)
|
181 |
-
audio_data = p2a.audio_to_normalized_mels(torch.Tensor(audio_data).unsqueeze(0), qt=qt_transform).numpy()[0]
|
182 |
-
|
183 |
-
pitch_data = hz_to_cents(pitch_data, min_norm_pitch=min_norm_pitch, pitch_downsample=pitch_downsample, min_clip=min_clip, max_clip=max_clip)
|
184 |
-
|
185 |
-
if audio_seq_len is not None:
|
186 |
-
# linearly interpolate pitch data to match audio sequence length, if audio_seq_len is given
|
187 |
-
pitch_inds = np.linspace(0, pitch_data.shape[0], num=audio_seq_len//hop_size, endpoint=False) #check here
|
188 |
-
pitch_data = np.interp(pitch_inds, np.arange(0, pitch_data.shape[0]), pitch_data)
|
189 |
-
|
190 |
-
# replace nan (aka silences) with min_clip - 4
|
191 |
-
pitch_data[np.isnan(pitch_data)] = min_clip - 4
|
192 |
-
|
193 |
-
return audio_data, pitch_data
|
194 |
-
def running_average(signal, window_size):
|
195 |
-
|
196 |
-
weights = np.ones(int(window_size)) / window_size
|
197 |
-
pad_width = len(weights) // 2
|
198 |
-
padded_signal = np.pad(signal, pad_width, mode='symmetric')
|
199 |
-
# Perform the convolution
|
200 |
-
smoothed_signal = np.convolve(padded_signal, weights, mode='valid')
|
201 |
-
if window_size % 2 == 0:
|
202 |
-
smoothed_signal = smoothed_signal[:-1]
|
203 |
-
return smoothed_signal
|
204 |
-
|
205 |
-
@gin.configurable
|
206 |
-
def pitch_coarse_condition(
|
207 |
-
inputs: TensorDict,
|
208 |
-
min_norm_pitch: int,
|
209 |
-
pitch_seq_len: int=None,
|
210 |
-
pitch_downsample: int = 1,
|
211 |
-
time_downsample: int = 1,
|
212 |
-
qt_transform: Optional[QuantileTransformer] = None,
|
213 |
-
min_clip: int = 200,
|
214 |
-
max_clip: int = 600,
|
215 |
-
add_noise: bool = True,
|
216 |
-
avg_window_size: float = 1 # window size in seconds
|
217 |
-
):
|
218 |
-
|
219 |
-
pitch_data = inputs['pitch']['data']
|
220 |
-
if pitch_seq_len is not None:
|
221 |
-
pitch_start = randint(0, pitch_data.shape[0] - pitch_seq_len*time_downsample - 1)
|
222 |
-
pitch_end = pitch_start + pitch_seq_len*time_downsample
|
223 |
-
pitch_data = pitch_data[pitch_start:pitch_end:time_downsample]
|
224 |
-
pitch_data = hz_to_cents(pitch_data, min_norm_pitch=min_norm_pitch, pitch_downsample=pitch_downsample, min_clip=min_clip, max_clip=max_clip)
|
225 |
-
|
226 |
-
# extract coarse pitch condition
|
227 |
-
pitch_sr = inputs['pitch']['sampling_rate'] // time_downsample
|
228 |
-
avg_pitch = running_average(pitch_data, np.around(pitch_sr * avg_window_size).astype(int))
|
229 |
-
# replace nan (aka silences) with min_clip - 4
|
230 |
-
if add_noise:
|
231 |
-
pitch_data[np.isnan(pitch_data)] = min_clip - 4 + np.clip(np.random.normal(size=pitch_data[np.isnan(pitch_data)].shape), -3, 3) # making sure noise is between -3 and 3 and thus won't spill into pitched values
|
232 |
-
avg_pitch[np.isnan(avg_pitch)] = min_clip - 4 + np.clip(np.random.normal(size=avg_pitch[np.isnan(avg_pitch)].shape), -3, 3) # making sure noise is between -3 and 3 and thus won't spill into pitched values
|
233 |
-
else:
|
234 |
-
pitch_data[np.isnan(pitch_data)] = min_clip - 4
|
235 |
-
|
236 |
-
if qt_transform:
|
237 |
-
# apply qt transform
|
238 |
-
qt_inp = pitch_data.reshape(-1, 1)
|
239 |
-
pitch_data = qt_transform.transform(qt_inp).reshape(-1)
|
240 |
-
avg_qt_inp = avg_pitch.reshape(-1, 1)
|
241 |
-
avg_pitch = qt_transform.transform(avg_qt_inp).reshape(-1)
|
242 |
-
# pdb.set_trace()
|
243 |
-
return pitch_data, avg_pitch
|
244 |
-
|
245 |
-
@gin.configurable
|
246 |
-
def mel_pitch_coarse_condition(
|
247 |
-
inputs: TensorDict,
|
248 |
-
min_norm_pitch: int,
|
249 |
-
audio_seq_len: int=None,
|
250 |
-
pitch_downsample: int = 1,
|
251 |
-
qt_transform: Optional[QuantileTransformer] = None,
|
252 |
-
min_clip: int = 200,
|
253 |
-
max_clip: int = 600,
|
254 |
-
nfft: int = 2048,
|
255 |
-
avg_window_size: float = 1 # duration of avg window in seconds
|
256 |
-
):
|
257 |
-
mel, pitch = mel_pitch(inputs, min_norm_pitch, audio_seq_len, pitch_downsample, qt_transform, min_clip, max_clip, nfft)
|
258 |
-
silence_token = min_clip - 4
|
259 |
-
avg_pitch = pitch.copy()
|
260 |
-
avg_pitch[pitch == silence_token] = np.nan
|
261 |
-
|
262 |
-
time = mel.shape[1]/inputs['audio']['sampling_rate']
|
263 |
-
pitch_sr = pitch.shape[0]/time
|
264 |
-
|
265 |
-
avg_pitch = running_average(avg_pitch, np.around(pitch_sr*avg_window_size))
|
266 |
-
avg_pitch[np.isnan(avg_pitch)] = silence_token
|
267 |
-
|
268 |
-
return mel, pitch, avg_pitch
|
269 |
-
|
270 |
-
def load_cached_audio(
|
271 |
-
inputs: TensorDict,
|
272 |
-
audio_len: Optional[float] = None,
|
273 |
-
) -> torch.Tensor:
|
274 |
-
|
275 |
-
audio_data = inputs['audio']['data']
|
276 |
-
if audio_len is not None:
|
277 |
-
audio_start = randint(0, audio_data.shape[1] - audio_len - 1)
|
278 |
-
audio_end = audio_start + audio_len
|
279 |
-
audio_data = audio_data[:, audio_start:audio_end]
|
280 |
-
return torch.Tensor(audio_data)
|
281 |
-
|
282 |
-
# need to add a silence token / range, calculate pitch avg
|
283 |
-
def load_cached_dataset(
|
284 |
-
inputs: TensorDict,
|
285 |
-
audio_len: float,
|
286 |
-
return_singer: bool = False
|
287 |
-
) -> Tuple[torch.Tensor, torch.Tensor]:
|
288 |
-
# pdb.set_trace()
|
289 |
-
audio_sr = inputs['audio']['sampling_rate']
|
290 |
-
audio_data = inputs['audio']['data']
|
291 |
-
audio_start = randint(0, audio_data.shape[1] - audio_len - 1)
|
292 |
-
audio_end = audio_start + audio_len
|
293 |
-
audio_data = audio_data[:, audio_start:audio_end]
|
294 |
-
|
295 |
-
pitch_sr = inputs['pitch']['sampling_rate']
|
296 |
-
pitch_len = np.floor(audio_len / audio_sr * pitch_sr).astype(int)
|
297 |
-
pitch_data = inputs['pitch']['data']
|
298 |
-
pitch_start = np.floor(audio_start * pitch_sr / audio_sr).astype(int)
|
299 |
-
pitch_end = pitch_start + pitch_len
|
300 |
-
pitch_data = pitch_data[pitch_start:pitch_end]
|
301 |
-
|
302 |
-
# interpolate data to match audio length
|
303 |
-
pitch_inds = np.linspace(0, pitch_data.shape[0], num=audio_len, endpoint=False) #check here
|
304 |
-
pitch_data = np.interp(pitch_inds, np.arange(0, pitch_data.shape[0]), pitch_data)
|
305 |
-
|
306 |
-
if return_singer:
|
307 |
-
singer = torch.Tensor([inputs['global_conditions']['singer']])
|
308 |
-
else:
|
309 |
-
singer = None
|
310 |
-
|
311 |
-
# print(audio_data.shape, pitch_data.shape, singer.shape if singer is not None else None)
|
312 |
-
return torch.Tensor(audio_data), torch.Tensor(pitch_data), singer
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/generate_utils.py
DELETED
@@ -1,88 +0,0 @@
|
|
1 |
-
import numpy as np
|
2 |
-
from typing import Optional
|
3 |
-
from sklearn.preprocessing import QuantileTransformer
|
4 |
-
import sys
|
5 |
-
import pdb
|
6 |
-
sys.path.append('../pitch-diffusion')
|
7 |
-
import torch
|
8 |
-
import gin
|
9 |
-
from src.model import UNet, UNetPitchConditioned
|
10 |
-
from functools import partial
|
11 |
-
import joblib
|
12 |
-
from src.dataset import hz_to_cents, pitch_read_w_downsample
|
13 |
-
|
14 |
-
def invert_pitch_read(pitch,
|
15 |
-
min_norm_pitch: int,
|
16 |
-
time_downsample: int,
|
17 |
-
pitch_downsample: int,
|
18 |
-
qt_transform: Optional[QuantileTransformer],
|
19 |
-
min_clip: int,
|
20 |
-
max_clip: int):
|
21 |
-
try:
|
22 |
-
pitch = pitch.detach().cpu().numpy()
|
23 |
-
except:
|
24 |
-
pass
|
25 |
-
if qt_transform is not None:
|
26 |
-
pitch = qt_transform.inverse_transform(pitch.reshape(-1, 1))
|
27 |
-
pitch.reshape(1, -1)
|
28 |
-
pitch[pitch < min_clip] = np.nan
|
29 |
-
pitch[~np.isnan(pitch)] = (pitch[~np.isnan(pitch)] - 1) * pitch_downsample
|
30 |
-
pitch[~np.isnan(pitch)] = pitch[~np.isnan(pitch)] + min_norm_pitch
|
31 |
-
pitch[~np.isnan(pitch)] = 440 * 2**(pitch[~np.isnan(pitch)] / 1200)
|
32 |
-
pitch[np.isnan(pitch)] = 0
|
33 |
-
|
34 |
-
return pitch, 200//time_downsample
|
35 |
-
|
36 |
-
def invert_tonic(tonic: Optional[int] = None,
|
37 |
-
min_norm_pitch: int = 0,
|
38 |
-
min_clip: int = 200,
|
39 |
-
pitch_downsample: int = 1,
|
40 |
-
):
|
41 |
-
tonic += min_clip
|
42 |
-
tonic = pitch_downsample * (tonic - 1)
|
43 |
-
tonic += min_norm_pitch
|
44 |
-
tonic = 440 * 2**(tonic / 1200)
|
45 |
-
|
46 |
-
return tonic
|
47 |
-
|
48 |
-
def load_processed_pitch(pitch,
|
49 |
-
audio_seq_len: int,
|
50 |
-
min_norm_pitch: int,
|
51 |
-
pitch_downsample: int,
|
52 |
-
min_clip: int,
|
53 |
-
max_clip: int,
|
54 |
-
):
|
55 |
-
# pdb.set_trace()
|
56 |
-
pitch = hz_to_cents(pitch, min_norm_pitch=min_norm_pitch, min_clip=min_clip, max_clip=max_clip, pitch_downsample=pitch_downsample, silence_token=min_clip-4)
|
57 |
-
pitch_inds = np.linspace(0, pitch.shape[0], num=audio_seq_len, endpoint=False)
|
58 |
-
pitch = np.interp(pitch_inds, np.arange(0, pitch.shape[0]), pitch)
|
59 |
-
return pitch
|
60 |
-
|
61 |
-
def load_pitch_model(config, ckpt, qt = None, prime_file=None, device='cuda'):
|
62 |
-
gin.parse_config_file(config)
|
63 |
-
model = UNet()
|
64 |
-
model.load_state_dict(torch.load(ckpt, map_location='cuda')['state_dict'])
|
65 |
-
model.to(device)
|
66 |
-
if qt is not None:
|
67 |
-
qt = joblib.load(qt)
|
68 |
-
if prime_file is not None:
|
69 |
-
with gin.config_scope('val'): # probably have to change this
|
70 |
-
with gin.unlock_config():
|
71 |
-
gin.bind_parameter('dataset.pitch_read_w_downsample.qt_transform', qt)
|
72 |
-
primes = np.load(prime_file, allow_pickle=True)['concatenated_array'][:, 0]
|
73 |
-
else:
|
74 |
-
primes = None
|
75 |
-
task_fn = None
|
76 |
-
task_fn = partial(pitch_read_w_downsample,
|
77 |
-
seq_len=None)
|
78 |
-
return model, qt, primes, task_fn
|
79 |
-
|
80 |
-
def load_audio_model(config, ckpt, qt = None, device='cuda'):
|
81 |
-
gin.parse_config_file(config)
|
82 |
-
model = UNetPitchConditioned() # there are no gin parameters for some reason
|
83 |
-
model.load_state_dict(torch.load(ckpt, map_location='cuda')['state_dict'])
|
84 |
-
model.to(device)
|
85 |
-
if qt is not None:
|
86 |
-
qt = joblib.load(qt)
|
87 |
-
|
88 |
-
return model, qt
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/model.py
DELETED
@@ -1,1130 +0,0 @@
|
|
1 |
-
import torch
|
2 |
-
import torch.nn as nn
|
3 |
-
import torch.optim as optim
|
4 |
-
import pytorch_lightning as pl
|
5 |
-
import torch.nn.functional as F
|
6 |
-
import math
|
7 |
-
from typing import Optional, Union
|
8 |
-
import numpy as np
|
9 |
-
import wandb
|
10 |
-
import matplotlib.pyplot as plt
|
11 |
-
import gin
|
12 |
-
import os
|
13 |
-
import pandas as pd
|
14 |
-
import src.pitch_to_audio_utils as p2a
|
15 |
-
import torchaudio
|
16 |
-
from typing import Callable
|
17 |
-
from pytorch_lightning.utilities import grad_norm
|
18 |
-
|
19 |
-
import sys
|
20 |
-
sys.path.append('..')
|
21 |
-
sys.path.append('../x-transformers/')
|
22 |
-
from src.utils import prob_mask_like
|
23 |
-
from x_transformers.x_transformers import AttentionLayers
|
24 |
-
import pdb
|
25 |
-
|
26 |
-
def get_activation(act: str = 'mish'):
|
27 |
-
act = act.lower()
|
28 |
-
if act == 'mish':
|
29 |
-
return nn.Mish()
|
30 |
-
elif act == 'relu':
|
31 |
-
return nn.ReLU()
|
32 |
-
elif act == 'leaky_relu':
|
33 |
-
return nn.LeakyReLU()
|
34 |
-
elif act == 'gelu':
|
35 |
-
return nn.GELU()
|
36 |
-
elif act == 'swish':
|
37 |
-
return nn.SiLU()
|
38 |
-
else:
|
39 |
-
raise ValueError(f'Activation {act} not supported')
|
40 |
-
|
41 |
-
def get_weight_norm(layer):
|
42 |
-
return torch.nn.utils.parametrizations.weight_norm(layer)
|
43 |
-
|
44 |
-
def get_layer(layer, norm: bool):
|
45 |
-
if norm:
|
46 |
-
return get_weight_norm(layer)
|
47 |
-
else:
|
48 |
-
return layer
|
49 |
-
|
50 |
-
class PositionalEncoding(nn.Module):
|
51 |
-
def __init__(self, dim):
|
52 |
-
super(PositionalEncoding, self).__init__()
|
53 |
-
self.dim = dim
|
54 |
-
|
55 |
-
def forward(self, x):
|
56 |
-
shape = x.shape
|
57 |
-
x = x * 100
|
58 |
-
w = torch.pow(10000, (2 * torch.arange(self.dim // 2).float() / self.dim)).to(x)
|
59 |
-
x = x.unsqueeze(-1) / w
|
60 |
-
embed = torch.cat([torch.cos(x), torch.sin(x)], -1)
|
61 |
-
embed = embed.reshape(*shape, -1)
|
62 |
-
if len(shape) == 2: # f0 embedding, else time embedding
|
63 |
-
embed = embed.permute(0, 2, 1)
|
64 |
-
return embed
|
65 |
-
|
66 |
-
class ConvBlock(nn.Module):
|
67 |
-
def __init__(self,
|
68 |
-
inp_dim,
|
69 |
-
out_dim,
|
70 |
-
kernel_size: int = 3,
|
71 |
-
stride: int = 1,
|
72 |
-
padding: Union[str, int] = "same",
|
73 |
-
norm: bool = True,
|
74 |
-
nonlinearity: Optional[str] = None,
|
75 |
-
up: bool = False,
|
76 |
-
dropout: float = 0.0,
|
77 |
-
):
|
78 |
-
super(ConvBlock, self).__init__()
|
79 |
-
self.inp_dim = inp_dim
|
80 |
-
self.out_dim = out_dim
|
81 |
-
# self.norm = norm
|
82 |
-
# pdb.set_trace()
|
83 |
-
if nonlinearity is not None:
|
84 |
-
self.nonlinearity = get_activation(nonlinearity)
|
85 |
-
else:
|
86 |
-
self.nonlinearity = None
|
87 |
-
if up:
|
88 |
-
self.conv = get_layer(nn.ConvTranspose1d(inp_dim, out_dim, kernel_size=kernel_size, stride=stride, padding=padding), norm)
|
89 |
-
else:
|
90 |
-
self.conv = get_layer(nn.Conv1d(inp_dim, out_dim, kernel_size=kernel_size, stride=stride, padding=padding), norm)
|
91 |
-
|
92 |
-
self.layers = nn.ModuleList()
|
93 |
-
if self.nonlinearity is not None:
|
94 |
-
self.layers.append(self.nonlinearity)
|
95 |
-
if dropout > 0:
|
96 |
-
self.layers.append(nn.Dropout(dropout))
|
97 |
-
self.layers.append(self.conv)
|
98 |
-
|
99 |
-
def forward(self, x):
|
100 |
-
for layer in self.layers:
|
101 |
-
x = layer(x)
|
102 |
-
return x
|
103 |
-
class UpSampleLayer(nn.Module):
|
104 |
-
def __init__(self,
|
105 |
-
inp_dim,
|
106 |
-
out_dim,
|
107 |
-
kernel_size: int = 3,
|
108 |
-
stride: int = 1,
|
109 |
-
padding: Union[str, int] = "same",
|
110 |
-
num_convs: int = 2,
|
111 |
-
norm: bool = True,
|
112 |
-
nonlinearity: Optional[str] = None,
|
113 |
-
dropout: float = 0.0,
|
114 |
-
):
|
115 |
-
super(UpSampleLayer, self).__init__()
|
116 |
-
assert num_convs > 0, "Number of convolutions must be greater than 0"
|
117 |
-
self.num_convs = num_convs
|
118 |
-
|
119 |
-
self.convs = nn.ModuleList([])
|
120 |
-
|
121 |
-
self.convs.append(ConvBlock(inp_dim, out_dim, kernel_size=stride*2, stride=stride, padding=padding, norm=norm, nonlinearity=nonlinearity, up=True)) # first convolutional layer to upsample
|
122 |
-
for ind in range(1, num_convs):
|
123 |
-
self.convs.append(ConvBlock(out_dim, out_dim, kernel_size=kernel_size, stride=1, padding="same", norm=norm, nonlinearity=nonlinearity, up=False, dropout=dropout if ind == num_convs-1 else 0))
|
124 |
-
|
125 |
-
def forward(self, x):
|
126 |
-
for conv in self.convs:
|
127 |
-
x = conv(x)
|
128 |
-
return x
|
129 |
-
|
130 |
-
class DownSampleLayer(nn.Module):
|
131 |
-
def __init__(self,
|
132 |
-
inp_dim,
|
133 |
-
out_dim,
|
134 |
-
kernel_size: int = 3,
|
135 |
-
stride: int = 1,
|
136 |
-
padding: Union[str, int] = "same",
|
137 |
-
num_convs: int = 2,
|
138 |
-
norm: bool = True,
|
139 |
-
nonlinearity: Optional[str] = None,
|
140 |
-
dropout: float = 0.0,
|
141 |
-
):
|
142 |
-
super(DownSampleLayer, self).__init__()
|
143 |
-
assert num_convs > 0, "Number of convolutions must be greater than 0"
|
144 |
-
self.num_convs = num_convs
|
145 |
-
|
146 |
-
self.convs = nn.ModuleList([])
|
147 |
-
|
148 |
-
self.convs.append(ConvBlock(inp_dim, out_dim, kernel_size=stride*2, stride=stride, padding=padding, norm=norm, nonlinearity=nonlinearity, up=False)) # first convolutional layer to upsample
|
149 |
-
for ind in range(1, num_convs):
|
150 |
-
self.convs.append(ConvBlock(out_dim, out_dim, kernel_size=kernel_size, stride=1, padding="same", norm=norm, nonlinearity=nonlinearity, up=False, dropout=dropout if ind == num_convs-1 else 0))
|
151 |
-
|
152 |
-
def forward(self, x):
|
153 |
-
for conv in self.convs:
|
154 |
-
x = conv(x)
|
155 |
-
return x
|
156 |
-
|
157 |
-
# class Attention(nn.Module):
|
158 |
-
# def __init__(self,
|
159 |
-
# num_heads,
|
160 |
-
# num_channels,
|
161 |
-
# dropout=0.0):
|
162 |
-
# super(Attention, self).__init__()
|
163 |
-
# self.num_heads = num_heads
|
164 |
-
# self.num_channels = num_channels
|
165 |
-
# self.layer_norm1 = nn.LayerNorm(self.num_channels)
|
166 |
-
# self.layer_norm2 = nn.LayerNorm(self.num_channels)
|
167 |
-
# self.qkv_proj = nn.Linear(self.num_channels, self.num_channels * 3, bias=False)
|
168 |
-
# self.head_dim = self.num_channels // self.num_heads
|
169 |
-
# self.final_proj = nn.Linear(self.num_channels, self.num_channels)
|
170 |
-
# self.dropout = nn.Dropout(dropout)
|
171 |
-
|
172 |
-
# def split_heads(self, x):
|
173 |
-
# # input shape bs, time, channels
|
174 |
-
# x = x.view(x.shape[0], x.shape[1], self.num_heads, self.head_dim)
|
175 |
-
# return x.permute(0, 2, 1, 3) # bs, num_heads, time, head_dim
|
176 |
-
|
177 |
-
# def forward(self, x):
|
178 |
-
# # pdb.set_trace()
|
179 |
-
# x = torch.permute(x, (0, 2, 1)) # bs, time, channels
|
180 |
-
# residual = x
|
181 |
-
# x = self.layer_norm1(x)
|
182 |
-
# x = self.qkv_proj(x)
|
183 |
-
# q, k, v = x.chunk(3, dim=-1)
|
184 |
-
|
185 |
-
# # split heads
|
186 |
-
# q = self.split_heads(q)
|
187 |
-
# k = self.split_heads(k)
|
188 |
-
# v = self.split_heads(v)
|
189 |
-
|
190 |
-
# # calculate attention
|
191 |
-
# x = torch.einsum("...td,...sd->...ts", q, k) / math.sqrt(self.head_dim)
|
192 |
-
# x = self.dropout(x)
|
193 |
-
# x = torch.einsum("...ts,...sd->...td", F.softmax(x, dim=-1), v) # bs, num_heads, time, head_dim
|
194 |
-
|
195 |
-
# # combine heads
|
196 |
-
# x = torch.permute(x, (0, 2, 1, 3)) # bs, time, num_heads, head_dim
|
197 |
-
# x = x.reshape(x.shape[0], x.shape[1], self.num_heads * self.head_dim)
|
198 |
-
|
199 |
-
# # final projection
|
200 |
-
# x = self.final_proj(x)
|
201 |
-
# x = self.layer_norm2(x + residual)
|
202 |
-
# return torch.permute(x, (0, 2, 1)) # bs, channels, time
|
203 |
-
|
204 |
-
class ResNetBlock(nn.Module):
|
205 |
-
def __init__(self,
|
206 |
-
in_channels: int,
|
207 |
-
out_channels: int,
|
208 |
-
dropout: float = 0.0,
|
209 |
-
nonlinearity: Optional[str] = None,
|
210 |
-
kernel_size: int = 3,
|
211 |
-
stride: int = 1,
|
212 |
-
norm: bool = True,
|
213 |
-
up: bool = False,
|
214 |
-
num_convs: int = 2,
|
215 |
-
):
|
216 |
-
super(ResNetBlock, self).__init__()
|
217 |
-
|
218 |
-
self.input_layers = nn.ModuleList([])
|
219 |
-
if nonlinearity is not None:
|
220 |
-
self.input_layers.append(get_activation(nonlinearity))
|
221 |
-
|
222 |
-
if up:
|
223 |
-
self.input_layers.append(get_layer(nn.ConvTranspose1d(in_channels, out_channels, kernel_size=stride*2, stride=stride, padding=stride//2), norm))
|
224 |
-
else:
|
225 |
-
if in_channels != out_channels:
|
226 |
-
self.input_layers.append(get_layer(nn.Conv1d(in_channels, out_channels, kernel_size=stride*2, stride=stride, padding=stride//2), norm))
|
227 |
-
elif stride > 1:
|
228 |
-
self.input_layers.append(nn.AvgPool1d(stride*2, stride=stride, padding=stride//2))
|
229 |
-
else:
|
230 |
-
self.input_layers.append(nn.Identity())
|
231 |
-
|
232 |
-
if up:
|
233 |
-
self.process_layer = UpSampleLayer(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=stride//2, num_convs=num_convs, norm=norm, nonlinearity=nonlinearity, dropout=dropout)
|
234 |
-
else:
|
235 |
-
self.process_layer = DownSampleLayer(in_channels, out_channels, kernel_size=kernel_size, stride=stride, padding=stride//2, num_convs=num_convs, norm=norm, nonlinearity=nonlinearity, dropout=dropout)
|
236 |
-
|
237 |
-
def forward(self, x):
|
238 |
-
# pdb.set_trace()
|
239 |
-
inputs = x.clone()
|
240 |
-
for layer in self.input_layers:
|
241 |
-
inputs = layer(inputs)
|
242 |
-
x = self.process_layer(x)
|
243 |
-
return x + inputs
|
244 |
-
|
245 |
-
@gin.configurable
|
246 |
-
class UNetBase(pl.LightningModule):
|
247 |
-
def __init__(self, log_grad_norms_every=10):
|
248 |
-
super(UNetBase, self).__init__()
|
249 |
-
self.log_grad_norms_every = log_grad_norms_every
|
250 |
-
|
251 |
-
@gin.configurable
|
252 |
-
def configure_optimizers(self, optimizer_cls: Callable[[], torch.optim.Optimizer],
|
253 |
-
scheduler_cls: Callable[[],
|
254 |
-
torch.optim.lr_scheduler._LRScheduler]):
|
255 |
-
# pdb.set_trace()
|
256 |
-
optimizer = optimizer_cls(self.parameters())
|
257 |
-
scheduler = scheduler_cls(optimizer)
|
258 |
-
|
259 |
-
return [optimizer], [{'scheduler': scheduler, 'interval': 'step'}]
|
260 |
-
|
261 |
-
@gin.configurable
|
262 |
-
class UNet(UNetBase):
|
263 |
-
def __init__(self,
|
264 |
-
inp_dim,
|
265 |
-
time_dim,
|
266 |
-
features,
|
267 |
-
strides,
|
268 |
-
kernel_size,
|
269 |
-
seq_len,
|
270 |
-
project_dim=None,
|
271 |
-
dropout=0.0,
|
272 |
-
nonlinearity=None,
|
273 |
-
norm=True,
|
274 |
-
num_convs=2,
|
275 |
-
num_attns=2,
|
276 |
-
num_heads=8,
|
277 |
-
log_samples_every=10,
|
278 |
-
ckpt=None,
|
279 |
-
loss_w_padding=False,
|
280 |
-
groups=None,
|
281 |
-
nfft=None,
|
282 |
-
log_grad_norms_every=10
|
283 |
-
):
|
284 |
-
super(UNet, self).__init__()
|
285 |
-
self.time_dim = time_dim
|
286 |
-
self.features = features
|
287 |
-
self.strides = strides
|
288 |
-
self.kernel_size = kernel_size
|
289 |
-
self.seq_len = seq_len
|
290 |
-
self.log_samples_every = log_samples_every
|
291 |
-
self.ckpt = ckpt
|
292 |
-
self.strides_prod = np.prod(strides)
|
293 |
-
self.loss_w_padding = loss_w_padding
|
294 |
-
|
295 |
-
if log_grad_norms_every is not None:
|
296 |
-
assert log_grad_norms_every > 0, "log_grad_norms_every must be greater than 0"
|
297 |
-
self.log_grad_norms_every = log_grad_norms_every
|
298 |
-
|
299 |
-
if project_dim is None:
|
300 |
-
project_dim = features[0]
|
301 |
-
self.initial_projection = nn.Conv1d(inp_dim, project_dim, kernel_size=1)
|
302 |
-
self.positional_encoding = PositionalEncoding(time_dim)
|
303 |
-
|
304 |
-
features = [project_dim] + features
|
305 |
-
strides = [None] + strides
|
306 |
-
|
307 |
-
self.downsample_layers = nn.ModuleList([
|
308 |
-
ResNetBlock(features[ind-1] + time_dim,
|
309 |
-
features[ind],
|
310 |
-
kernel_size=kernel_size,
|
311 |
-
stride=strides[ind],
|
312 |
-
dropout=dropout,
|
313 |
-
nonlinearity=nonlinearity,
|
314 |
-
norm=norm,
|
315 |
-
num_convs=num_convs,
|
316 |
-
) for ind in range(1, len(features))
|
317 |
-
])
|
318 |
-
|
319 |
-
# self.attention_layers = nn.ModuleList(
|
320 |
-
# [Attention(num_heads=num_heads, num_channels=features[-1], dropout=dropout) for _ in range(num_attns)]
|
321 |
-
# )
|
322 |
-
|
323 |
-
self.attention_layers = AttentionLayers(
|
324 |
-
dim = features[-1],
|
325 |
-
heads = num_heads,
|
326 |
-
depth = num_attns,
|
327 |
-
)
|
328 |
-
|
329 |
-
self.upsample_layers = nn.ModuleList([
|
330 |
-
ResNetBlock(features[ind] * 2 + time_dim, # input size defined by features + skip dimension + time dimension
|
331 |
-
features[ind-1],
|
332 |
-
kernel_size=kernel_size,
|
333 |
-
stride=strides[ind],
|
334 |
-
dropout=dropout,
|
335 |
-
nonlinearity=nonlinearity,
|
336 |
-
norm=norm,
|
337 |
-
num_convs=num_convs,
|
338 |
-
up=True
|
339 |
-
) for ind in range(len(features) - 1, 0, -1)
|
340 |
-
])
|
341 |
-
self.final_projection = nn.Conv1d(2*project_dim, inp_dim, kernel_size=1)
|
342 |
-
|
343 |
-
def pad_to(self, x, strides):
|
344 |
-
# modified from: https://stackoverflow.com/questions/66028743/how-to-handle-odd-resolutions-in-unet-architecture-pytorch
|
345 |
-
l = x.shape[-1]
|
346 |
-
|
347 |
-
if l % strides > 0:
|
348 |
-
new_l = l + strides - l % strides
|
349 |
-
else:
|
350 |
-
new_l = l
|
351 |
-
|
352 |
-
ll, ul = int((new_l-l) / 2), int(new_l-l) - int((new_l-l) / 2)
|
353 |
-
pads = (ll, ul)
|
354 |
-
|
355 |
-
out = F.pad(x, pads, "reflect").to(x)
|
356 |
-
|
357 |
-
return out, pads
|
358 |
-
|
359 |
-
def unpad(self, x, pad):
|
360 |
-
# modified from: https://stackoverflow.com/questions/66028743/how-to-handle-odd-resolutions-in-unet-architecture-pytorch
|
361 |
-
if pad[0]+pad[1] > 0:
|
362 |
-
x = x[:,:,pad[0]:-pad[1]]
|
363 |
-
return x
|
364 |
-
|
365 |
-
def forward(self, x, time):
|
366 |
-
|
367 |
-
# INITIAL PROJECTION
|
368 |
-
x = self.initial_projection(x)
|
369 |
-
|
370 |
-
# TIME CONDITIONING
|
371 |
-
time = self.positional_encoding(time)
|
372 |
-
|
373 |
-
def _concat_time(x, time):
|
374 |
-
time = time.unsqueeze(2).expand(-1, -1, x.shape[-1])
|
375 |
-
x = torch.cat([x, time], -2)
|
376 |
-
return x
|
377 |
-
|
378 |
-
skips = []
|
379 |
-
|
380 |
-
# DOWNSAMPLING
|
381 |
-
for ind, downsample_layer in enumerate(self.downsample_layers):
|
382 |
-
# print(f'Down sample layer {ind}')
|
383 |
-
skips.append(x)
|
384 |
-
x = _concat_time(x, time)
|
385 |
-
x = downsample_layer(x)
|
386 |
-
skips.append(x)
|
387 |
-
|
388 |
-
# BOTTLENECK ATTENTION
|
389 |
-
x = torch.permute(x, (0, 2, 1))
|
390 |
-
x = self.attention_layers(x)
|
391 |
-
x = torch.permute(x, (0, 2, 1))
|
392 |
-
# pdb.set_trace()
|
393 |
-
# UPSAMPLING
|
394 |
-
for ind, upsample_layer in enumerate(self.upsample_layers):
|
395 |
-
# print(f'Up sample layer {ind}')
|
396 |
-
x = _concat_time(x, time)
|
397 |
-
x = torch.cat([x, skips.pop(-1)], 1)
|
398 |
-
x = upsample_layer(x)
|
399 |
-
x = torch.cat([x, skips.pop(-1)], 1)
|
400 |
-
|
401 |
-
# FINAL PROJECTION
|
402 |
-
x = self.final_projection(x)
|
403 |
-
return x
|
404 |
-
|
405 |
-
def loss(self, x):
|
406 |
-
# pdb.set_trace()
|
407 |
-
padded_x, padding = self.pad_to(x, self.strides_prod)
|
408 |
-
t = torch.rand((padded_x.shape[0],)).to(padded_x)
|
409 |
-
noise = torch.normal(0, 1, padded_x.shape).to(padded_x)
|
410 |
-
# print(t.device, noise.device, x.device)
|
411 |
-
x_t = t[:, None, None] * padded_x + (1 - t[:, None, None]) * noise
|
412 |
-
# print(t.device, noise.device, x_t.device, x.device)
|
413 |
-
padded_y = self.forward(x_t, t)
|
414 |
-
unpadded_y = self.unpad(padded_y, padding)
|
415 |
-
|
416 |
-
if self.loss_w_padding:
|
417 |
-
target = padded_x - noise
|
418 |
-
return torch.mean((padded_y - target) ** 2)
|
419 |
-
else:
|
420 |
-
target = x - self.unpad(noise, padding) # x1 - x0
|
421 |
-
return torch.mean((unpadded_y - target) ** 2)
|
422 |
-
|
423 |
-
|
424 |
-
def on_before_optimizer_step(self, optimizer, *_):
|
425 |
-
def calculate_grad_norm(module_list, norm_type=2):
|
426 |
-
total_norm = 0
|
427 |
-
if isinstance(module_list, nn.Module):
|
428 |
-
module_list = [module_list]
|
429 |
-
for module in module_list:
|
430 |
-
for name, param in module.named_parameters():
|
431 |
-
if param.requires_grad:
|
432 |
-
param_norm = torch.norm(param.grad.detach(), p=norm_type)
|
433 |
-
total_norm += param_norm**2
|
434 |
-
# pdb.set_trace()
|
435 |
-
total_norm = torch.sqrt(total_norm)
|
436 |
-
return total_norm
|
437 |
-
|
438 |
-
if self.log_grad_norms_every is not None and self.global_step % self.log_grad_norms_every == 0:
|
439 |
-
self.log('Grad Norm/Downsample Layers', calculate_grad_norm(self.downsample_layers))
|
440 |
-
self.log('Grad Norm/Attention Layers', calculate_grad_norm(self.attention_layers))
|
441 |
-
self.log('Grad Norm/Upsample Layers', calculate_grad_norm(self.upsample_layers))
|
442 |
-
|
443 |
-
def training_step(self, batch, batch_idx):
|
444 |
-
# print('\n', batch_idx, batch.shape)
|
445 |
-
x = batch
|
446 |
-
loss = self.loss(x)
|
447 |
-
|
448 |
-
# log grad_norms
|
449 |
-
# if self.log_grad_norms_every > 0 and self.current_epoch % self.log_grad_norms_every == 0:
|
450 |
-
|
451 |
-
# for ind, attention_layer in enumerate(self.attention_layers):
|
452 |
-
# self.log(f'Grad Norm/Attention Layer {ind}', grad_norm(attention_layer, norm_type=2))
|
453 |
-
# for ind, downsample_layer in enumerate(self.downsample_layers):
|
454 |
-
# self.log(f'Grad Norm/Downsample Layer {ind}', grad_norm(downsample_layer, norm_type=2))
|
455 |
-
|
456 |
-
self.log('train_loss', loss)
|
457 |
-
|
458 |
-
return loss
|
459 |
-
|
460 |
-
def validation_step(self, batch, batch_idx):
|
461 |
-
x = batch
|
462 |
-
loss = self.loss(x)
|
463 |
-
self.log('val_loss', loss)
|
464 |
-
return loss
|
465 |
-
|
466 |
-
def sample_fn(self, batch_size: int, num_steps: int, prime: Optional[torch.Tensor] = None):
|
467 |
-
# CREATE INITIAL NOISE
|
468 |
-
if prime is not None:
|
469 |
-
prime = prime.to(self.device)
|
470 |
-
noise = torch.normal(mean=0, std=1, size=(batch_size, 1, self.seq_len)).to(self.device)
|
471 |
-
x_alpha_t = noise.clone()
|
472 |
-
t_array = torch.ones((batch_size,)).to(self.device)
|
473 |
-
# x_alpha_ts = {}
|
474 |
-
with torch.no_grad():
|
475 |
-
# SAMPLE FROM MODEL
|
476 |
-
for t in np.linspace(0, 1, num_steps + 1)[:-1]:
|
477 |
-
t_tensor = torch.tensor(t)
|
478 |
-
alpha_t = t_tensor * t_array
|
479 |
-
alpha_t = alpha_t.unsqueeze(1).unsqueeze(2).to(self.device)
|
480 |
-
if prime is not None:
|
481 |
-
x_alpha_t[:, :, :prime.shape[-1]] = ((1 - alpha_t) * noise[:, :, :prime.shape[-1]]) + (alpha_t * prime) # fill in the prime in the beginning of each x_t
|
482 |
-
diff = self.forward(x_alpha_t, t_tensor * t_array)
|
483 |
-
x_alpha_t = x_alpha_t + 1 / num_steps * diff
|
484 |
-
# x_alpha_ts[t] = x_alpha_t
|
485 |
-
# if prime is not None:
|
486 |
-
# x_alpha_t[:, :, :prime.shape[-1]] = prime
|
487 |
-
return x_alpha_t
|
488 |
-
|
489 |
-
def sample_sdedit(self, cond, batch_size, num_steps, t0=0.5):
|
490 |
-
# pdb.set_trace()
|
491 |
-
t0_steps = int(t0*num_steps)
|
492 |
-
# iterate to get x0
|
493 |
-
t_array = torch.ones((batch_size,)).to(self.device)
|
494 |
-
x_alpha_t = cond.clone()
|
495 |
-
with torch.no_grad():
|
496 |
-
for t in np.linspace(t0, 0, t0_steps + 1)[:-1]:
|
497 |
-
t_tensor = torch.tensor(t)
|
498 |
-
x_alpha_t = x_alpha_t - (1 / num_steps) * self.forward(x_alpha_t, t_tensor * t_array)
|
499 |
-
# x_alpha_t is x0 now
|
500 |
-
# iterate to get x1
|
501 |
-
for t in np.linspace(0, 1, num_steps + 1)[:-1]:
|
502 |
-
t_tensor = torch.tensor(t)
|
503 |
-
# print(unet.device, noise.device, t_tensor.device, t_array.device)
|
504 |
-
x_alpha_t = x_alpha_t + 1 / num_steps * self.forward(x_alpha_t, t_tensor * t_array)
|
505 |
-
|
506 |
-
return x_alpha_t
|
507 |
-
|
508 |
-
|
509 |
-
|
510 |
-
def on_validation_epoch_end(self) -> None:
|
511 |
-
if self.current_epoch % self.log_samples_every == 0:
|
512 |
-
samples = self.sample_fn(16, 100).detach().cpu().numpy()
|
513 |
-
if self.ckpt is not None:
|
514 |
-
os.makedirs(os.path.join(self.ckpt, 'samples', str(self.current_epoch)), exist_ok=True)
|
515 |
-
fig, axs = plt.subplots(4, 4, figsize=(16, 16))
|
516 |
-
for i in range(4):
|
517 |
-
for j in range(4):
|
518 |
-
axs[i, j].plot(samples[i*4+j].squeeze())
|
519 |
-
pd.DataFrame(samples[i*4+j].squeeze(), columns=['normalized_pitch']).to_csv(os.path.join(self.ckpt, 'samples', str(self.current_epoch), f'sample_{i*4+j}.csv'))
|
520 |
-
if self.logger:
|
521 |
-
wandb.log({"samples": [wandb.Image(fig, caption="Samples")]})
|
522 |
-
else:
|
523 |
-
fig.savefig(os.path.join(self.ckpt, 'samples', str(self.current_epoch), 'samples.png'))
|
524 |
-
plt.close(fig)
|
525 |
-
|
526 |
-
|
527 |
-
@gin.configurable
|
528 |
-
class UNetAudio(UNetBase):
|
529 |
-
def __init__(self,
|
530 |
-
inp_dim,
|
531 |
-
time_dim,
|
532 |
-
features,
|
533 |
-
strides,
|
534 |
-
kernel_size,
|
535 |
-
seq_len,
|
536 |
-
project_dim=None,
|
537 |
-
dropout=0.0,
|
538 |
-
nonlinearity=None,
|
539 |
-
norm=True,
|
540 |
-
num_convs=2,
|
541 |
-
num_attns=2,
|
542 |
-
num_heads=8,
|
543 |
-
ckpt=None,
|
544 |
-
qt = None,
|
545 |
-
log_samples_every = 10,
|
546 |
-
log_wandb_samples_every = 50,
|
547 |
-
sr=16000,
|
548 |
-
loss_w_padding=False,
|
549 |
-
log_grad_norms_every=10
|
550 |
-
):
|
551 |
-
super(UNetAudio, self).__init__()
|
552 |
-
self.inp_dim = inp_dim
|
553 |
-
self.time_dim = time_dim
|
554 |
-
self.features = features
|
555 |
-
self.strides = strides
|
556 |
-
self.kernel_size = kernel_size
|
557 |
-
self.seq_len = seq_len
|
558 |
-
self.log_samples_every = log_samples_every
|
559 |
-
self.log_wandb_samples_every = log_wandb_samples_every
|
560 |
-
self.ckpt = ckpt
|
561 |
-
self.qt = qt
|
562 |
-
self.sr = sr
|
563 |
-
self.strides_prod = np.prod(strides)
|
564 |
-
self.loss_w_padding = loss_w_padding
|
565 |
-
self.log_grad_norms_every = log_grad_norms_every
|
566 |
-
|
567 |
-
if project_dim is None:
|
568 |
-
project_dim = features[0]
|
569 |
-
self.initial_projection = nn.Conv1d(inp_dim, project_dim, kernel_size=1)
|
570 |
-
self.positional_encoding = PositionalEncoding(time_dim)
|
571 |
-
|
572 |
-
features = [project_dim] + features
|
573 |
-
strides = [None] + strides
|
574 |
-
|
575 |
-
self.downsample_layers = nn.ModuleList([
|
576 |
-
ResNetBlock(features[ind-1] + time_dim,
|
577 |
-
features[ind],
|
578 |
-
kernel_size=kernel_size,
|
579 |
-
stride=strides[ind],
|
580 |
-
dropout=dropout,
|
581 |
-
nonlinearity=nonlinearity,
|
582 |
-
norm=norm,
|
583 |
-
num_convs=num_convs,
|
584 |
-
) for ind in range(1, len(features))
|
585 |
-
])
|
586 |
-
|
587 |
-
self.attention_layers = AttentionLayers(
|
588 |
-
dim = features[-1],
|
589 |
-
heads = num_heads,
|
590 |
-
depth = num_attns,
|
591 |
-
)
|
592 |
-
|
593 |
-
self.upsample_layers = nn.ModuleList([
|
594 |
-
ResNetBlock(features[ind] * 2 + time_dim, # input size defined by features + skip dimension + time dimension
|
595 |
-
features[ind-1],
|
596 |
-
kernel_size=kernel_size,
|
597 |
-
stride=strides[ind],
|
598 |
-
dropout=dropout,
|
599 |
-
nonlinearity=nonlinearity,
|
600 |
-
norm=norm,
|
601 |
-
num_convs=num_convs,
|
602 |
-
up=True
|
603 |
-
) for ind in range(len(features) - 1, 0, -1)
|
604 |
-
])
|
605 |
-
self.final_projection = nn.Conv1d(2*project_dim, inp_dim, kernel_size=1)
|
606 |
-
self.losses = []
|
607 |
-
|
608 |
-
def forward(self, x, time):
|
609 |
-
# INITIAL PROJECTION
|
610 |
-
x = self.initial_projection(x)
|
611 |
-
|
612 |
-
# TIME CONDITIONING
|
613 |
-
time = self.positional_encoding(time)
|
614 |
-
|
615 |
-
def _concat_time(x, time):
|
616 |
-
time = time.unsqueeze(2).expand(-1, -1, x.shape[-1])
|
617 |
-
x = torch.cat([x, time], -2)
|
618 |
-
return x
|
619 |
-
|
620 |
-
skips = []
|
621 |
-
|
622 |
-
# DOWNSAMPLING
|
623 |
-
for ind, downsample_layer in enumerate(self.downsample_layers):
|
624 |
-
# print(f'Down sample layer {ind}')
|
625 |
-
skips.append(x)
|
626 |
-
x = _concat_time(x, time)
|
627 |
-
x = downsample_layer(x)
|
628 |
-
skips.append(x)
|
629 |
-
# BOTTLENECK ATTENTION
|
630 |
-
x = torch.permute(x, (0, 2, 1))
|
631 |
-
x = self.attention_layers(x)
|
632 |
-
x = torch.permute(x, (0, 2, 1))
|
633 |
-
|
634 |
-
# pdb.set_trace()
|
635 |
-
# UPSAMPLING
|
636 |
-
for ind, upsample_layer in enumerate(self.upsample_layers):
|
637 |
-
# print(f'Up sample layer {ind}')
|
638 |
-
x = _concat_time(x, time)
|
639 |
-
x = torch.cat([x, skips.pop(-1)], 1)
|
640 |
-
x = upsample_layer(x)
|
641 |
-
x = torch.cat([x, skips.pop(-1)], 1)
|
642 |
-
|
643 |
-
# FINAL PROJECTION
|
644 |
-
x = self.final_projection(x)
|
645 |
-
return x
|
646 |
-
|
647 |
-
def pad_to(self, x, strides):
|
648 |
-
# modified from: https://stackoverflow.com/questions/66028743/how-to-handle-odd-resolutions-in-unet-architecture-pytorch
|
649 |
-
l = x.shape[-1]
|
650 |
-
|
651 |
-
if l % strides > 0:
|
652 |
-
new_l = l + strides - l % strides
|
653 |
-
else:
|
654 |
-
new_l = l
|
655 |
-
|
656 |
-
ll, ul = int((new_l-l) / 2), int(new_l-l) - int((new_l-l) / 2)
|
657 |
-
pads = (ll, ul)
|
658 |
-
|
659 |
-
out = F.pad(x, pads, "reflect").to(x)
|
660 |
-
|
661 |
-
return out, pads
|
662 |
-
|
663 |
-
def unpad(self, x, pad):
|
664 |
-
# modified from: https://stackoverflow.com/questions/66028743/how-to-handle-odd-resolutions-in-unet-architecture-pytorch
|
665 |
-
if pad[0]+pad[1] > 0:
|
666 |
-
x = x[:,:,pad[0]:-pad[1]]
|
667 |
-
return x
|
668 |
-
|
669 |
-
def loss(self, x):
|
670 |
-
padded_x, padding = self.pad_to(x, self.strides_prod)
|
671 |
-
t = torch.rand((padded_x.shape[0],)).to(padded_x)
|
672 |
-
noise = torch.normal(0, 1, padded_x.shape).to(padded_x)
|
673 |
-
# print(t.device, noise.device, x.device)
|
674 |
-
x_t = t[:, None, None] * padded_x + (1 - t[:, None, None]) * noise
|
675 |
-
# print(t.device, noise.device, x_t.device, x.device)
|
676 |
-
padded_y = self.forward(x_t, t)
|
677 |
-
unpadded_y = self.unpad(padded_y, padding)
|
678 |
-
|
679 |
-
if self.loss_w_padding:
|
680 |
-
target = padded_x - noise
|
681 |
-
return torch.mean((padded_y - target) ** 2)
|
682 |
-
else:
|
683 |
-
target = x - self.unpad(noise, padding) # x1 - x0
|
684 |
-
return torch.mean((unpadded_y - target) ** 2)
|
685 |
-
|
686 |
-
def training_step(self, batch, batch_idx):
|
687 |
-
# print('\n', batch_idx, batch.shape)
|
688 |
-
x = batch
|
689 |
-
loss = self.loss(x)
|
690 |
-
self.log('train_loss', loss)
|
691 |
-
return loss
|
692 |
-
|
693 |
-
def validation_step(self, batch, batch_idx):
|
694 |
-
x = batch
|
695 |
-
loss = self.loss(x)
|
696 |
-
self.log('val_loss', loss)
|
697 |
-
return loss
|
698 |
-
|
699 |
-
def sample_fn(self, batch_size: int, num_steps: int, prime=None):
|
700 |
-
if prime is not None:
|
701 |
-
prime = prime.to(self.device)
|
702 |
-
# CREATE INITIAL NOISE
|
703 |
-
noise = torch.normal(mean=0, std=1, size=(batch_size, self.inp_dim, self.seq_len)).to(self.device)
|
704 |
-
padded_noise, padding = self.pad_to(noise, self.strides_prod)
|
705 |
-
x_alpha_t = padded_noise.clone()
|
706 |
-
t_array = torch.ones((batch_size,)).to(self.device)
|
707 |
-
with torch.no_grad():
|
708 |
-
# SAMPLE FROM MODEL
|
709 |
-
for t in np.linspace(0, 1, num_steps + 1)[:-1]:
|
710 |
-
t_tensor = torch.tensor(t)
|
711 |
-
alpha_t = t_tensor * t_array
|
712 |
-
alpha_t = alpha_t.unsqueeze(1).unsqueeze(2).to(self.device)
|
713 |
-
if prime is not None:
|
714 |
-
x_alpha_t[:, :, :prime.shape[-1]] = ((1 - alpha_t) * noise[:, :, :prime.shape[-1]]) + (alpha_t * prime) # fill in the prime in the beginning of each x_t
|
715 |
-
diff = self.forward(x_alpha_t, t_tensor * t_array)
|
716 |
-
x_alpha_t = x_alpha_t + 1 / num_steps * diff
|
717 |
-
|
718 |
-
padded_y = x_alpha_t
|
719 |
-
unpadded_y = self.unpad(padded_y, padding)
|
720 |
-
|
721 |
-
return unpadded_y
|
722 |
-
|
723 |
-
def on_validation_epoch_end(self) -> None:
|
724 |
-
if self.current_epoch % self.log_samples_every == 0:
|
725 |
-
if self.ckpt is not None:
|
726 |
-
os.makedirs(os.path.join(self.ckpt, 'samples', str(self.current_epoch)), exist_ok=True)
|
727 |
-
samples = self.sample_fn(16, 100)
|
728 |
-
audio = p2a.normalized_mels_to_audio(samples, qt=self.qt)
|
729 |
-
beep = torch.sin(2 * torch.pi * 220 * torch.arange(0, 0.1 * self.sr) / self.sr).to(audio)
|
730 |
-
concat_audios = []
|
731 |
-
for sample in audio:
|
732 |
-
concat_audios.append(torch.cat([sample, beep]))
|
733 |
-
concat_audio = torch.cat(concat_audios, dim=-1).reshape(1, -1).to('cpu')
|
734 |
-
output_file = os.path.join(self.ckpt, 'samples', f'samples_{self.current_epoch}.wav')
|
735 |
-
torchaudio.save(output_file, concat_audio, self.sr)
|
736 |
-
if self.current_epoch % self.log_wandb_samples_every == 0:
|
737 |
-
if self.logger:
|
738 |
-
wandb.log({
|
739 |
-
"samples": [wandb.Audio(output_file, self.sr, caption="Samples")]})
|
740 |
-
|
741 |
-
def on_before_optimizer_step(self, optimizer, *_):
|
742 |
-
def calculate_grad_norm(module_list, norm_type=2):
|
743 |
-
total_norm = 0
|
744 |
-
if isinstance(module_list, nn.Module):
|
745 |
-
module_list = [module_list]
|
746 |
-
for module in module_list:
|
747 |
-
for name, param in module.named_parameters():
|
748 |
-
if param.requires_grad:
|
749 |
-
param_norm = torch.norm(param.grad.detach(), p=norm_type)
|
750 |
-
total_norm += param_norm**2
|
751 |
-
# pdb.set_trace()
|
752 |
-
total_norm = torch.sqrt(total_norm)
|
753 |
-
return total_norm
|
754 |
-
|
755 |
-
if self.log_grad_norms_every is not None and self.global_step % self.log_grad_norms_every == 0:
|
756 |
-
self.log('Grad Norm/Downsample Layers', calculate_grad_norm(self.downsample_layers))
|
757 |
-
self.log('Grad Norm/Attention Layers', calculate_grad_norm(self.attention_layers))
|
758 |
-
self.log('Grad Norm/Upsample Layers', calculate_grad_norm(self.upsample_layers))
|
759 |
-
# def configure_optimizers(self):
|
760 |
-
# return optim.Adam(self.parameters(), lr=1e-4)
|
761 |
-
|
762 |
-
@gin.configurable
|
763 |
-
class UNetPitchConditioned(UNetBase):
|
764 |
-
def __init__(self,
|
765 |
-
inp_dim,
|
766 |
-
time_dim,
|
767 |
-
f0_dim,
|
768 |
-
features,
|
769 |
-
strides,
|
770 |
-
kernel_size,
|
771 |
-
audio_seq_len,
|
772 |
-
project_dim=None,
|
773 |
-
dropout=0.0,
|
774 |
-
nonlinearity=None,
|
775 |
-
norm=True,
|
776 |
-
num_convs=2,
|
777 |
-
num_attns=2,
|
778 |
-
num_heads=8,
|
779 |
-
log_samples_every=10,
|
780 |
-
log_wandb_samples_every=10,
|
781 |
-
ckpt=None,
|
782 |
-
val_data=None,
|
783 |
-
qt=None,
|
784 |
-
singer_conditioning=False,
|
785 |
-
singer_dim=128,
|
786 |
-
singer_vocab=56,
|
787 |
-
sr = 44100,
|
788 |
-
cfg = False,
|
789 |
-
f0_mask = 0,
|
790 |
-
cond_drop_prob = 0.0,
|
791 |
-
groups = None,
|
792 |
-
nfft = None,
|
793 |
-
loss_w_padding = False,
|
794 |
-
log_grad_norms_every=10
|
795 |
-
):
|
796 |
-
super(UNetPitchConditioned, self).__init__()
|
797 |
-
self.inp_dim = inp_dim
|
798 |
-
self.time_dim = time_dim
|
799 |
-
self.features = features
|
800 |
-
self.strides = strides
|
801 |
-
self.kernel_size = kernel_size
|
802 |
-
self.seq_len = audio_seq_len
|
803 |
-
self.log_samples_every = log_samples_every
|
804 |
-
self.log_wandb_samples_every = log_wandb_samples_every
|
805 |
-
self.ckpt = ckpt
|
806 |
-
self.qt = qt
|
807 |
-
self.singer_conditioning = singer_conditioning
|
808 |
-
self.sr = sr # used for logging audio to wandb
|
809 |
-
self.cond_drop_prob = cond_drop_prob
|
810 |
-
self.f0_masked_token = f0_mask
|
811 |
-
self.cfg = cfg
|
812 |
-
self.strides_prod = np.prod(strides)
|
813 |
-
self.loss_w_padding = loss_w_padding
|
814 |
-
self.log_grad_norms_every = log_grad_norms_every
|
815 |
-
|
816 |
-
conditioning_dim = time_dim
|
817 |
-
if singer_conditioning:
|
818 |
-
conditioning_dim += singer_dim
|
819 |
-
|
820 |
-
if project_dim is None:
|
821 |
-
project_dim = features[0]
|
822 |
-
self.initial_projection = nn.Conv1d(inp_dim, project_dim, kernel_size=1)
|
823 |
-
self.time_positional_encoding = PositionalEncoding(time_dim)
|
824 |
-
self.f0_positional_encoding = PositionalEncoding(f0_dim)
|
825 |
-
|
826 |
-
if singer_conditioning:
|
827 |
-
self.singer_embedding = nn.Embedding(singer_vocab + 1*self.cfg, singer_dim) # if cfg, add 1 to the singer vocabulary
|
828 |
-
self.singer_masked_token = singer_vocab
|
829 |
-
else:
|
830 |
-
self.singer_embedding = None
|
831 |
-
|
832 |
-
features = [project_dim] + features
|
833 |
-
f0_features = features.copy()
|
834 |
-
f0_features[0] = f0_dim # first layer should be the f0 dimension
|
835 |
-
strides = [None] + strides
|
836 |
-
|
837 |
-
self.downsample_layers = nn.ModuleList([
|
838 |
-
ResNetBlock(features[ind-1] + conditioning_dim,
|
839 |
-
features[ind],
|
840 |
-
kernel_size=kernel_size,
|
841 |
-
stride=strides[ind],
|
842 |
-
dropout=dropout,
|
843 |
-
nonlinearity=nonlinearity,
|
844 |
-
norm=norm,
|
845 |
-
num_convs=num_convs,
|
846 |
-
) for ind in range(1, len(features))
|
847 |
-
])
|
848 |
-
|
849 |
-
self.f0_conv_layers = nn.ModuleList([
|
850 |
-
nn.Conv1d(
|
851 |
-
f0_dim,
|
852 |
-
f0_dim,
|
853 |
-
kernel_size=2 * strides[ind],
|
854 |
-
stride=strides[ind],
|
855 |
-
padding=strides[ind]//2,
|
856 |
-
) for ind in range(1, len(features))
|
857 |
-
])
|
858 |
-
|
859 |
-
self.attention_layers = AttentionLayers(
|
860 |
-
dim = features[-1],
|
861 |
-
heads = num_heads,
|
862 |
-
depth = num_attns,
|
863 |
-
)
|
864 |
-
|
865 |
-
self.upsample_layers = nn.ModuleList([
|
866 |
-
ResNetBlock((features[ind] * 2) + (conditioning_dim) + f0_dim, # input size defined by features + skip dimension + time dimension
|
867 |
-
features[ind-1],
|
868 |
-
kernel_size=kernel_size,
|
869 |
-
stride=strides[ind],
|
870 |
-
dropout=dropout,
|
871 |
-
nonlinearity=nonlinearity,
|
872 |
-
norm=norm,
|
873 |
-
num_convs=num_convs,
|
874 |
-
up=True
|
875 |
-
) for ind in range(len(features) - 1, 0, -1)
|
876 |
-
])
|
877 |
-
self.final_projection = nn.Conv1d(2*project_dim + f0_dim, inp_dim, kernel_size=1)
|
878 |
-
# save 16 f0 values from to sample on
|
879 |
-
if val_data is not None:
|
880 |
-
val_ids = np.random.choice(len(val_data), 16)
|
881 |
-
val_samples = [val_data[i] for i in val_ids]
|
882 |
-
self.val_f0 = torch.stack([v[1] for v in val_samples], 0).to(self.device)
|
883 |
-
if self.singer_conditioning:
|
884 |
-
self.val_singer = torch.tensor([v[2] for v in val_samples]).long().to(self.device)
|
885 |
-
else:
|
886 |
-
self.val_singer = None
|
887 |
-
val_audio = torch.stack([v[0] for v in val_samples], 0).to(self.device)
|
888 |
-
if self.ckpt is not None:
|
889 |
-
# log the f0 and audio to wandb
|
890 |
-
os.makedirs(os.path.join(self.ckpt, 'samples'), exist_ok=True)
|
891 |
-
concat_audios = []
|
892 |
-
beep = torch.sin(2 * torch.pi * 220 * torch.arange(0, 0.1 * self.sr) / self.sr).to(val_audio)
|
893 |
-
recon_audios = p2a.normalized_mels_to_audio(val_audio, qt=self.qt)
|
894 |
-
fig, axs = plt.subplots(4, 4, figsize=(16, 16))
|
895 |
-
for i in range(4):
|
896 |
-
for j in range(4):
|
897 |
-
axs[i, j].plot(self.val_f0[i*4+j].squeeze())
|
898 |
-
if self.singer_conditioning:
|
899 |
-
axs[i, j].set_title(f'Singer {self.val_singer[i*4+j].item()}')
|
900 |
-
concat_audios.append(torch.cat((recon_audios[i*4+j].squeeze(), beep)))
|
901 |
-
concat_audios = torch.cat(concat_audios, dim=-1).reshape(1, -1).to('cpu')
|
902 |
-
output_file = os.path.join(self.ckpt, 'samples', f'gt_samples.wav')
|
903 |
-
torchaudio.save(output_file, concat_audios, self.sr)
|
904 |
-
|
905 |
-
try:
|
906 |
-
wandb.log({"sample f0 input": [wandb.Image(fig, caption="f0 conditioning on samples")]})
|
907 |
-
wandb.log({
|
908 |
-
"sample audio ground truth": [wandb.Audio(output_file, self.sr, caption="Samples")]})
|
909 |
-
except:
|
910 |
-
pass
|
911 |
-
|
912 |
-
fig.savefig(os.path.join(self.ckpt, 'samples', 'f0_inputs.png'))
|
913 |
-
|
914 |
-
def pad_to(self, x, strides):
|
915 |
-
# modified from: https://stackoverflow.com/questions/66028743/how-to-handle-odd-resolutions-in-unet-architecture-pytorch
|
916 |
-
l = x.shape[-1]
|
917 |
-
|
918 |
-
if l % strides > 0:
|
919 |
-
new_l = l + strides - l % strides
|
920 |
-
else:
|
921 |
-
new_l = l
|
922 |
-
|
923 |
-
ll, ul = int((new_l-l) / 2), int(new_l-l) - int((new_l-l) / 2)
|
924 |
-
pads = (ll, ul)
|
925 |
-
|
926 |
-
out = F.pad(x, pads, "reflect").to(x)
|
927 |
-
|
928 |
-
return out, pads
|
929 |
-
|
930 |
-
def unpad(self, x, pad):
|
931 |
-
# modified from: https://stackoverflow.com/questions/66028743/how-to-handle-odd-resolutions-in-unet-architecture-pytorch
|
932 |
-
if pad[0]+pad[1] > 0:
|
933 |
-
x = x[:,:,pad[0]:-pad[1]]
|
934 |
-
return x
|
935 |
-
|
936 |
-
def forward(self, x, time, f0, singer, drop_tokens=True, drop_all=False):
|
937 |
-
# INITIAL PROJECTION
|
938 |
-
x = self.initial_projection(x)
|
939 |
-
|
940 |
-
bs = x.shape[0]
|
941 |
-
if self.cfg:
|
942 |
-
# pdb.set_trace()
|
943 |
-
if drop_all:
|
944 |
-
prob_keep_mask_pitch = torch.zeros((bs)).unsqueeze(1).repeat(1, f0.shape[1]).to(self.device).bool()
|
945 |
-
prob_keep_mask_singer = torch.zeros((bs)).to(self.device).bool()
|
946 |
-
elif drop_tokens:
|
947 |
-
prob_keep_mask_pitch = prob_mask_like((bs), 1. - self.cond_drop_prob, device = self.device).unsqueeze(1).repeat(1, f0.shape[1])
|
948 |
-
prob_keep_mask_singer = prob_mask_like((bs), 1. - self.cond_drop_prob, device = self.device)
|
949 |
-
else:
|
950 |
-
prob_keep_mask_pitch = torch.ones((bs)).unsqueeze(1).repeat(1, f0.shape[1]).to(self.device).bool()
|
951 |
-
prob_keep_mask_singer = torch.ones((bs)).to(self.device).bool()
|
952 |
-
f0 = torch.where(prob_keep_mask_pitch, f0, torch.empty((f0.shape[0], f0.shape[1])).fill_(self.f0_masked_token).to(self.device).long())
|
953 |
-
if self.singer_conditioning:
|
954 |
-
singer = torch.where(prob_keep_mask_singer, singer, torch.empty((bs)).fill_(self.singer_masked_token).to(self.device).long())
|
955 |
-
|
956 |
-
# TIME and F0 CONDITIONING
|
957 |
-
conditions = [self.time_positional_encoding(time)]
|
958 |
-
if self.singer_conditioning:
|
959 |
-
conditions.append(self.singer_embedding(singer))
|
960 |
-
f0 = self.f0_positional_encoding(f0)
|
961 |
-
|
962 |
-
def _concat_condition(x, condition):
|
963 |
-
condition = condition.unsqueeze(2).expand(-1, -1, x.shape[-1])
|
964 |
-
x = torch.cat([x, condition], -2)
|
965 |
-
return x
|
966 |
-
|
967 |
-
skips = []
|
968 |
-
|
969 |
-
# DOWNSAMPLING
|
970 |
-
# pdb.set_trace()
|
971 |
-
for ind, downsample_layer in enumerate(self.downsample_layers):
|
972 |
-
# print(f'Down sample layer {ind}')
|
973 |
-
# pdb.set_trace()
|
974 |
-
skips.append(torch.cat([x, f0], -2))
|
975 |
-
for cond in conditions:
|
976 |
-
x = _concat_condition(x, cond)
|
977 |
-
# print(x.shape, time.shape, f0.shape, skips[-1].shape)
|
978 |
-
x = downsample_layer(x)
|
979 |
-
f0 = self.f0_conv_layers[ind](f0)
|
980 |
-
skips.append(torch.cat([x, f0], -2))
|
981 |
-
# BOTTLENECK ATTENTION
|
982 |
-
x = torch.permute(x, (0, 2, 1))
|
983 |
-
x = self.attention_layers(x)
|
984 |
-
x = torch.permute(x, (0, 2, 1))
|
985 |
-
# print(x.shape, time.shape, f0.shape, skips[-1].shape)
|
986 |
-
# pdb.set_trace()
|
987 |
-
# UPSAMPLING
|
988 |
-
for ind, upsample_layer in enumerate(self.upsample_layers):
|
989 |
-
# print(f'Up sample layer {ind}')
|
990 |
-
for cond in conditions:
|
991 |
-
x = _concat_condition(x, cond)
|
992 |
-
x = torch.cat([x, skips.pop(-1)], 1)
|
993 |
-
# print(x.shape, time.shape, f0.shape)
|
994 |
-
x = upsample_layer(x)
|
995 |
-
x = torch.cat([x, skips.pop(-1)], 1)
|
996 |
-
|
997 |
-
# FINAL PROJECTION
|
998 |
-
x = self.final_projection(x)
|
999 |
-
return x
|
1000 |
-
|
1001 |
-
def loss(self, x, f0, singer, drop_tokens):
|
1002 |
-
# pdb.set_trace()
|
1003 |
-
padded_x, padding = self.pad_to(x, self.strides_prod)
|
1004 |
-
padded_f0, _ = self.pad_to(f0, self.strides_prod)
|
1005 |
-
t = torch.rand((padded_x.shape[0],)).to(padded_x)
|
1006 |
-
noise = torch.normal(0, 1, padded_x.shape).to(padded_x)
|
1007 |
-
# print(t.device, noise.device, x.device)
|
1008 |
-
x_t = t[:, None, None] * padded_x + (1 - t[:, None, None]) * noise
|
1009 |
-
# print(t.device, noise.device, x_t.device, x.device)
|
1010 |
-
padded_y = self.forward(x_t, t, padded_f0, singer, drop_tokens)
|
1011 |
-
unpadded_y = self.unpad(padded_y, padding)
|
1012 |
-
|
1013 |
-
if self.loss_w_padding:
|
1014 |
-
target = padded_x - noise
|
1015 |
-
return torch.mean((padded_y - target) ** 2)
|
1016 |
-
else:
|
1017 |
-
target = x - self.unpad(noise, padding) # x1 - x0
|
1018 |
-
return torch.mean((unpadded_y - target) ** 2)
|
1019 |
-
|
1020 |
-
def training_step(self, batch, batch_idx):
|
1021 |
-
# print('\n', batch_idx, batch.shape)
|
1022 |
-
x, f0, singer = batch
|
1023 |
-
x = x.to(self.device)
|
1024 |
-
f0 = f0.to(self.device)
|
1025 |
-
singer = singer.reshape(-1).long().to(self.device) if self.singer_conditioning else None
|
1026 |
-
loss = self.loss(x, f0, singer, drop_tokens=True)
|
1027 |
-
self.log('train_loss', loss, batch_size=x.shape[0])
|
1028 |
-
return loss
|
1029 |
-
|
1030 |
-
def validation_step(self, batch, batch_idx):
|
1031 |
-
# pdb.set_trace()
|
1032 |
-
x, f0, singer = batch
|
1033 |
-
x = x.to(self.device)
|
1034 |
-
f0 = f0.to(self.device)
|
1035 |
-
singer = singer.reshape(-1).long().to(self.device) if self.singer_conditioning else None
|
1036 |
-
loss = self.loss(x, f0, singer, drop_tokens=False)
|
1037 |
-
self.log('val_loss', loss, batch_size=x.shape[0])
|
1038 |
-
return loss
|
1039 |
-
|
1040 |
-
def sample_fn(self, f0, singer, batch_size: int, num_steps: int):
|
1041 |
-
# CREATE INITIAL NOISE
|
1042 |
-
noise = torch.normal(mean=0, std=1, size=(batch_size, self.inp_dim, self.seq_len)).to(self.device)
|
1043 |
-
padded_noise, padding = self.pad_to(noise, self.strides_prod)
|
1044 |
-
t_array = torch.ones((batch_size,)).to(self.device)
|
1045 |
-
f0 = f0.to(self.device)
|
1046 |
-
padded_f0, _ = self.pad_to(f0, self.strides_prod)
|
1047 |
-
singer = singer.to(self.device)
|
1048 |
-
with torch.no_grad():
|
1049 |
-
# SAMPLE FROM MODEL
|
1050 |
-
for t in np.linspace(0, 1, num_steps + 1)[:-1]:
|
1051 |
-
t_tensor = torch.tensor(t)
|
1052 |
-
padded_noise = padded_noise + 1 / num_steps * self.forward(padded_noise, t_tensor * t_array, padded_f0, singer, drop_tokens=False)
|
1053 |
-
noise = self.unpad(padded_noise, padding)
|
1054 |
-
return noise
|
1055 |
-
|
1056 |
-
def sample_cfg(self, batch_size: int, num_steps: int, f0=None, singer=[4, 25, 45, 32], strength=1):
|
1057 |
-
# CREATE INITIAL NOISE
|
1058 |
-
noise = torch.normal(mean=0, std=1, size=(batch_size, self.inp_dim, self.seq_len)).to(self.device)
|
1059 |
-
padded_noise, padding = self.pad_to(noise, self.strides_prod)
|
1060 |
-
t_array = torch.ones((batch_size,)).to(self.device)
|
1061 |
-
if f0 is None:
|
1062 |
-
val_idx = np.random.choice(len(self.val_dataloader), batch_size)
|
1063 |
-
val_samples = [self.val_dataloader[i][1] for i in val_idx]
|
1064 |
-
f0 = torch.stack([sample for sample in val_samples]).to(self.device)
|
1065 |
-
else:
|
1066 |
-
assert len(f0) == batch_size
|
1067 |
-
f0 = f0.to(self.device)
|
1068 |
-
singer = singer.to(self.device)
|
1069 |
-
# f0 = torch.tensor(f0).to(self.device)
|
1070 |
-
# singer = torch.Tensor(np.choice(singer, batch_size, replace=True)).to(self.device)
|
1071 |
-
padded_f0, _ = self.pad_to(f0, self.strides_prod)
|
1072 |
-
with torch.no_grad():
|
1073 |
-
# SAMPLE FROM MODEL
|
1074 |
-
for t in np.linspace(0, 1, num_steps + 1)[:-1]:
|
1075 |
-
t_tensor = torch.tensor(t)
|
1076 |
-
unconditioned_logits = self.forward(padded_noise, t_tensor * t_array, padded_f0, singer, drop_tokens=False, drop_all=True)
|
1077 |
-
conditioned_logits = self.forward(padded_noise, t_tensor * t_array, padded_f0, singer, drop_tokens=False, drop_all=False)
|
1078 |
-
total_logits = strength * conditioned_logits + (1 - strength) * unconditioned_logits
|
1079 |
-
padded_noise = padded_noise + 1 / num_steps * total_logits
|
1080 |
-
|
1081 |
-
noise = self.unpad(padded_noise, padding)
|
1082 |
-
return noise, f0, singer
|
1083 |
-
|
1084 |
-
def on_validation_epoch_end(self) -> None:
|
1085 |
-
with torch.no_grad():
|
1086 |
-
# pdb.set_trace()
|
1087 |
-
if self.current_epoch % self.log_samples_every == 0:
|
1088 |
-
samples = self.sample_fn(self.val_f0, self.val_singer, 16, 100)
|
1089 |
-
if self.ckpt is not None:
|
1090 |
-
audio = p2a.normalized_mels_to_audio(samples, qt=self.qt)
|
1091 |
-
beep = torch.sin(2 * torch.pi * 220 * torch.arange(0, 0.1 * self.sr) / self.sr).to(audio)
|
1092 |
-
concat_audio = []
|
1093 |
-
for sample in audio:
|
1094 |
-
concat_audio.append(torch.cat([sample, beep]))
|
1095 |
-
concat_audio = torch.cat(concat_audio, dim=-1).reshape(1, -1).to('cpu')
|
1096 |
-
output_file = os.path.join(self.ckpt, 'samples', f'samples_{self.current_epoch}.wav')
|
1097 |
-
torchaudio.save(output_file, concat_audio, self.sr)
|
1098 |
-
if self.current_epoch % self.log_wandb_samples_every == 0:
|
1099 |
-
if self.logger:
|
1100 |
-
wandb.log({
|
1101 |
-
"samples": [wandb.Audio(output_file, self.sr, caption="Samples")]},
|
1102 |
-
step = self.global_step)
|
1103 |
-
def on_before_optimizer_step(self, optimizer, *_):
|
1104 |
-
def calculate_grad_norm(module_list, norm_type=2):
|
1105 |
-
total_norm = 0
|
1106 |
-
if isinstance(module_list, nn.Module):
|
1107 |
-
module_list = [module_list]
|
1108 |
-
for module in module_list:
|
1109 |
-
for name, param in module.named_parameters():
|
1110 |
-
if param.requires_grad:
|
1111 |
-
param_norm = torch.norm(param.grad.detach(), p=norm_type)
|
1112 |
-
total_norm += param_norm**2
|
1113 |
-
# pdb.set_trace()
|
1114 |
-
total_norm = torch.sqrt(total_norm)
|
1115 |
-
return total_norm
|
1116 |
-
|
1117 |
-
if self.log_grad_norms_every is not None and self.global_step % self.log_grad_norms_every == 0:
|
1118 |
-
self.log('Grad Norm/Downsample Layers', calculate_grad_norm(self.downsample_layers))
|
1119 |
-
self.log('Grad Norm/Attention Layers', calculate_grad_norm(self.attention_layers))
|
1120 |
-
self.log('Grad Norm/Upsample Layers', calculate_grad_norm(self.upsample_layers))
|
1121 |
-
|
1122 |
-
# @gin.configurable
|
1123 |
-
# def configure_optimizers(self, optimizer_cls: Callable[[], torch.optim.Optimizer],
|
1124 |
-
# scheduler_cls: Callable[[],
|
1125 |
-
# torch.optim.lr_scheduler._LRScheduler]):
|
1126 |
-
# # pdb.set_trace()
|
1127 |
-
# optimizer = optimizer_cls(self.parameters())
|
1128 |
-
# scheduler = scheduler_cls(optimizer)
|
1129 |
-
|
1130 |
-
# return [optimizer], [{'scheduler': scheduler, 'interval': 'step'}]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/pitch_to_audio_utils.py
DELETED
@@ -1,121 +0,0 @@
|
|
1 |
-
import math
|
2 |
-
import librosa as li
|
3 |
-
import torch
|
4 |
-
from tqdm import tqdm
|
5 |
-
import numpy as np
|
6 |
-
import gin
|
7 |
-
import logging
|
8 |
-
|
9 |
-
import pdb
|
10 |
-
|
11 |
-
@gin.configurable
|
12 |
-
def torch_stft(x, nfft):
|
13 |
-
window = torch.hann_window(nfft).to(x)
|
14 |
-
x = torch.stft(
|
15 |
-
x,
|
16 |
-
n_fft=nfft,
|
17 |
-
hop_length=nfft // 4,
|
18 |
-
win_length=nfft,
|
19 |
-
window=window,
|
20 |
-
center=True,
|
21 |
-
return_complex=True,
|
22 |
-
)
|
23 |
-
x = 2 * x / torch.mean(window)
|
24 |
-
return x
|
25 |
-
|
26 |
-
@gin.configurable
|
27 |
-
def torch_istft(x, nfft):
|
28 |
-
# pdb.set_trace()
|
29 |
-
window = torch.hann_window(nfft).to(x.device)
|
30 |
-
x = x / 2 * torch.mean(window)
|
31 |
-
return torch.istft(
|
32 |
-
x,
|
33 |
-
n_fft=nfft,
|
34 |
-
hop_length=nfft // 4,
|
35 |
-
win_length=nfft,
|
36 |
-
window=window,
|
37 |
-
center=True,
|
38 |
-
)
|
39 |
-
|
40 |
-
@gin.configurable
|
41 |
-
def to_mels(stft, nfft, num_mels, sr, eps=1e-2):
|
42 |
-
mels = li.filters.mel(
|
43 |
-
sr=sr,
|
44 |
-
n_fft=nfft,
|
45 |
-
n_mels=num_mels,
|
46 |
-
fmin=40,
|
47 |
-
)
|
48 |
-
# pdb.set_trace()
|
49 |
-
mels = torch.from_numpy(mels).to(stft)
|
50 |
-
mel_stft = torch.einsum("mf,bft->bmt", mels, stft)
|
51 |
-
mel_stft = torch.log(mel_stft + eps)
|
52 |
-
return mel_stft
|
53 |
-
|
54 |
-
@gin.configurable
|
55 |
-
def from_mels(mel_stft, nfft, num_mels, sr, eps=1e-2):
|
56 |
-
mels = li.filters.mel(
|
57 |
-
sr=sr,
|
58 |
-
n_fft=nfft,
|
59 |
-
n_mels=num_mels,
|
60 |
-
fmin=40,
|
61 |
-
)
|
62 |
-
mels = torch.from_numpy(mels).to(mel_stft)
|
63 |
-
mels = torch.pinverse(mels)
|
64 |
-
mel_stft = torch.exp(mel_stft) - eps
|
65 |
-
stft = torch.einsum("fm,bmt->bft", mels, mel_stft)
|
66 |
-
return stft
|
67 |
-
|
68 |
-
@gin.configurable
|
69 |
-
def torch_gl(stft, nfft, sr, n_iter):
|
70 |
-
|
71 |
-
def _gl_iter(phase, xs, stft):
|
72 |
-
del xs
|
73 |
-
# pdb.set_trace()
|
74 |
-
c_stft = stft * torch.exp(1j * phase)
|
75 |
-
rec = torch_istft(c_stft, nfft)
|
76 |
-
r_stft = torch_stft(rec, nfft)
|
77 |
-
phase = torch.angle(r_stft)
|
78 |
-
return phase, None
|
79 |
-
|
80 |
-
phase = torch.rand_like(stft) * 2 * torch.pi
|
81 |
-
|
82 |
-
for _ in tqdm(range(n_iter)):
|
83 |
-
phase, _ = _gl_iter(phase, None, stft)
|
84 |
-
|
85 |
-
c_stft = stft * torch.exp(1j * phase)
|
86 |
-
audio = torch_istft(c_stft, nfft)
|
87 |
-
|
88 |
-
return audio
|
89 |
-
|
90 |
-
@gin.configurable
|
91 |
-
def normalize(x, qt=None):
|
92 |
-
x_flat = x.reshape(-1, 1)
|
93 |
-
if qt is None:
|
94 |
-
logging.warning('No quantile transformer found, returning input')
|
95 |
-
return x
|
96 |
-
return torch.Tensor(qt.transform(x_flat).reshape(x.shape))
|
97 |
-
|
98 |
-
@gin.configurable
|
99 |
-
def unnormalize(x, qt=None):
|
100 |
-
x_flat = x.reshape(-1, 1)
|
101 |
-
if qt is None:
|
102 |
-
logging.warning('No quantile transformer found, returning input')
|
103 |
-
return x
|
104 |
-
if isinstance(x_flat, torch.Tensor):
|
105 |
-
x_flat = x_flat.detach().cpu().numpy()
|
106 |
-
return torch.Tensor(qt.inverse_transform(x_flat).reshape(x.shape))
|
107 |
-
|
108 |
-
@gin.configurable
|
109 |
-
def audio_to_normalized_mels(x, nfft, num_mels, sr, qt):
|
110 |
-
# pdb.set_trace()
|
111 |
-
stfts = torch_stft(x, nfft=nfft).abs()[..., :-1]
|
112 |
-
mel_stfts = to_mels(stfts, nfft, num_mels, sr)
|
113 |
-
return normalize(mel_stfts, qt).to(x)
|
114 |
-
|
115 |
-
@gin.configurable
|
116 |
-
def normalized_mels_to_audio(x, nfft, num_mels, sr, qt, n_iter=20):
|
117 |
-
x = unnormalize(x, qt).to(x)
|
118 |
-
x = from_mels(x, nfft, num_mels, sr)
|
119 |
-
x = torch.clamp(x, 0, nfft)
|
120 |
-
x = torch_gl(x, nfft, sr, n_iter=n_iter)
|
121 |
-
return x
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/preprocess_utils.py
DELETED
@@ -1,127 +0,0 @@
|
|
1 |
-
import subprocess
|
2 |
-
import numpy as np
|
3 |
-
import pandas as pd
|
4 |
-
from typing import Iterable, Tuple, Callable
|
5 |
-
import multiprocessing
|
6 |
-
import functools
|
7 |
-
from itertools import repeat
|
8 |
-
from protobuf.data_example import AudioExample, DTYPE_TO_PRECISION
|
9 |
-
import librosa
|
10 |
-
import pdb
|
11 |
-
# from memory_profiler import profile
|
12 |
-
|
13 |
-
# @profile
|
14 |
-
def load_chunk(
|
15 |
-
row: pd.Series,
|
16 |
-
n_signal_audio: int,
|
17 |
-
n_signal_pitch: int,
|
18 |
-
sr_audio: int,
|
19 |
-
sr_pitch: int,
|
20 |
-
error_path: str = None,
|
21 |
-
) -> Iterable[np.ndarray]:
|
22 |
-
audio_path = row['audio_path']
|
23 |
-
csv_path = row['pitch_path']
|
24 |
-
# print (audio_path, csv_path)
|
25 |
-
# pdb.set_trace()
|
26 |
-
try:
|
27 |
-
chunk_csv = pd.read_csv(csv_path, chunksize=n_signal_pitch)
|
28 |
-
except:
|
29 |
-
if error_path is not None:
|
30 |
-
with open(error_path, 'a') as f:
|
31 |
-
f.write(f'Error reading {csv_path}\n')
|
32 |
-
return
|
33 |
-
chunk_iter = iter(chunk_csv)
|
34 |
-
|
35 |
-
chunk_pitch = next(chunk_iter)
|
36 |
-
f0 = chunk_pitch['filtered_f0'].fillna(0).to_numpy()
|
37 |
-
|
38 |
-
# print('Number of chunks: ', pd.read_csv(csv_path).shape[0]//n_signal_pitch, '\n')
|
39 |
-
while len(f0) == n_signal_pitch:
|
40 |
-
start_time = chunk_pitch['time'].values[0]
|
41 |
-
# print(start_time, chunk_pitch['time'].values[-1] - ((n_signal_pitch - 1)/sr_pitch))
|
42 |
-
assert abs(start_time - (chunk_pitch['time'].values[-1] - ((n_signal_pitch - 1)/sr_pitch))) < 1e-6 # check that no time stamps were skipped
|
43 |
-
chunk_audio = librosa.load(audio_path, sr=sr_audio, offset=start_time, duration=n_signal_audio/sr_audio, dtype=np.float32)[0]
|
44 |
-
assert chunk_audio.shape[0] == n_signal_audio
|
45 |
-
# and len(f0) == n_signal_pitch:
|
46 |
-
# chunk_audio /= 2**15
|
47 |
-
# pdb.set_trace()
|
48 |
-
yield chunk_audio, f0, row, start_time
|
49 |
-
try:
|
50 |
-
chunk_pitch = next(chunk_iter)
|
51 |
-
f0 = chunk_pitch['filtered_f0'].fillna(0).to_numpy()
|
52 |
-
except StopIteration:
|
53 |
-
return
|
54 |
-
|
55 |
-
|
56 |
-
def flatmap(
|
57 |
-
pool: multiprocessing.Pool,
|
58 |
-
func: Callable,
|
59 |
-
iterable: Iterable,
|
60 |
-
queue_size: int,
|
61 |
-
chunksize=None,
|
62 |
-
):
|
63 |
-
queue = multiprocessing.Manager().Queue(maxsize=queue_size)
|
64 |
-
pool.map_async(
|
65 |
-
functools.partial(flat_mappper, func),
|
66 |
-
zip(iterable, repeat(queue)),
|
67 |
-
chunksize,
|
68 |
-
lambda _: queue.put(None),
|
69 |
-
lambda *e: print(e),
|
70 |
-
)
|
71 |
-
|
72 |
-
item = queue.get()
|
73 |
-
while item is not None:
|
74 |
-
# print(item)
|
75 |
-
yield item
|
76 |
-
item = queue.get()
|
77 |
-
|
78 |
-
def flat_mappper(func, arg):
|
79 |
-
data, queue = arg
|
80 |
-
for item in func(data):
|
81 |
-
queue.put(item)
|
82 |
-
|
83 |
-
def batch(iterator: Iterable, batch_size: int):
|
84 |
-
batch = []
|
85 |
-
for elm in iterator:
|
86 |
-
batch.append(elm)
|
87 |
-
if len(batch) == batch_size:
|
88 |
-
yield batch
|
89 |
-
batch = []
|
90 |
-
if len(batch):
|
91 |
-
yield batch
|
92 |
-
|
93 |
-
def preprocess_batch(
|
94 |
-
preprocessed_array,
|
95 |
-
sr_audio: int,
|
96 |
-
sr_pitch: int,
|
97 |
-
):
|
98 |
-
# pdb.set_trace()
|
99 |
-
dtype = np.float32
|
100 |
-
data_examples = [AudioExample() for _ in range(len(preprocessed_array))]
|
101 |
-
for ae, data in zip(data_examples, preprocessed_array):
|
102 |
-
# pdb.set_trace()
|
103 |
-
audio_data, csv_data, row, start_time = data
|
104 |
-
buffer_audio = ae.ae.buffers['audio']
|
105 |
-
buffer_audio.data = audio_data.astype(dtype).tobytes()
|
106 |
-
buffer_audio.shape.extend(audio_data.shape)
|
107 |
-
buffer_audio.precision = DTYPE_TO_PRECISION[dtype]
|
108 |
-
buffer_audio.sampling_rate = sr_audio
|
109 |
-
buffer_audio.data_path = row['audio_path']
|
110 |
-
buffer_audio.start_time = start_time
|
111 |
-
|
112 |
-
buffer_csv = ae.ae.buffers['pitch']
|
113 |
-
buffer_csv.data = csv_data.astype(dtype).tobytes()
|
114 |
-
buffer_csv.shape.extend(csv_data.shape)
|
115 |
-
buffer_csv.precision = DTYPE_TO_PRECISION[dtype]
|
116 |
-
buffer_csv.sampling_rate = sr_pitch
|
117 |
-
buffer_csv.data_path = row['pitch_path']
|
118 |
-
buffer_csv.start_time = start_time
|
119 |
-
|
120 |
-
ae.ae.global_conditions.tonic = row['tonic']
|
121 |
-
ae.ae.global_conditions.raga = row['raga']
|
122 |
-
ae.ae.global_conditions.singer = row['singer']
|
123 |
-
|
124 |
-
return data_examples
|
125 |
-
|
126 |
-
|
127 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/process_encodec.py
DELETED
@@ -1,22 +0,0 @@
|
|
1 |
-
import gin
|
2 |
-
from sklearn.preprocessing import QuantileTransformer
|
3 |
-
from transformers import EncodecModel, AutoProcessor
|
4 |
-
import librosa as li
|
5 |
-
|
6 |
-
import pdb
|
7 |
-
|
8 |
-
@gin.configurable
|
9 |
-
def read_tokens(
|
10 |
-
inputs,
|
11 |
-
encodec_model: EncodecModel,
|
12 |
-
encodec_processor: AutoProcessor,
|
13 |
-
target_bandwidth: int = 3
|
14 |
-
):
|
15 |
-
# pdb.set_trace()
|
16 |
-
audio = inputs['audio']['data']
|
17 |
-
audio = li.resample(y=audio, orig_sr= inputs['audio']['sampling_rate'], target_sr=encodec_processor.sampling_rate)
|
18 |
-
|
19 |
-
encodec_inputs = encodec_processor(raw_audio=audio, sampling_rate=encodec_processor.sampling_rate, return_tensors='pt')
|
20 |
-
encodec_tokens = encodec_model.encode(encodec_inputs['input_values'], bandwidth=target_bandwidth).audio_codes
|
21 |
-
|
22 |
-
return encodec_tokens.detach().cpu().numpy()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
src/utils.py
DELETED
@@ -1,65 +0,0 @@
|
|
1 |
-
from pathlib import Path
|
2 |
-
import os
|
3 |
-
import random
|
4 |
-
import torch
|
5 |
-
import numpy as np
|
6 |
-
import gin
|
7 |
-
|
8 |
-
def search_for_run(run_path, mode="last"):
|
9 |
-
if run_path is None: return None
|
10 |
-
if ".ckpt" in run_path: return run_path
|
11 |
-
ckpts = map(str, Path(run_path).rglob("*.ckpt"))
|
12 |
-
ckpts = filter(lambda e: mode in os.path.basename(str(e)), ckpts)
|
13 |
-
ckpts = sorted(ckpts)
|
14 |
-
if len(ckpts):
|
15 |
-
if len(ckpts) > 1 and 'last.ckpt' in ckpts:
|
16 |
-
return ckpts[-2] # last.ckpt is always at the end, so we take the second last
|
17 |
-
else:
|
18 |
-
return ckpts[-1]
|
19 |
-
else: return None
|
20 |
-
|
21 |
-
def set_seed(seed: int):
|
22 |
-
"""Set seed"""
|
23 |
-
random.seed(seed)
|
24 |
-
np.random.seed(seed)
|
25 |
-
torch.manual_seed(seed)
|
26 |
-
if torch.cuda.is_available():
|
27 |
-
torch.cuda.manual_seed(seed)
|
28 |
-
torch.cuda.manual_seed_all(seed)
|
29 |
-
torch.backends.cudnn.deterministic = True
|
30 |
-
torch.backends.cudnn.benchmark = False
|
31 |
-
os.environ["PYTHONHASHSEED"] = str(seed)
|
32 |
-
|
33 |
-
@gin.configurable
|
34 |
-
def build_warmed_exponential_lr_scheduler(
|
35 |
-
optim: torch.optim.Optimizer, start_factor: float, peak_iteration: int,
|
36 |
-
decay_factor: float=None, cycle_length: int=None, eta_min: float=None, eta_max: float=None) -> torch.optim.lr_scheduler._LRScheduler:
|
37 |
-
linear = torch.optim.lr_scheduler.LinearLR(
|
38 |
-
optim,
|
39 |
-
start_factor=start_factor,
|
40 |
-
end_factor=1.,
|
41 |
-
total_iters=peak_iteration,
|
42 |
-
)
|
43 |
-
if decay_factor:
|
44 |
-
exp = torch.optim.lr_scheduler.ExponentialLR(
|
45 |
-
optim,
|
46 |
-
gamma=decay_factor,
|
47 |
-
)
|
48 |
-
return torch.optim.lr_scheduler.SequentialLR(optim, [linear, exp],
|
49 |
-
milestones=[peak_iteration])
|
50 |
-
if cycle_length:
|
51 |
-
cosine = torch.optim.lr_scheduler.CosineAnnealingLR(
|
52 |
-
optim,
|
53 |
-
T_max=cycle_length,
|
54 |
-
eta_min = eta_min * eta_max
|
55 |
-
)
|
56 |
-
return torch.optim.lr_scheduler.SequentialLR(optim, [linear, cosine],
|
57 |
-
milestones=[peak_iteration])
|
58 |
-
|
59 |
-
def prob_mask_like(shape, prob, device):
|
60 |
-
if prob == 1:
|
61 |
-
return torch.ones(shape, device = device, dtype = torch.bool)
|
62 |
-
elif prob == 0:
|
63 |
-
return torch.zeros(shape, device = device, dtype = torch.bool)
|
64 |
-
else:
|
65 |
-
return torch.zeros(shape, device = device).float().uniform_(0, 1) < prob
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|