Spaces:
Sleeping
Sleeping
Hugo Flores
commited on
Commit
·
b54865d
1
Parent(s):
275afd0
interface
Browse files- requirements.txt +0 -1
- setup.py +1 -2
- vampnet/__init__.py +1 -1
- vampnet/enchilada.py +0 -179
- vampnet/interface.py +332 -0
- vampnet/modules/base.py +28 -3
requirements.txt
CHANGED
@@ -26,5 +26,4 @@ jupyter-client==6.1.12
|
|
26 |
tensorboardX
|
27 |
gradio
|
28 |
einops
|
29 |
-
flash-attn
|
30 |
frechet_audio_distance
|
|
|
26 |
tensorboardX
|
27 |
gradio
|
28 |
einops
|
|
|
29 |
frechet_audio_distance
|
setup.py
CHANGED
@@ -20,7 +20,7 @@ setup(
|
|
20 |
description="Generative Music Modeling.",
|
21 |
long_description=long_description,
|
22 |
long_description_content_type="text/markdown",
|
23 |
-
author="Hugo Flores García",
|
24 |
author_email="[email protected]",
|
25 |
url="https://github.com/descriptinc/lyrebird-vampnet",
|
26 |
license="MIT",
|
@@ -37,7 +37,6 @@ setup(
|
|
37 |
"google-cloud-logging==2.2.0",
|
38 |
"torchmetrics>=0.7.3",
|
39 |
"einops",
|
40 |
-
"flash-attn",
|
41 |
"frechet_audio_distance"
|
42 |
],
|
43 |
)
|
|
|
20 |
description="Generative Music Modeling.",
|
21 |
long_description=long_description,
|
22 |
long_description_content_type="text/markdown",
|
23 |
+
author="Hugo Flores García, Prem Seetharaman",
|
24 |
author_email="[email protected]",
|
25 |
url="https://github.com/descriptinc/lyrebird-vampnet",
|
26 |
license="MIT",
|
|
|
37 |
"google-cloud-logging==2.2.0",
|
38 |
"torchmetrics>=0.7.3",
|
39 |
"einops",
|
|
|
40 |
"frechet_audio_distance"
|
41 |
],
|
42 |
)
|
vampnet/__init__.py
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
|
2 |
from . import modules
|
3 |
from . import scheduler
|
4 |
-
from . import
|
5 |
|
6 |
__version__ = "0.0.1"
|
|
|
1 |
|
2 |
from . import modules
|
3 |
from . import scheduler
|
4 |
+
from .interface import Interface
|
5 |
|
6 |
__version__ = "0.0.1"
|
vampnet/enchilada.py
DELETED
@@ -1,179 +0,0 @@
|
|
1 |
-
import os
|
2 |
-
from pathlib import Path
|
3 |
-
|
4 |
-
import torch
|
5 |
-
from audiotools import AudioSignal
|
6 |
-
|
7 |
-
from .modules.transformer import VampNet
|
8 |
-
from lac.model.lac import LAC
|
9 |
-
|
10 |
-
|
11 |
-
class TheWholeEnchilada:
|
12 |
-
def __init__(
|
13 |
-
self,
|
14 |
-
coarse_ckpt: str,
|
15 |
-
coarse2fine_ckpt: str,
|
16 |
-
codec_ckpt: str,
|
17 |
-
device: str = "cpu",
|
18 |
-
):
|
19 |
-
self.codec = LAC.load(Path(codec_ckpt))
|
20 |
-
self.codec.eval()
|
21 |
-
self.codec.to(device)
|
22 |
-
|
23 |
-
self.coarse = VampNet.load(location=Path(coarse_ckpt), map_location="cpu")
|
24 |
-
self.coarse.to(device)
|
25 |
-
self.coarse.eval()
|
26 |
-
|
27 |
-
self.coarse2fine = VampNet.load(
|
28 |
-
location=Path(coarse2fine_ckpt), map_location="cpu"
|
29 |
-
)
|
30 |
-
# FIXME
|
31 |
-
print(
|
32 |
-
f"WARNING: PATCHING coarse2fine seq_len to 288, for backwards compatibility with a specific jazzpop model. it used to be {self.coarse2fine.seq_len}"
|
33 |
-
)
|
34 |
-
self.coarse2fine.seq_len = 288
|
35 |
-
|
36 |
-
self.coarse2fine.to(device)
|
37 |
-
self.coarse2fine.eval()
|
38 |
-
|
39 |
-
self.device = device
|
40 |
-
|
41 |
-
def seconds_to_tokens(self, seconds: float):
|
42 |
-
return int(seconds * self.codec.sample_rate / self.codec.hop_length)
|
43 |
-
|
44 |
-
def to(self, device):
|
45 |
-
self.device = device
|
46 |
-
self.coarse.to(device)
|
47 |
-
self.coarse2fine.to(device)
|
48 |
-
self.codec.to(device)
|
49 |
-
return self
|
50 |
-
|
51 |
-
def encode(self, signal: AudioSignal):
|
52 |
-
with torch.inference_mode():
|
53 |
-
# coarse z
|
54 |
-
cz = self.codec.encode(signal.samples, signal.sample_rate)["codes"]
|
55 |
-
|
56 |
-
return cz
|
57 |
-
|
58 |
-
def vamp(
|
59 |
-
self,
|
60 |
-
signal,
|
61 |
-
prefix_dur_s: float = 1.25,
|
62 |
-
suffix_dur_s: float = 1.25,
|
63 |
-
downsample_hint: bool = True,
|
64 |
-
downsample_factor: int = 4,
|
65 |
-
num_loops: int = 3,
|
66 |
-
**kwargs,
|
67 |
-
):
|
68 |
-
"""
|
69 |
-
Loop imputation of a signal.
|
70 |
-
"""
|
71 |
-
signal.to(self.device).resample(self.codec.sample_rate).to_mono()
|
72 |
-
|
73 |
-
z = self.encode(signal)
|
74 |
-
|
75 |
-
cz = z[:, : self.coarse.n_codebooks, :].clone()
|
76 |
-
original_cz = cz.clone()
|
77 |
-
seq_len = original_cz.shape[-1]
|
78 |
-
assert (
|
79 |
-
seq_len == self.coarse.seq_len
|
80 |
-
), f"expected seq_len {self.coarse.seq_len}, got {seq_len} for token sequence length. Is your signal the same duration as the model was trained with? "
|
81 |
-
|
82 |
-
vamp_hop_s = prefix_dur_s
|
83 |
-
vamp_hop = self.seconds_to_tokens(vamp_hop_s)
|
84 |
-
|
85 |
-
cmask = torch.ones_like(cz)
|
86 |
-
|
87 |
-
if downsample_hint:
|
88 |
-
# downsample by factor of 4
|
89 |
-
for i in range(cmask.shape[-1]):
|
90 |
-
if i % downsample_factor == 0:
|
91 |
-
cmask[:, :, i] = 0
|
92 |
-
|
93 |
-
if prefix_dur_s > 0:
|
94 |
-
prefix_len = self.seconds_to_tokens(prefix_dur_s)
|
95 |
-
cmask[:, :, :prefix_len] = 0
|
96 |
-
print(f"prefix_len: {prefix_len}")
|
97 |
-
else:
|
98 |
-
prefix_len = 0
|
99 |
-
|
100 |
-
if suffix_dur_s > 0:
|
101 |
-
suffix_len = self.seconds_to_tokens(suffix_dur_s)
|
102 |
-
cmask[:, :, -suffix_len:] = 0
|
103 |
-
print(f"suffix_len: {suffix_len}")
|
104 |
-
else:
|
105 |
-
suffix_len = 0
|
106 |
-
|
107 |
-
prefix_z = cz[:, :, :prefix_len]
|
108 |
-
|
109 |
-
coarse_vamp = [prefix_z.clone()]
|
110 |
-
for i in range(num_loops):
|
111 |
-
sampled_cz = self.coarse.sample(
|
112 |
-
codec=self.codec,
|
113 |
-
time_steps=seq_len,
|
114 |
-
mask=cmask,
|
115 |
-
start_tokens=cz,
|
116 |
-
return_signal=False,
|
117 |
-
**kwargs,
|
118 |
-
)
|
119 |
-
|
120 |
-
new_prefix = sampled_cz[:, :, prefix_len : prefix_len + vamp_hop]
|
121 |
-
coarse_vamp.append(new_prefix.clone())
|
122 |
-
|
123 |
-
# replace the prefix in cz with the new prefix
|
124 |
-
# don't worry about a copy of the prefix still being
|
125 |
-
# in the mask area, since that will be masked out
|
126 |
-
cz[:, :, :vamp_hop] = new_prefix.clone()
|
127 |
-
print("to append and to prefix")
|
128 |
-
|
129 |
-
# we're done, so add the suffix
|
130 |
-
coarse_vamp.append(sampled_cz[:, :, prefix_len + vamp_hop :])
|
131 |
-
|
132 |
-
# concatenate the vamps
|
133 |
-
coarse_vamp = torch.cat(coarse_vamp, dim=-1)
|
134 |
-
|
135 |
-
# add a layer of
|
136 |
-
fine_prefix = z[:, self.coarse.n_codebooks :, :prefix_len]
|
137 |
-
fine_suffix = z[:, self.coarse.n_codebooks :, -suffix_len:]
|
138 |
-
fine_vamp = torch.randint(
|
139 |
-
0,
|
140 |
-
self.coarse2fine.vocab_size,
|
141 |
-
(
|
142 |
-
coarse_vamp.shape[0],
|
143 |
-
self.coarse2fine.n_predict_codebooks,
|
144 |
-
coarse_vamp.shape[-1],
|
145 |
-
),
|
146 |
-
).to(self.device)
|
147 |
-
fine_vamp[:, :, :prefix_len] = fine_prefix
|
148 |
-
fine_vamp[:, :, -suffix_len:] = fine_suffix
|
149 |
-
|
150 |
-
vamp_z = torch.cat([coarse_vamp, fine_vamp], dim=1)
|
151 |
-
|
152 |
-
# now we sample from the coarse2fine model
|
153 |
-
# to get the fine details
|
154 |
-
start_pos = 0
|
155 |
-
|
156 |
-
c2f_vamp = []
|
157 |
-
while start_pos < vamp_z.shape[-1]:
|
158 |
-
end_pos = min(start_pos + self.coarse2fine.seq_len, vamp_z.shape[-1])
|
159 |
-
|
160 |
-
c2fz = vamp_z[:, :, start_pos:end_pos]
|
161 |
-
self.coarse2fine: VampNet
|
162 |
-
sampled_c2fz = self.coarse2fine.sample(
|
163 |
-
codec=self.codec,
|
164 |
-
start_tokens=c2fz,
|
165 |
-
return_signal=False,
|
166 |
-
mask=None,
|
167 |
-
)
|
168 |
-
c2f_vamp.append(sampled_c2fz)
|
169 |
-
start_pos += self.coarse2fine.seq_len
|
170 |
-
|
171 |
-
c2f_vamp = torch.cat(c2f_vamp, dim=-1)
|
172 |
-
|
173 |
-
# make it a signal
|
174 |
-
vamp_signal = self.coarse2fine.to_signal(c2f_vamp, self.codec)
|
175 |
-
|
176 |
-
return {
|
177 |
-
"full": vamp_signal,
|
178 |
-
"coarse": self.coarse.to_signal(coarse_vamp, self.codec),
|
179 |
-
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
vampnet/interface.py
ADDED
@@ -0,0 +1,332 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
from pathlib import Path
|
3 |
+
import math
|
4 |
+
|
5 |
+
import torch
|
6 |
+
from audiotools import AudioSignal
|
7 |
+
|
8 |
+
from .modules.transformer import VampNet
|
9 |
+
from lac.model.lac import LAC
|
10 |
+
|
11 |
+
|
12 |
+
class Interface:
|
13 |
+
def __init__(
|
14 |
+
self,
|
15 |
+
coarse_ckpt: str,
|
16 |
+
coarse2fine_ckpt: str,
|
17 |
+
codec_ckpt: str,
|
18 |
+
device: str = "cpu",
|
19 |
+
coarse_chunk_size_s: int = 5,
|
20 |
+
coarse2fine_chunk_size_s: int = 3,
|
21 |
+
):
|
22 |
+
self.codec = LAC.load(Path(codec_ckpt))
|
23 |
+
self.codec.eval()
|
24 |
+
self.codec.to(device)
|
25 |
+
|
26 |
+
self.coarse = VampNet.load(location=Path(coarse_ckpt), map_location="cpu")
|
27 |
+
self.coarse.to(device)
|
28 |
+
self.coarse.eval()
|
29 |
+
self.coarse.chunk_size_s = coarse_chunk_size_s
|
30 |
+
|
31 |
+
self.c2f = VampNet.load(
|
32 |
+
location=Path(coarse2fine_ckpt), map_location="cpu"
|
33 |
+
)
|
34 |
+
self.c2f.to(device)
|
35 |
+
self.c2f.eval()
|
36 |
+
self.c2f.chunk_size_s = coarse2fine_chunk_size_s
|
37 |
+
|
38 |
+
self.device = device
|
39 |
+
|
40 |
+
def s2t(self, seconds: float):
|
41 |
+
"""seconds to tokens"""
|
42 |
+
return int(seconds * self.codec.sample_rate / self.codec.hop_length)
|
43 |
+
|
44 |
+
def to(self, device):
|
45 |
+
self.device = device
|
46 |
+
self.coarse.to(device)
|
47 |
+
self.c2f.to(device)
|
48 |
+
self.codec.to(device)
|
49 |
+
return self
|
50 |
+
|
51 |
+
def to_signal(self, z: torch.Tensor):
|
52 |
+
return self.coarse.to_signal(z, self.codec)
|
53 |
+
|
54 |
+
@torch.inference_mode()
|
55 |
+
def encode(self, signal: AudioSignal):
|
56 |
+
signal = signal.clone().to(self.device).resample(self.codec.sample_rate).to_mono()
|
57 |
+
z = self.codec.encode(signal.samples, signal.sample_rate)["codes"]
|
58 |
+
return z
|
59 |
+
|
60 |
+
def coarse_to_fine(
|
61 |
+
self,
|
62 |
+
coarse_z: torch.Tensor,
|
63 |
+
**kwargs
|
64 |
+
):
|
65 |
+
length = coarse_z.shape[-1]
|
66 |
+
chunk_len = self.s2t(self.c2f.chunk_size_s)
|
67 |
+
n_chunks = math.ceil(coarse_z.shape[-1] / chunk_len)
|
68 |
+
|
69 |
+
# zero pad to chunk_len
|
70 |
+
if length % chunk_len != 0:
|
71 |
+
pad_len = chunk_len - (length % chunk_len)
|
72 |
+
coarse_z = torch.nn.functional.pad(coarse_z, (0, pad_len))
|
73 |
+
|
74 |
+
n_codebooks_to_append = self.c2f.n_codebooks - coarse_z.shape[1]
|
75 |
+
if n_codebooks_to_append > 0:
|
76 |
+
coarse_z = torch.cat([
|
77 |
+
coarse_z,
|
78 |
+
torch.zeros(coarse_z.shape[0], n_codebooks_to_append, coarse_z.shape[-1]).long().to(self.device)
|
79 |
+
], dim=1)
|
80 |
+
|
81 |
+
fine_z = []
|
82 |
+
for i in range(n_chunks):
|
83 |
+
chunk = coarse_z[:, :, i * chunk_len : (i + 1) * chunk_len]
|
84 |
+
chunk = self.c2f.sample(
|
85 |
+
codec=self.codec,
|
86 |
+
time_steps=chunk_len,
|
87 |
+
start_tokens=chunk,
|
88 |
+
return_signal=False,
|
89 |
+
)
|
90 |
+
fine_z.append(chunk)
|
91 |
+
|
92 |
+
fine_z = torch.cat(fine_z, dim=-1)
|
93 |
+
return fine_z[:, :, :length].clone()
|
94 |
+
|
95 |
+
def coarse_vamp(
|
96 |
+
self,
|
97 |
+
signal,
|
98 |
+
prefix_dur_s: float = 1.25,
|
99 |
+
suffix_dur_s: float = 1.25,
|
100 |
+
num_loops: int = 3,
|
101 |
+
mode="impute",
|
102 |
+
downsample_factor: int = None,
|
103 |
+
debug=False,
|
104 |
+
**kwargs
|
105 |
+
):
|
106 |
+
z = self.encode(signal)
|
107 |
+
|
108 |
+
assert signal.duration == self.coarse.chunk_size_s, "signal duration must match coarse chunk size for now"
|
109 |
+
|
110 |
+
# coarse z
|
111 |
+
cz = z[:, : self.coarse.n_codebooks, :].clone()
|
112 |
+
c_seq_len = cz.shape[-1]
|
113 |
+
n_prefix = self.s2t(prefix_dur_s)
|
114 |
+
n_suffix = self.s2t(suffix_dur_s)
|
115 |
+
|
116 |
+
# we'll keep the final codes sequence here
|
117 |
+
c_vamp = {
|
118 |
+
'prefix': [cz[:, :, :n_prefix].clone()],
|
119 |
+
'suffix': [cz[:, :, c_seq_len-n_suffix:].clone()]
|
120 |
+
}
|
121 |
+
|
122 |
+
_cz = cz.clone()
|
123 |
+
for _ in range(num_loops):
|
124 |
+
# add noise
|
125 |
+
cz_masked, cz_mask = self.coarse.add_noise(
|
126 |
+
_cz, r=0.0,
|
127 |
+
n_prefix=n_prefix,
|
128 |
+
n_suffix=n_suffix,
|
129 |
+
downsample_factor=downsample_factor
|
130 |
+
)
|
131 |
+
if debug:
|
132 |
+
print("tokens to infer")
|
133 |
+
self.to_signal(cz_masked).cpu().widget()
|
134 |
+
|
135 |
+
# sample!
|
136 |
+
cz_sampled = self.coarse.sample(
|
137 |
+
codec=self.codec,
|
138 |
+
time_steps=self.s2t(self.coarse.chunk_size_s),
|
139 |
+
start_tokens=_cz,
|
140 |
+
mask=cz_mask,
|
141 |
+
return_signal=False,
|
142 |
+
**kwargs
|
143 |
+
)
|
144 |
+
|
145 |
+
if debug:
|
146 |
+
print("tokens sampled")
|
147 |
+
self.to_signal(cz_sampled).cpu().widget()
|
148 |
+
|
149 |
+
cz_imputed = cz_sampled[:, :, n_prefix:c_seq_len-n_suffix].clone()
|
150 |
+
|
151 |
+
if mode == "impute":
|
152 |
+
# split the imputed codes into two halves
|
153 |
+
cz_imputed_a = cz_imputed[:, :, : cz_imputed.shape[-1] // 2].clone()
|
154 |
+
cz_imputed_b = cz_imputed[:, :, cz_imputed.shape[-1] // 2 :].clone()
|
155 |
+
elif mode == "continue":
|
156 |
+
cz_imputed_a = cz_imputed[:, :, : cz_imputed.shape[-1]].clone()
|
157 |
+
cz_imputed_b = _cz[:, :, :0].clone() # empty
|
158 |
+
elif mode == "reverse-continue":
|
159 |
+
cz_imputed_a = _cz[:, :, :0].clone() # empty
|
160 |
+
cz_imputed_b = cz_imputed[:, :, : cz_imputed.shape[-1]].clone()
|
161 |
+
else:
|
162 |
+
raise ValueError(f"mode {mode} not supported")
|
163 |
+
|
164 |
+
if debug:
|
165 |
+
# add to our c_vamp
|
166 |
+
if cz_imputed_a.shape[-1] > 0:
|
167 |
+
print("new_prefix added")
|
168 |
+
self.to_signal(cz_imputed_a).cpu().widget()
|
169 |
+
if cz_imputed_b.shape[-1] > 0:
|
170 |
+
print("new_suffix added")
|
171 |
+
self.to_signal(cz_imputed_b).cpu().widget()
|
172 |
+
|
173 |
+
c_vamp['prefix'].append(cz_imputed_a.clone())
|
174 |
+
c_vamp['suffix'].insert(0, cz_imputed_b.clone())
|
175 |
+
|
176 |
+
n_to_insert = c_seq_len - (cz_imputed_a.shape[-1] + cz_imputed_b.shape[-1])
|
177 |
+
to_insert = torch.zeros(cz_imputed_a.shape[0], cz_imputed_a.shape[1], n_to_insert).long().to(self.device)
|
178 |
+
_cz = torch.cat([cz_imputed_a, to_insert, cz_imputed_b], dim=-1)
|
179 |
+
|
180 |
+
if debug:
|
181 |
+
print("tokens to infer next round (area to insert in the middle)")
|
182 |
+
self.to_signal(_cz).cpu().widget()
|
183 |
+
|
184 |
+
|
185 |
+
|
186 |
+
|
187 |
+
prefix_codes = torch.cat(c_vamp['prefix'], dim=-1)
|
188 |
+
suffix_codes = torch.cat(c_vamp['suffix'], dim=-1)
|
189 |
+
c_vamp = torch.cat([prefix_codes, suffix_codes], dim=-1)
|
190 |
+
return c_vamp
|
191 |
+
|
192 |
+
|
193 |
+
def coarse_vamp_v2(
|
194 |
+
self,
|
195 |
+
signal,
|
196 |
+
prefix_dur_s: float = 1.25,
|
197 |
+
suffix_dur_s: float = 1.25,
|
198 |
+
num_loops: int = 3,
|
199 |
+
downsample_factor: int = None,
|
200 |
+
debug=False,
|
201 |
+
**kwargs
|
202 |
+
):
|
203 |
+
z = self.encode(signal)
|
204 |
+
|
205 |
+
assert signal.duration == self.coarse.chunk_size_s, "signal duration must match coarse chunk size for now"
|
206 |
+
|
207 |
+
# coarse z
|
208 |
+
cz = z[:, : self.coarse.n_codebooks, :].clone()
|
209 |
+
c_seq_len = cz.shape[-1]
|
210 |
+
n_prefix = self.s2t(prefix_dur_s)
|
211 |
+
n_suffix = self.s2t(suffix_dur_s)
|
212 |
+
|
213 |
+
assert n_prefix + n_suffix < c_seq_len, "prefix and suffix must be smaller than the chunk size"
|
214 |
+
|
215 |
+
# we'll keep the final codes sequence here
|
216 |
+
c_vamp = {
|
217 |
+
'prefix': [cz[:, :, :n_prefix].clone()],
|
218 |
+
'suffix': [cz[:, :, c_seq_len-n_suffix:].clone()]
|
219 |
+
}
|
220 |
+
|
221 |
+
_cz = cz.clone()
|
222 |
+
cz_mask = None
|
223 |
+
for _ in range(num_loops):
|
224 |
+
# add noise
|
225 |
+
cz_masked, cz_mask = self.coarse.add_noise(
|
226 |
+
_cz, r=0.0,
|
227 |
+
n_prefix=n_prefix,
|
228 |
+
n_suffix=n_suffix,
|
229 |
+
downsample_factor=downsample_factor,
|
230 |
+
mask=cz_mask
|
231 |
+
)
|
232 |
+
if debug:
|
233 |
+
print("tokens to infer")
|
234 |
+
self.to_signal(cz_masked).cpu().widget()
|
235 |
+
|
236 |
+
# sample!
|
237 |
+
if debug:
|
238 |
+
print(f"mask: {cz_mask[:,0,:]}")
|
239 |
+
print(f"z: {_cz[:,0,:]}")
|
240 |
+
cz_sampled = self.coarse.sample(
|
241 |
+
codec=self.codec,
|
242 |
+
time_steps=self.s2t(self.coarse.chunk_size_s),
|
243 |
+
start_tokens=_cz,
|
244 |
+
mask=cz_mask,
|
245 |
+
return_signal=False,
|
246 |
+
**kwargs
|
247 |
+
)
|
248 |
+
|
249 |
+
if debug:
|
250 |
+
print("tokens sampled")
|
251 |
+
self.to_signal(cz_sampled).cpu().widget()
|
252 |
+
|
253 |
+
# the z that was generated
|
254 |
+
cz_generated = cz_sampled[:, :, n_prefix:c_seq_len-n_suffix].clone()
|
255 |
+
n_generated = cz_generated.shape[-1]
|
256 |
+
|
257 |
+
# create the new prefix and suffix
|
258 |
+
# we'll make sure that the number of prefix and suffix
|
259 |
+
# tokens is the same as the original
|
260 |
+
# but we do want to advance the sequence as much as we can
|
261 |
+
if n_prefix > 0 and n_suffix > 0:
|
262 |
+
# we have both prefix and suffix, so we'll split the generated
|
263 |
+
# codes in two halves
|
264 |
+
prefix_start_idx = n_generated // 2
|
265 |
+
prefix_stop_idx = prefix_start_idx + n_prefix
|
266 |
+
assert prefix_start_idx >= 0, "internal error"
|
267 |
+
|
268 |
+
suffix_start_idx = n_prefix + n_generated // 2
|
269 |
+
suffix_stop_idx = suffix_start_idx + n_suffix
|
270 |
+
assert suffix_stop_idx <= cz_sampled.shape[-1], "internal error"
|
271 |
+
|
272 |
+
cz_new_prefix = cz_sampled[:, :, prefix_start_idx:prefix_stop_idx].clone()
|
273 |
+
cz_new_suffix = cz_sampled[:, :, suffix_start_idx:suffix_stop_idx].clone()
|
274 |
+
|
275 |
+
c_vamp['prefix'].append(cz_generated[:,:,:n_generated//2])
|
276 |
+
c_vamp['suffix'].insert(0, cz_generated[:,:,n_generated//2:])
|
277 |
+
|
278 |
+
elif n_prefix > 0:
|
279 |
+
# we only have a prefix
|
280 |
+
prefix_start_idx = n_generated
|
281 |
+
prefix_stop_idx = prefix_start_idx + n_prefix
|
282 |
+
|
283 |
+
cz_new_prefix = cz_sampled[:, :, prefix_start_idx:prefix_stop_idx].clone()
|
284 |
+
cz_new_suffix = _cz[:, :, :0].clone()
|
285 |
+
|
286 |
+
|
287 |
+
c_vamp['prefix'].append(cz_generated)
|
288 |
+
|
289 |
+
elif n_suffix > 0:
|
290 |
+
# we only have a suffix, so everything starting at 0 is generated
|
291 |
+
suffix_stop_idx = max(n_generated, n_suffix)
|
292 |
+
suffix_start_idx = suffix_stop_idx - n_suffix
|
293 |
+
|
294 |
+
cz_new_prefix = _cz[:, :, :0].clone()
|
295 |
+
cz_new_suffix = cz_sampled[:, :, suffix_start_idx:suffix_stop_idx].clone()
|
296 |
+
|
297 |
+
c_vamp['suffix'].insert(0, cz_generated)
|
298 |
+
|
299 |
+
|
300 |
+
n_to_insert = c_seq_len - (cz_new_prefix.shape[-1] + cz_new_suffix.shape[-1])
|
301 |
+
to_insert = torch.zeros(cz_new_prefix.shape[0], cz_new_prefix.shape[1], n_to_insert).long().to(self.device)
|
302 |
+
_cz = torch.cat([cz_new_prefix, to_insert, cz_new_suffix], dim=-1)
|
303 |
+
|
304 |
+
to_insert_mask = torch.zeros_like(_cz).long().to(self.device)
|
305 |
+
to_insert_mask[:, :, cz_new_prefix.shape[-1]:cz_new_prefix.shape[-1]+n_to_insert] = 1
|
306 |
+
cz_mask = (cz_mask + to_insert_mask).bool().long()
|
307 |
+
|
308 |
+
|
309 |
+
if debug:
|
310 |
+
print("tokens to infer next round (area to insert in the middle)")
|
311 |
+
self.to_signal(_cz).cpu().widget()
|
312 |
+
|
313 |
+
|
314 |
+
prefix_codes = torch.cat(c_vamp['prefix'], dim=-1)
|
315 |
+
suffix_codes = torch.cat(c_vamp['suffix'], dim=-1)
|
316 |
+
c_vamp = torch.cat([prefix_codes, suffix_codes], dim=-1)
|
317 |
+
return c_vamp
|
318 |
+
|
319 |
+
|
320 |
+
|
321 |
+
|
322 |
+
|
323 |
+
|
324 |
+
|
325 |
+
|
326 |
+
|
327 |
+
|
328 |
+
|
329 |
+
|
330 |
+
|
331 |
+
|
332 |
+
|
vampnet/modules/base.py
CHANGED
@@ -24,6 +24,9 @@ def gumbel_sample(t, temperature=1.0, dim=-1):
|
|
24 |
return ((t / max(temperature, 1e-10)) + gumbel_noise(t)).argmax(dim=dim)
|
25 |
|
26 |
|
|
|
|
|
|
|
27 |
class VampBase(at.ml.BaseModel):
|
28 |
def forward(self, x: torch.Tensor, r: torch.Tensor):
|
29 |
raise NotImplementedError
|
@@ -36,20 +39,40 @@ class VampBase(at.ml.BaseModel):
|
|
36 |
mask: Optional[torch.Tensor] = None,
|
37 |
n_prefix: Optional[torch.Tensor] = None,
|
38 |
n_suffix: Optional[torch.Tensor] = None,
|
|
|
39 |
) -> Tuple[torch.Tensor, torch.Tensor]:
|
40 |
assert x.ndim == 3, "x must be (batch, n_codebooks, seq)"
|
41 |
|
42 |
if mask is None:
|
|
|
|
|
43 |
r = self.gamma(r)[:, None, None]
|
44 |
probs = torch.ones_like(x) * r
|
45 |
|
46 |
# if we have a prefix or suffix, set their mask prob to 0
|
47 |
if n_prefix is not None:
|
|
|
|
|
48 |
for i, n in enumerate(n_prefix):
|
49 |
-
|
|
|
50 |
if n_suffix is not None:
|
|
|
|
|
51 |
for i, n in enumerate(n_suffix):
|
52 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
53 |
|
54 |
mask = torch.bernoulli(probs)
|
55 |
mask = mask.round().long()
|
@@ -347,7 +370,9 @@ class VampBase(at.ml.BaseModel):
|
|
347 |
if num_to_keep > 0:
|
348 |
probs = logits.softmax(dim=-1)
|
349 |
|
350 |
-
|
|
|
|
|
351 |
|
352 |
probs = rearrange(
|
353 |
probs, "b (t c) p -> b c t p", c=n_infer_codebooks
|
|
|
24 |
return ((t / max(temperature, 1e-10)) + gumbel_noise(t)).argmax(dim=dim)
|
25 |
|
26 |
|
27 |
+
def scalar_to_batch_tensor(x, batch_size):
|
28 |
+
return torch.tensor(x).repeat(batch_size)
|
29 |
+
|
30 |
class VampBase(at.ml.BaseModel):
|
31 |
def forward(self, x: torch.Tensor, r: torch.Tensor):
|
32 |
raise NotImplementedError
|
|
|
39 |
mask: Optional[torch.Tensor] = None,
|
40 |
n_prefix: Optional[torch.Tensor] = None,
|
41 |
n_suffix: Optional[torch.Tensor] = None,
|
42 |
+
downsample_factor: Optional[int] = None,
|
43 |
) -> Tuple[torch.Tensor, torch.Tensor]:
|
44 |
assert x.ndim == 3, "x must be (batch, n_codebooks, seq)"
|
45 |
|
46 |
if mask is None:
|
47 |
+
if not isinstance(r, torch.Tensor):
|
48 |
+
r = scalar_to_batch_tensor(r, x.shape[0]).to(x.device)
|
49 |
r = self.gamma(r)[:, None, None]
|
50 |
probs = torch.ones_like(x) * r
|
51 |
|
52 |
# if we have a prefix or suffix, set their mask prob to 0
|
53 |
if n_prefix is not None:
|
54 |
+
if not isinstance(n_prefix, torch.Tensor):
|
55 |
+
n_prefix = scalar_to_batch_tensor(n_prefix, x.shape[0]).to(x.device)
|
56 |
for i, n in enumerate(n_prefix):
|
57 |
+
if n > 0:
|
58 |
+
probs[i, :, :n] = 0.0
|
59 |
if n_suffix is not None:
|
60 |
+
if not isinstance(n_suffix, torch.Tensor):
|
61 |
+
n_suffix = scalar_to_batch_tensor(n_suffix, x.shape[0]).to(x.device)
|
62 |
for i, n in enumerate(n_suffix):
|
63 |
+
if n > 0:
|
64 |
+
probs[i, :, -n:] = 0.0
|
65 |
+
|
66 |
+
# if we have a downsample factor, set the mask prob to 0
|
67 |
+
if downsample_factor is not None:
|
68 |
+
if not isinstance(downsample_factor, torch.Tensor):
|
69 |
+
downsample_factor = scalar_to_batch_tensor(downsample_factor, x.shape[0])
|
70 |
+
for i, factor in enumerate(downsample_factor):
|
71 |
+
if factor == 0:
|
72 |
+
continue
|
73 |
+
for j in range(probs.shape[-1]):
|
74 |
+
if j % factor == 0:
|
75 |
+
probs[i, :, j] = 0.0
|
76 |
|
77 |
mask = torch.bernoulli(probs)
|
78 |
mask = mask.round().long()
|
|
|
370 |
if num_to_keep > 0:
|
371 |
probs = logits.softmax(dim=-1)
|
372 |
|
373 |
+
# do mod self.vocab_size to make sure we don't sample from the mask token
|
374 |
+
# in case the mask token was in the og z
|
375 |
+
keep_probs = F.one_hot(z%self.vocab_size, self.vocab_size)[:, :, :]
|
376 |
|
377 |
probs = rearrange(
|
378 |
probs, "b (t c) p -> b c t p", c=n_infer_codebooks
|