Spaces:
Sleeping
Sleeping
Hugo Flores Garcia
commited on
Commit
·
d98455c
1
Parent(s):
c068a29
cleanup
Browse files- scripts/exp/train.py +0 -6
- scripts/utils/vamp_folder.py +51 -81
- vampnet/interface.py +1 -1
- vampnet/modules/transformer.py +33 -248
scripts/exp/train.py
CHANGED
@@ -342,8 +342,6 @@ def train(
|
|
342 |
dtype = torch.bfloat16 if accel.amp else None
|
343 |
with accel.autocast(dtype=dtype):
|
344 |
z_hat = model(z_mask_latent, r)
|
345 |
-
# for mask mode
|
346 |
-
z_hat = vn.add_truth_to_logits(z, z_hat, mask)
|
347 |
|
348 |
target = codebook_flatten(
|
349 |
z[:, vn.n_conditioning_codebooks :, :],
|
@@ -414,8 +412,6 @@ def train(
|
|
414 |
z_mask_latent = vn.embedding.from_codes(z_mask, codec)
|
415 |
|
416 |
z_hat = model(z_mask_latent, r)
|
417 |
-
# for mask mode
|
418 |
-
z_hat = vn.add_truth_to_logits(z, z_hat, mask)
|
419 |
|
420 |
target = codebook_flatten(
|
421 |
z[:, vn.n_conditioning_codebooks :, :],
|
@@ -573,8 +569,6 @@ def train(
|
|
573 |
z_mask_latent = vn.embedding.from_codes(z_mask, codec)
|
574 |
|
575 |
z_hat = model(z_mask_latent, r)
|
576 |
-
# for mask mode
|
577 |
-
z_hat = vn.add_truth_to_logits(z, z_hat, mask)
|
578 |
|
579 |
z_pred = torch.softmax(z_hat, dim=1).argmax(dim=1)
|
580 |
z_pred = codebook_unflatten(z_pred, n_c=vn.n_predict_codebooks)
|
|
|
342 |
dtype = torch.bfloat16 if accel.amp else None
|
343 |
with accel.autocast(dtype=dtype):
|
344 |
z_hat = model(z_mask_latent, r)
|
|
|
|
|
345 |
|
346 |
target = codebook_flatten(
|
347 |
z[:, vn.n_conditioning_codebooks :, :],
|
|
|
412 |
z_mask_latent = vn.embedding.from_codes(z_mask, codec)
|
413 |
|
414 |
z_hat = model(z_mask_latent, r)
|
|
|
|
|
415 |
|
416 |
target = codebook_flatten(
|
417 |
z[:, vn.n_conditioning_codebooks :, :],
|
|
|
569 |
z_mask_latent = vn.embedding.from_codes(z_mask, codec)
|
570 |
|
571 |
z_hat = model(z_mask_latent, r)
|
|
|
|
|
572 |
|
573 |
z_pred = torch.softmax(z_hat, dim=1).argmax(dim=1)
|
574 |
z_pred = codebook_unflatten(z_pred, n_c=vn.n_predict_codebooks)
|
scripts/utils/vamp_folder.py
CHANGED
@@ -9,10 +9,13 @@ from tqdm import tqdm
|
|
9 |
import torch
|
10 |
|
11 |
from vampnet.interface import Interface
|
|
|
12 |
import audiotools as at
|
13 |
|
14 |
Interface: Interface = argbind.bind(Interface)
|
15 |
|
|
|
|
|
16 |
def calculate_bitrate(
|
17 |
interface, num_codebooks,
|
18 |
downsample_factor
|
@@ -38,29 +41,19 @@ def coarse2fine(sig, interface):
|
|
38 |
z = interface.coarse_to_fine(z)
|
39 |
return interface.to_signal(z)
|
40 |
|
41 |
-
def coarse2fine_argmax(sig, interface):
|
42 |
-
z = interface.encode(sig)
|
43 |
-
z = z[:, :interface.c2f.n_conditioning_codebooks, :]
|
44 |
-
|
45 |
-
z = interface.coarse_to_fine(z,
|
46 |
-
sample="argmax", sampling_steps=1,
|
47 |
-
temperature=1.0
|
48 |
-
)
|
49 |
-
return interface.to_signal(z)
|
50 |
-
|
51 |
class CoarseCond:
|
52 |
|
53 |
-
def __init__(self,
|
54 |
-
self.
|
55 |
self.downsample_factor = downsample_factor
|
56 |
|
57 |
def __call__(self, sig, interface):
|
58 |
-
|
59 |
-
|
60 |
-
|
61 |
-
|
62 |
-
)
|
63 |
|
|
|
64 |
zv = interface.coarse_to_fine(zv)
|
65 |
return interface.to_signal(zv)
|
66 |
|
@@ -97,24 +90,24 @@ def opus(sig, interface, bitrate=128):
|
|
97 |
|
98 |
def mask_ratio_1_step(ratio=1.0):
|
99 |
def wrapper(sig, interface):
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
zv = interface.coarse_vamp(
|
104 |
-
|
105 |
-
|
106 |
sampling_steps=1,
|
107 |
-
intensity=intensity
|
108 |
)
|
109 |
|
110 |
return interface.to_signal(zv)
|
111 |
return wrapper
|
112 |
|
113 |
def num_sampling_steps(num_steps=1):
|
114 |
-
def wrapper(sig, interface):
|
|
|
|
|
115 |
zv = interface.coarse_vamp(
|
116 |
-
|
117 |
-
|
118 |
sampling_steps=num_steps,
|
119 |
)
|
120 |
|
@@ -130,9 +123,9 @@ def beat_mask(ctx_time):
|
|
130 |
after_beat_s=ctx_time,
|
131 |
invert=True
|
132 |
)
|
|
|
133 |
zv = interface.coarse_vamp(
|
134 |
-
|
135 |
-
ext_mask=beat_mask,
|
136 |
)
|
137 |
|
138 |
zv = interface.coarse_to_fine(zv)
|
@@ -140,17 +133,28 @@ def beat_mask(ctx_time):
|
|
140 |
return wrapper
|
141 |
|
142 |
def inpaint(ctx_time):
|
143 |
-
def wrapper(sig, interface):
|
144 |
-
|
145 |
-
|
146 |
-
prefix_dur_s=ctx_time,
|
147 |
-
suffix_dur_s=ctx_time,
|
148 |
-
)
|
149 |
|
|
|
150 |
zv = interface.coarse_to_fine(zv)
|
|
|
151 |
return interface.to_signal(zv)
|
152 |
return wrapper
|
153 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
154 |
EXP_REGISTRY = {}
|
155 |
|
156 |
EXP_REGISTRY["gen-compression"] = {
|
@@ -158,62 +162,27 @@ EXP_REGISTRY["gen-compression"] = {
|
|
158 |
"reconstructed": reconstructed,
|
159 |
"coarse2fine": coarse2fine,
|
160 |
**{
|
161 |
-
f"{n}_codebooks_downsampled_{x}x": CoarseCond(
|
162 |
for (n, x) in (
|
163 |
-
(
|
164 |
-
(2, 2), # 2 codebooks, downsampled 2x
|
165 |
-
(1, None), # 1 codebook, no downsampling
|
166 |
(4, 4), # 4 codebooks, downsampled 4x
|
167 |
-
(1, 2), # 1 codebook, downsampled 2x,
|
168 |
-
(4, 6), # 4 codebooks, downsampled 6x
|
169 |
-
(4, 8), # 4 codebooks, downsampled 8x
|
170 |
(4, 16), # 4 codebooks, downsampled 16x
|
171 |
(4, 32), # 4 codebooks, downsampled 16x
|
172 |
)
|
173 |
},
|
|
|
|
|
|
|
|
|
174 |
|
175 |
}
|
176 |
|
177 |
-
EXP_REGISTRY["opus-jazzpop"] = {
|
178 |
-
f"opus_{bitrate}": lambda sig, interface: opus(sig, interface, bitrate=bitrate)
|
179 |
-
for bitrate in [5620, 1875, 1250, 625]
|
180 |
-
}
|
181 |
-
|
182 |
-
EXP_REGISTRY["opus-spotdl"] = {
|
183 |
-
f"opus_{bitrate}": lambda sig, interface: opus(sig, interface, bitrate=bitrate)
|
184 |
-
for bitrate in [8036, 2296, 1148, 574]
|
185 |
-
}
|
186 |
-
|
187 |
-
EXP_REGISTRY["opus-baseline"] = {
|
188 |
-
f"opus_{bitrate}": lambda sig, interface: opus(sig, interface, bitrate=bitrate)
|
189 |
-
for bitrate in [8000, 12000, 16000]
|
190 |
-
}
|
191 |
-
|
192 |
-
EXP_REGISTRY["c2f"] = {
|
193 |
-
"baseline": baseline,
|
194 |
-
"reconstructed": reconstructed,
|
195 |
-
"coarse2fine": coarse2fine,
|
196 |
-
"coarse2fine_argmax": coarse2fine_argmax,
|
197 |
-
}
|
198 |
-
|
199 |
-
EXP_REGISTRY["token-noise"] = {
|
200 |
-
f"token_noise_{r}": token_noise(r) for r in [0.25, 0.5, 0.75, 1.0]
|
201 |
-
}
|
202 |
-
|
203 |
-
EXP_REGISTRY["mask-ratio"] = {
|
204 |
-
"codec": reconstructed,
|
205 |
-
**{f"mask_ratio_{r}": mask_ratio_1_step(r) for r in [0.25, 0.5, 0.75, 0.9]}
|
206 |
-
}
|
207 |
|
208 |
EXP_REGISTRY["sampling-steps"] = {
|
209 |
-
"codec": reconstructed,
|
210 |
-
**{f"steps_{n}": num_sampling_steps(n) for n in [1, 4, 12,
|
211 |
}
|
212 |
|
213 |
-
EXP_REGISTRY["baseline"] = {
|
214 |
-
"baseline": baseline,
|
215 |
-
"codec": reconstructed,
|
216 |
-
}
|
217 |
|
218 |
EXP_REGISTRY["musical-sampling"] = {
|
219 |
"baseline": baseline,
|
@@ -226,12 +195,13 @@ EXP_REGISTRY["musical-sampling"] = {
|
|
226 |
@argbind.bind(without_prefix=True)
|
227 |
def main(
|
228 |
sources=[
|
229 |
-
"/
|
230 |
],
|
231 |
output_dir: str = "./samples",
|
232 |
-
max_excerpts: int =
|
233 |
-
exp_type: str = "
|
234 |
seed: int = 0,
|
|
|
235 |
):
|
236 |
at.util.seed(seed)
|
237 |
interface = Interface()
|
@@ -241,7 +211,7 @@ def main(
|
|
241 |
|
242 |
from audiotools.data.datasets import AudioLoader, AudioDataset
|
243 |
|
244 |
-
loader = AudioLoader(sources=sources, shuffle_state=seed)
|
245 |
dataset = AudioDataset(loader,
|
246 |
sample_rate=interface.codec.sample_rate,
|
247 |
duration=interface.coarse.chunk_size_s,
|
|
|
9 |
import torch
|
10 |
|
11 |
from vampnet.interface import Interface
|
12 |
+
from vampnet import mask as pmask
|
13 |
import audiotools as at
|
14 |
|
15 |
Interface: Interface = argbind.bind(Interface)
|
16 |
|
17 |
+
|
18 |
+
|
19 |
def calculate_bitrate(
|
20 |
interface, num_codebooks,
|
21 |
downsample_factor
|
|
|
41 |
z = interface.coarse_to_fine(z)
|
42 |
return interface.to_signal(z)
|
43 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
44 |
class CoarseCond:
|
45 |
|
46 |
+
def __init__(self, num_conditioning_codebooks, downsample_factor):
|
47 |
+
self.num_conditioning_codebooks = num_conditioning_codebooks
|
48 |
self.downsample_factor = downsample_factor
|
49 |
|
50 |
def __call__(self, sig, interface):
|
51 |
+
z = interface.encode(sig)
|
52 |
+
mask = pmask.full_mask(z)
|
53 |
+
mask = pmask.codebook_unmask(mask, self.num_conditioning_codebooks)
|
54 |
+
mask = pmask.periodic_mask(mask, self.downsample_factor)
|
|
|
55 |
|
56 |
+
zv = interface.coarse_vamp(z, mask)
|
57 |
zv = interface.coarse_to_fine(zv)
|
58 |
return interface.to_signal(zv)
|
59 |
|
|
|
90 |
|
91 |
def mask_ratio_1_step(ratio=1.0):
|
92 |
def wrapper(sig, interface):
|
93 |
+
z = interface.encode(sig)
|
94 |
+
mask = pmask.linear_random(z, ratio)
|
|
|
95 |
zv = interface.coarse_vamp(
|
96 |
+
z,
|
97 |
+
mask,
|
98 |
sampling_steps=1,
|
|
|
99 |
)
|
100 |
|
101 |
return interface.to_signal(zv)
|
102 |
return wrapper
|
103 |
|
104 |
def num_sampling_steps(num_steps=1):
|
105 |
+
def wrapper(sig, interface: Interface):
|
106 |
+
z = interface.encode(sig)
|
107 |
+
mask = pmask.periodic_mask(z, 16)
|
108 |
zv = interface.coarse_vamp(
|
109 |
+
z,
|
110 |
+
mask,
|
111 |
sampling_steps=num_steps,
|
112 |
)
|
113 |
|
|
|
123 |
after_beat_s=ctx_time,
|
124 |
invert=True
|
125 |
)
|
126 |
+
z = interface.encode(sig)
|
127 |
zv = interface.coarse_vamp(
|
128 |
+
z, beat_mask,
|
|
|
129 |
)
|
130 |
|
131 |
zv = interface.coarse_to_fine(zv)
|
|
|
133 |
return wrapper
|
134 |
|
135 |
def inpaint(ctx_time):
|
136 |
+
def wrapper(sig, interface: Interface):
|
137 |
+
z = interface.encode(sig)
|
138 |
+
mask = pmask.inpaint(z, interface.s2t(ctx_time), interface.s2t(ctx_time))
|
|
|
|
|
|
|
139 |
|
140 |
+
zv = interface.coarse_vamp(z, mask)
|
141 |
zv = interface.coarse_to_fine(zv)
|
142 |
+
|
143 |
return interface.to_signal(zv)
|
144 |
return wrapper
|
145 |
|
146 |
+
def token_noise(noise_amt):
|
147 |
+
def wrapper(sig, interface: Interface):
|
148 |
+
z = interface.encode(sig)
|
149 |
+
mask = pmask.random(z, noise_amt)
|
150 |
+
z = torch.where(
|
151 |
+
mask,
|
152 |
+
torch.randint_like(z, 0, interface.coarse.vocab_size),
|
153 |
+
z
|
154 |
+
)
|
155 |
+
return interface.to_signal(z)
|
156 |
+
return wrapper
|
157 |
+
|
158 |
EXP_REGISTRY = {}
|
159 |
|
160 |
EXP_REGISTRY["gen-compression"] = {
|
|
|
162 |
"reconstructed": reconstructed,
|
163 |
"coarse2fine": coarse2fine,
|
164 |
**{
|
165 |
+
f"{n}_codebooks_downsampled_{x}x": CoarseCond(num_conditioning_codebooks=n, downsample_factor=x)
|
166 |
for (n, x) in (
|
167 |
+
(1, 1), # 1 codebook, no downsampling
|
|
|
|
|
168 |
(4, 4), # 4 codebooks, downsampled 4x
|
|
|
|
|
|
|
169 |
(4, 16), # 4 codebooks, downsampled 16x
|
170 |
(4, 32), # 4 codebooks, downsampled 16x
|
171 |
)
|
172 |
},
|
173 |
+
**{
|
174 |
+
f"token_noise_{x}": mask_ratio_1_step(ratio=x)
|
175 |
+
for x in [0.25, 0.5, 0.75]
|
176 |
+
},
|
177 |
|
178 |
}
|
179 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
180 |
|
181 |
EXP_REGISTRY["sampling-steps"] = {
|
182 |
+
# "codec": reconstructed,
|
183 |
+
**{f"steps_{n}": num_sampling_steps(n) for n in [1, 4, 12, 36, 64, 72]},
|
184 |
}
|
185 |
|
|
|
|
|
|
|
|
|
186 |
|
187 |
EXP_REGISTRY["musical-sampling"] = {
|
188 |
"baseline": baseline,
|
|
|
195 |
@argbind.bind(without_prefix=True)
|
196 |
def main(
|
197 |
sources=[
|
198 |
+
"/media/CHONK/hugo/spotdl/audio-test",
|
199 |
],
|
200 |
output_dir: str = "./samples",
|
201 |
+
max_excerpts: int = 2000,
|
202 |
+
exp_type: str = "gen-compression",
|
203 |
seed: int = 0,
|
204 |
+
ext: str = [".mp3"],
|
205 |
):
|
206 |
at.util.seed(seed)
|
207 |
interface = Interface()
|
|
|
211 |
|
212 |
from audiotools.data.datasets import AudioLoader, AudioDataset
|
213 |
|
214 |
+
loader = AudioLoader(sources=sources, shuffle_state=seed, ext=ext)
|
215 |
dataset = AudioDataset(loader,
|
216 |
sample_rate=interface.codec.sample_rate,
|
217 |
duration=interface.coarse.chunk_size_s,
|
vampnet/interface.py
CHANGED
@@ -321,7 +321,7 @@ class Interface(torch.nn.Module):
|
|
321 |
cz_masked, mask = apply_mask(cz, mask, self.coarse.mask_token)
|
322 |
cz_masked = cz_masked[:, : self.coarse.n_codebooks, :]
|
323 |
|
324 |
-
gen_fn = gen_fn or self.coarse.
|
325 |
c_vamp = gen_fn(
|
326 |
codec=self.codec,
|
327 |
time_steps=cz.shape[-1],
|
|
|
321 |
cz_masked, mask = apply_mask(cz, mask, self.coarse.mask_token)
|
322 |
cz_masked = cz_masked[:, : self.coarse.n_codebooks, :]
|
323 |
|
324 |
+
gen_fn = gen_fn or self.coarse.generate
|
325 |
c_vamp = gen_fn(
|
326 |
codec=self.codec,
|
327 |
time_steps=cz.shape[-1],
|
vampnet/modules/transformer.py
CHANGED
@@ -572,173 +572,13 @@ class VampNet(at.ml.BaseModel):
|
|
572 |
|
573 |
return signal
|
574 |
|
575 |
-
def add_truth_to_logits(
|
576 |
-
self,
|
577 |
-
z_true,
|
578 |
-
z_hat,
|
579 |
-
mask,
|
580 |
-
):
|
581 |
-
z_true = z_true[:, self.n_conditioning_codebooks :, :]
|
582 |
-
mask = mask[:, self.n_conditioning_codebooks :, :]
|
583 |
-
|
584 |
-
truth = F.one_hot(z_true, self.vocab_size)
|
585 |
-
mask = mask[:, :, :, None].expand(-1, -1, -1, self.vocab_size)
|
586 |
-
z_hat = rearrange(
|
587 |
-
z_hat,
|
588 |
-
"b p (t c) -> b c t p",
|
589 |
-
c=self.n_codebooks - self.n_conditioning_codebooks,
|
590 |
-
)
|
591 |
-
|
592 |
-
z_hat = z_hat * mask + truth * (1 - mask)
|
593 |
-
|
594 |
-
z_hat = rearrange(z_hat, "b c t p -> b p (t c)")
|
595 |
-
|
596 |
-
return z_hat
|
597 |
-
|
598 |
-
|
599 |
-
@torch.no_grad()
|
600 |
-
def sample(
|
601 |
-
self,
|
602 |
-
codec,
|
603 |
-
time_steps: int = 300,
|
604 |
-
sampling_steps: int = 36,
|
605 |
-
start_tokens: Optional[torch.Tensor] = None,
|
606 |
-
mask: Optional[torch.Tensor] = None,
|
607 |
-
temperature: Union[float, Tuple[float, float]] = 0.8,
|
608 |
-
top_k: int = None,
|
609 |
-
sample: str = "gumbel",
|
610 |
-
typical_filtering=True,
|
611 |
-
typical_mass=0.2,
|
612 |
-
typical_min_tokens=1,
|
613 |
-
return_signal=True,
|
614 |
-
):
|
615 |
-
if isinstance(temperature, float):
|
616 |
-
temperature = torch.tensor(temperature).repeat(sampling_steps)
|
617 |
-
elif isinstance(temperature, tuple):
|
618 |
-
assert len(temperature) == 2
|
619 |
-
l, h = temperature
|
620 |
-
temperature = torch.linspace(l, h, sampling_steps)
|
621 |
-
else:
|
622 |
-
raise TypeError(f"invalid type for temperature")
|
623 |
-
|
624 |
-
z = start_tokens
|
625 |
-
|
626 |
-
if z is None:
|
627 |
-
z = torch.full((1, self.n_codebooks, time_steps), self.mask_token).to(
|
628 |
-
self.device
|
629 |
-
)
|
630 |
-
|
631 |
-
if mask is None:
|
632 |
-
mask = torch.ones_like(z).to(self.device).int()
|
633 |
-
mask[:, : self.n_conditioning_codebooks, :] = 0.0
|
634 |
-
if mask.ndim == 2:
|
635 |
-
mask = mask[:, None, :].repeat(1, z.shape[1], 1)
|
636 |
-
|
637 |
-
# figure out which timesteps we're keeping
|
638 |
-
keep_mask = 1 - mask
|
639 |
-
|
640 |
-
# any conditioning codebook levels need to be in the keep mask
|
641 |
-
# if self.n_conditioning_codebooks > 0:
|
642 |
-
# cond_mask = torch.ones(z.shape[0], self.n_conditioning_codebooks, z.shape[-1]).to(z.device)
|
643 |
-
# keep_mask = torch.cat([cond_mask, keep_mask], dim=1)
|
644 |
-
|
645 |
-
# flatten
|
646 |
-
keep_mask = codebook_flatten(keep_mask)
|
647 |
-
|
648 |
-
# our r steps
|
649 |
-
r_steps = torch.linspace(0, 1, sampling_steps + 1)[1:].to(self.device)
|
650 |
-
|
651 |
-
# how many tokens did we keep on init?
|
652 |
-
num_kept_on_init = keep_mask.sum()
|
653 |
-
|
654 |
-
# how many codebooks are we inferring vs conditioning on?
|
655 |
-
n_infer_codebooks = self.n_codebooks - self.n_conditioning_codebooks
|
656 |
-
|
657 |
-
for i in range(sampling_steps):
|
658 |
-
# our current temperature
|
659 |
-
tmpt = temperature[i]
|
660 |
-
|
661 |
-
# our current schedule step
|
662 |
-
r = r_steps[i : i + 1]
|
663 |
-
|
664 |
-
with torch.inference_mode():
|
665 |
-
# mask our z
|
666 |
-
keep_mask_unflat = codebook_unflatten(keep_mask, n_c=self.n_codebooks)
|
667 |
-
z_masked = z.masked_fill(~keep_mask_unflat.bool(), self.mask_token)
|
668 |
-
|
669 |
-
# get latents
|
670 |
-
latents = self.embedding.from_codes(z_masked, codec)
|
671 |
-
|
672 |
-
# infer from latents
|
673 |
-
logits = self.forward(latents, r)
|
674 |
-
logits = logits.permute(0, 2, 1) # b, seq, prob
|
675 |
-
|
676 |
-
# the schedule determines how many samples to keep
|
677 |
-
num_tokens_to_infer = (z.shape[-1] * z.shape[-2]) - num_kept_on_init
|
678 |
-
num_to_keep = num_kept_on_init + int(
|
679 |
-
num_tokens_to_infer * (_gamma(1 - r))
|
680 |
-
)
|
681 |
-
|
682 |
-
# figure out which logits we wanna keep
|
683 |
-
if num_to_keep > 0:
|
684 |
-
probs = logits.softmax(dim=-1)
|
685 |
-
|
686 |
-
# do mod self.vocab_size to make sure we don't sample from the mask token
|
687 |
-
# in case the mask token was in the og z
|
688 |
-
keep_probs = F.one_hot(z%self.vocab_size, self.vocab_size)[:, :, :]
|
689 |
-
|
690 |
-
probs = rearrange(
|
691 |
-
probs, "b (t c) p -> b c t p", c=n_infer_codebooks
|
692 |
-
)
|
693 |
-
probs = torch.cat(
|
694 |
-
[keep_probs[:, : self.n_conditioning_codebooks, ...], probs],
|
695 |
-
dim=1,
|
696 |
-
)
|
697 |
-
|
698 |
-
keep_probs = rearrange(
|
699 |
-
keep_probs, "b c t p -> b (t c) p", c=self.n_codebooks
|
700 |
-
)
|
701 |
-
probs = rearrange(probs, "b c t p -> b (t c) p", c=self.n_codebooks)
|
702 |
-
|
703 |
-
keep_prob_mask = keep_mask.unsqueeze(-1).repeat(
|
704 |
-
1, 1, self.vocab_size
|
705 |
-
)
|
706 |
-
probs = (keep_prob_mask.long() * keep_probs) + (
|
707 |
-
1 - keep_prob_mask.long()
|
708 |
-
) * probs
|
709 |
-
|
710 |
-
highest_probs = probs.max(dim=-1, keepdim=False)[0]
|
711 |
-
v, _ = highest_probs.topk(num_to_keep, dim=-1)
|
712 |
-
|
713 |
-
keep_mask = torch.ones_like(keep_mask).bool().clone()
|
714 |
-
keep_mask[highest_probs < v[..., [-1]]] = 0
|
715 |
-
|
716 |
-
logits = torch.log(probs)
|
717 |
-
|
718 |
-
z_inferred = sample_from_logits(
|
719 |
-
logits=logits,
|
720 |
-
top_k=top_k,
|
721 |
-
temperature=tmpt,
|
722 |
-
sample=sample,
|
723 |
-
typical_filtering=typical_filtering,
|
724 |
-
typical_mass=typical_mass,
|
725 |
-
typical_min_tokens=typical_min_tokens,
|
726 |
-
)
|
727 |
-
|
728 |
-
z = codebook_unflatten(z_inferred, n_c=self.n_codebooks)
|
729 |
-
|
730 |
-
|
731 |
-
if return_signal:
|
732 |
-
return self.to_signal(z, codec)
|
733 |
-
else:
|
734 |
-
return z
|
735 |
|
736 |
@torch.no_grad()
|
737 |
def generate(
|
738 |
self,
|
739 |
codec,
|
740 |
time_steps: int = 300,
|
741 |
-
sampling_steps: int =
|
742 |
start_tokens: Optional[torch.Tensor] = None,
|
743 |
mask: Optional[torch.Tensor] = None,
|
744 |
temperature: Union[float, Tuple[float, float]] = 8.0,
|
@@ -747,7 +587,7 @@ class VampNet(at.ml.BaseModel):
|
|
747 |
typical_min_tokens=1,
|
748 |
return_signal=True,
|
749 |
):
|
750 |
-
logging.
|
751 |
|
752 |
#####################
|
753 |
# resolve temperature #
|
@@ -761,7 +601,7 @@ class VampNet(at.ml.BaseModel):
|
|
761 |
else:
|
762 |
raise TypeError(f"invalid type for temperature")
|
763 |
|
764 |
-
logging.
|
765 |
|
766 |
|
767 |
#####################
|
@@ -774,7 +614,7 @@ class VampNet(at.ml.BaseModel):
|
|
774 |
self.device
|
775 |
)
|
776 |
|
777 |
-
logging.
|
778 |
|
779 |
|
780 |
#################
|
@@ -788,7 +628,7 @@ class VampNet(at.ml.BaseModel):
|
|
788 |
mask = mask[:, None, :].repeat(1, z.shape[1], 1)
|
789 |
# init_mask = mask.clone()
|
790 |
|
791 |
-
logging.
|
792 |
|
793 |
|
794 |
###########
|
@@ -796,38 +636,38 @@ class VampNet(at.ml.BaseModel):
|
|
796 |
##########
|
797 |
# apply the mask to z
|
798 |
z_masked = z.masked_fill(mask.bool(), self.mask_token)
|
799 |
-
# logging.
|
800 |
|
801 |
# how many mask tokens to begin with?
|
802 |
num_mask_tokens_at_start = (z_masked == self.mask_token).sum()
|
803 |
-
logging.
|
804 |
|
805 |
# our r steps
|
806 |
r_steps = torch.linspace(1e-10, 1, sampling_steps+1)[1:].to(self.device)
|
807 |
-
logging.
|
808 |
|
809 |
# how many codebooks are we inferring vs conditioning on?
|
810 |
n_infer_codebooks = self.n_codebooks - self.n_conditioning_codebooks
|
811 |
-
logging.
|
812 |
|
813 |
#################
|
814 |
# begin sampling #
|
815 |
#################
|
816 |
|
817 |
for i in range(sampling_steps):
|
818 |
-
logging.
|
819 |
|
820 |
# our current temperature
|
821 |
tmpt = temperature[i]
|
822 |
-
logging.
|
823 |
|
824 |
# our current schedule step
|
825 |
r = r_steps[i : i + 1]
|
826 |
-
logging.
|
827 |
|
828 |
# get latents
|
829 |
latents = self.embedding.from_codes(z_masked, codec)
|
830 |
-
logging.
|
831 |
|
832 |
|
833 |
# infer from latents
|
@@ -841,12 +681,12 @@ class VampNet(at.ml.BaseModel):
|
|
841 |
)
|
842 |
|
843 |
|
844 |
-
logging.
|
845 |
|
846 |
|
847 |
# logits2probs
|
848 |
probs = torch.softmax(logits, dim=-1)
|
849 |
-
logging.
|
850 |
|
851 |
|
852 |
# sample from logits with multinomial sampling
|
@@ -857,7 +697,7 @@ class VampNet(at.ml.BaseModel):
|
|
857 |
|
858 |
sampled_z = rearrange(sampled_z, "(b seq)-> b seq", b=b)
|
859 |
probs = rearrange(probs, "(b seq) prob -> b seq prob", b=b)
|
860 |
-
logging.
|
861 |
|
862 |
|
863 |
# flatten z_masked and mask, so we can deal with the sampling logic
|
@@ -868,12 +708,12 @@ class VampNet(at.ml.BaseModel):
|
|
868 |
mask = (z_masked == self.mask_token).int()
|
869 |
|
870 |
# update the mask, remove conditioning codebooks from the mask
|
871 |
-
logging.
|
872 |
# add z back into sampled z where the mask was false
|
873 |
sampled_z = torch.where(
|
874 |
mask.bool(), sampled_z, z_masked
|
875 |
)
|
876 |
-
logging.
|
877 |
|
878 |
|
879 |
# get the confidences: which tokens did we sample?
|
@@ -891,7 +731,7 @@ class VampNet(at.ml.BaseModel):
|
|
891 |
|
892 |
# get the num tokens to mask, according to the schedule
|
893 |
num_to_mask = torch.floor(_gamma(r) * num_mask_tokens_at_start).unsqueeze(1).long()
|
894 |
-
logging.
|
895 |
|
896 |
num_to_mask = torch.maximum(
|
897 |
torch.tensor(1),
|
@@ -911,17 +751,17 @@ class VampNet(at.ml.BaseModel):
|
|
911 |
z_masked = torch.where(
|
912 |
mask.bool(), self.mask_token, sampled_z
|
913 |
)
|
914 |
-
logging.
|
915 |
|
916 |
z_masked = codebook_unflatten(z_masked, n_infer_codebooks)
|
917 |
mask = codebook_unflatten(mask, n_infer_codebooks)
|
918 |
-
logging.
|
919 |
|
920 |
# add conditioning codebooks back to z_masked
|
921 |
z_masked = torch.cat(
|
922 |
(z[:, :self.n_conditioning_codebooks, :], z_masked), dim=1
|
923 |
)
|
924 |
-
logging.
|
925 |
|
926 |
|
927 |
# add conditioning codebooks back to sampled_z
|
@@ -930,7 +770,7 @@ class VampNet(at.ml.BaseModel):
|
|
930 |
(z[:, :self.n_conditioning_codebooks, :], sampled_z), dim=1
|
931 |
)
|
932 |
|
933 |
-
logging.
|
934 |
|
935 |
if return_signal:
|
936 |
return self.to_signal(sampled_z, codec)
|
@@ -945,28 +785,28 @@ def mask_by_random_topk(num_to_mask: int, probs: torch.Tensor, temperature: floa
|
|
945 |
probs (torch.Tensor): probabilities for each sampled event, shape (batch, seq)
|
946 |
temperature (float, optional): temperature. Defaults to 1.0.
|
947 |
"""
|
948 |
-
logging.
|
949 |
-
logging.
|
950 |
-
logging.
|
951 |
-
logging.
|
952 |
-
logging.
|
953 |
|
954 |
confidence = torch.log(probs) + temperature * gumbel_noise_like(probs)
|
955 |
-
logging.
|
956 |
|
957 |
sorted_confidence, sorted_idx = confidence.sort(dim=-1)
|
958 |
-
logging.
|
959 |
-
logging.
|
960 |
|
961 |
# get the cut off threshold, given the mask length
|
962 |
cut_off = torch.take_along_dim(
|
963 |
sorted_confidence, num_to_mask, axis=-1
|
964 |
)
|
965 |
-
logging.
|
966 |
|
967 |
# mask out the tokens
|
968 |
mask = confidence < cut_off
|
969 |
-
logging.
|
970 |
|
971 |
return mask
|
972 |
|
@@ -999,61 +839,6 @@ def typical_filter(
|
|
999 |
logits = rearrange(x_flat, "(b t) l -> b t l", t=nt)
|
1000 |
return logits
|
1001 |
|
1002 |
-
def sample_from_logits(
|
1003 |
-
logits,
|
1004 |
-
top_k: int = None,
|
1005 |
-
temperature: float = 1.0,
|
1006 |
-
sample: str = "multinomial",
|
1007 |
-
typical_filtering=False,
|
1008 |
-
typical_mass=0.2,
|
1009 |
-
typical_min_tokens=1,
|
1010 |
-
):
|
1011 |
-
# add temperature
|
1012 |
-
logits = logits / temperature
|
1013 |
-
|
1014 |
-
# add topk
|
1015 |
-
if top_k is not None and typical_filtering == False:
|
1016 |
-
v, topk_idx = logits.topk(top_k)
|
1017 |
-
logits[logits < v[..., [-1]]] = -float("inf")
|
1018 |
-
|
1019 |
-
if typical_filtering:
|
1020 |
-
assert top_k is None
|
1021 |
-
nb, nt, _ = logits.shape
|
1022 |
-
x_flat = rearrange(logits, "b t l -> (b t ) l")
|
1023 |
-
x_flat_norm = torch.nn.functional.log_softmax(x_flat, dim=-1)
|
1024 |
-
x_flat_norm_p = torch.exp(x_flat_norm)
|
1025 |
-
entropy = -(x_flat_norm * x_flat_norm_p).nansum(-1, keepdim=True)
|
1026 |
-
|
1027 |
-
c_flat_shifted = torch.abs((-x_flat_norm) - entropy)
|
1028 |
-
c_flat_sorted, x_flat_indices = torch.sort(c_flat_shifted, descending=False)
|
1029 |
-
x_flat_cumsum = (
|
1030 |
-
x_flat.gather(-1, x_flat_indices).softmax(dim=-1).cumsum(dim=-1)
|
1031 |
-
)
|
1032 |
-
|
1033 |
-
last_ind = (x_flat_cumsum < typical_mass).sum(dim=-1)
|
1034 |
-
sorted_indices_to_remove = c_flat_sorted > c_flat_sorted.gather(
|
1035 |
-
1, last_ind.view(-1, 1)
|
1036 |
-
)
|
1037 |
-
if typical_min_tokens > 1:
|
1038 |
-
sorted_indices_to_remove[..., :typical_min_tokens] = 0
|
1039 |
-
indices_to_remove = sorted_indices_to_remove.scatter(
|
1040 |
-
1, x_flat_indices, sorted_indices_to_remove
|
1041 |
-
)
|
1042 |
-
x_flat = x_flat.masked_fill(indices_to_remove, -float("Inf"))
|
1043 |
-
logits = rearrange(x_flat, "(b t) l -> b t l", t=nt)
|
1044 |
-
|
1045 |
-
if sample == "multinomial":
|
1046 |
-
probs = torch.softmax(logits, dim=-1)
|
1047 |
-
inferred = torch.stack([pr.multinomial(1).squeeze(-1) for pr in probs])
|
1048 |
-
elif sample == "argmax":
|
1049 |
-
inferred = torch.softmax(logits, dim=-1).argmax(dim=-1)
|
1050 |
-
elif sample == "gumbel":
|
1051 |
-
inferred = gumbel_sample(logits, dim=-1)
|
1052 |
-
else:
|
1053 |
-
raise ValueError(f"invalid sampling method: {sample}")
|
1054 |
-
|
1055 |
-
return inferred
|
1056 |
-
|
1057 |
|
1058 |
if __name__ == "__main__":
|
1059 |
# import argbind
|
|
|
572 |
|
573 |
return signal
|
574 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
575 |
|
576 |
@torch.no_grad()
|
577 |
def generate(
|
578 |
self,
|
579 |
codec,
|
580 |
time_steps: int = 300,
|
581 |
+
sampling_steps: int = 24,
|
582 |
start_tokens: Optional[torch.Tensor] = None,
|
583 |
mask: Optional[torch.Tensor] = None,
|
584 |
temperature: Union[float, Tuple[float, float]] = 8.0,
|
|
|
587 |
typical_min_tokens=1,
|
588 |
return_signal=True,
|
589 |
):
|
590 |
+
logging.debug(f"beginning generation with {sampling_steps} steps")
|
591 |
|
592 |
#####################
|
593 |
# resolve temperature #
|
|
|
601 |
else:
|
602 |
raise TypeError(f"invalid type for temperature")
|
603 |
|
604 |
+
logging.debug(f"temperature: {temperature}")
|
605 |
|
606 |
|
607 |
#####################
|
|
|
614 |
self.device
|
615 |
)
|
616 |
|
617 |
+
logging.debug(f"created z with shape {z.shape}")
|
618 |
|
619 |
|
620 |
#################
|
|
|
628 |
mask = mask[:, None, :].repeat(1, z.shape[1], 1)
|
629 |
# init_mask = mask.clone()
|
630 |
|
631 |
+
logging.debug(f"created mask with shape {mask.shape}")
|
632 |
|
633 |
|
634 |
###########
|
|
|
636 |
##########
|
637 |
# apply the mask to z
|
638 |
z_masked = z.masked_fill(mask.bool(), self.mask_token)
|
639 |
+
# logging.debug(f"z_masked: {z_masked}")
|
640 |
|
641 |
# how many mask tokens to begin with?
|
642 |
num_mask_tokens_at_start = (z_masked == self.mask_token).sum()
|
643 |
+
logging.debug(f"num mask tokens at start: {num_mask_tokens_at_start}")
|
644 |
|
645 |
# our r steps
|
646 |
r_steps = torch.linspace(1e-10, 1, sampling_steps+1)[1:].to(self.device)
|
647 |
+
logging.debug(f"r steps: {r_steps}")
|
648 |
|
649 |
# how many codebooks are we inferring vs conditioning on?
|
650 |
n_infer_codebooks = self.n_codebooks - self.n_conditioning_codebooks
|
651 |
+
logging.debug(f"n infer codebooks: {n_infer_codebooks}")
|
652 |
|
653 |
#################
|
654 |
# begin sampling #
|
655 |
#################
|
656 |
|
657 |
for i in range(sampling_steps):
|
658 |
+
logging.debug(f"step {i} of {sampling_steps}")
|
659 |
|
660 |
# our current temperature
|
661 |
tmpt = temperature[i]
|
662 |
+
logging.debug(f"temperature: {tmpt}")
|
663 |
|
664 |
# our current schedule step
|
665 |
r = r_steps[i : i + 1]
|
666 |
+
logging.debug(f"r: {r}")
|
667 |
|
668 |
# get latents
|
669 |
latents = self.embedding.from_codes(z_masked, codec)
|
670 |
+
logging.debug(f"computed latents with shape: {latents.shape}")
|
671 |
|
672 |
|
673 |
# infer from latents
|
|
|
681 |
)
|
682 |
|
683 |
|
684 |
+
logging.debug(f"permuted logits with shape: {logits.shape}")
|
685 |
|
686 |
|
687 |
# logits2probs
|
688 |
probs = torch.softmax(logits, dim=-1)
|
689 |
+
logging.debug(f"computed probs with shape: {probs.shape}")
|
690 |
|
691 |
|
692 |
# sample from logits with multinomial sampling
|
|
|
697 |
|
698 |
sampled_z = rearrange(sampled_z, "(b seq)-> b seq", b=b)
|
699 |
probs = rearrange(probs, "(b seq) prob -> b seq prob", b=b)
|
700 |
+
logging.debug(f"sampled z with shape: {sampled_z.shape}")
|
701 |
|
702 |
|
703 |
# flatten z_masked and mask, so we can deal with the sampling logic
|
|
|
708 |
mask = (z_masked == self.mask_token).int()
|
709 |
|
710 |
# update the mask, remove conditioning codebooks from the mask
|
711 |
+
logging.debug(f"updated mask with shape: {mask.shape}")
|
712 |
# add z back into sampled z where the mask was false
|
713 |
sampled_z = torch.where(
|
714 |
mask.bool(), sampled_z, z_masked
|
715 |
)
|
716 |
+
logging.debug(f"added z back into sampled z with shape: {sampled_z.shape}")
|
717 |
|
718 |
|
719 |
# get the confidences: which tokens did we sample?
|
|
|
731 |
|
732 |
# get the num tokens to mask, according to the schedule
|
733 |
num_to_mask = torch.floor(_gamma(r) * num_mask_tokens_at_start).unsqueeze(1).long()
|
734 |
+
logging.debug(f"num to mask: {num_to_mask}")
|
735 |
|
736 |
num_to_mask = torch.maximum(
|
737 |
torch.tensor(1),
|
|
|
751 |
z_masked = torch.where(
|
752 |
mask.bool(), self.mask_token, sampled_z
|
753 |
)
|
754 |
+
logging.debug(f"updated z_masked with shape: {z_masked.shape}")
|
755 |
|
756 |
z_masked = codebook_unflatten(z_masked, n_infer_codebooks)
|
757 |
mask = codebook_unflatten(mask, n_infer_codebooks)
|
758 |
+
logging.debug(f"unflattened z_masked with shape: {z_masked.shape}")
|
759 |
|
760 |
# add conditioning codebooks back to z_masked
|
761 |
z_masked = torch.cat(
|
762 |
(z[:, :self.n_conditioning_codebooks, :], z_masked), dim=1
|
763 |
)
|
764 |
+
logging.debug(f"added conditioning codebooks back to z_masked with shape: {z_masked.shape}")
|
765 |
|
766 |
|
767 |
# add conditioning codebooks back to sampled_z
|
|
|
770 |
(z[:, :self.n_conditioning_codebooks, :], sampled_z), dim=1
|
771 |
)
|
772 |
|
773 |
+
logging.debug(f"finished sampling")
|
774 |
|
775 |
if return_signal:
|
776 |
return self.to_signal(sampled_z, codec)
|
|
|
785 |
probs (torch.Tensor): probabilities for each sampled event, shape (batch, seq)
|
786 |
temperature (float, optional): temperature. Defaults to 1.0.
|
787 |
"""
|
788 |
+
logging.debug(f"masking by random topk")
|
789 |
+
logging.debug(f"num to mask: {num_to_mask}")
|
790 |
+
logging.debug(f"probs shape: {probs.shape}")
|
791 |
+
logging.debug(f"temperature: {temperature}")
|
792 |
+
logging.debug("")
|
793 |
|
794 |
confidence = torch.log(probs) + temperature * gumbel_noise_like(probs)
|
795 |
+
logging.debug(f"confidence shape: {confidence.shape}")
|
796 |
|
797 |
sorted_confidence, sorted_idx = confidence.sort(dim=-1)
|
798 |
+
logging.debug(f"sorted confidence shape: {sorted_confidence.shape}")
|
799 |
+
logging.debug(f"sorted idx shape: {sorted_idx.shape}")
|
800 |
|
801 |
# get the cut off threshold, given the mask length
|
802 |
cut_off = torch.take_along_dim(
|
803 |
sorted_confidence, num_to_mask, axis=-1
|
804 |
)
|
805 |
+
logging.debug(f"cut off shape: {cut_off.shape}")
|
806 |
|
807 |
# mask out the tokens
|
808 |
mask = confidence < cut_off
|
809 |
+
logging.debug(f"mask shape: {mask.shape}")
|
810 |
|
811 |
return mask
|
812 |
|
|
|
839 |
logits = rearrange(x_flat, "(b t) l -> b t l", t=nt)
|
840 |
return logits
|
841 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
842 |
|
843 |
if __name__ == "__main__":
|
844 |
# import argbind
|