Upload seamless_communication/inference/generator.py with huggingface_hub
Browse files
seamless_communication/inference/generator.py
ADDED
@@ -0,0 +1,364 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# Copyright (c) Meta Platforms, Inc. and affiliates
|
2 |
+
# All rights reserved.
|
3 |
+
#
|
4 |
+
# This source code is licensed under the license found in the
|
5 |
+
# MIT_LICENSE file in the root directory of this source tree.
|
6 |
+
|
7 |
+
from dataclasses import dataclass
|
8 |
+
from typing import List, Optional, Tuple
|
9 |
+
|
10 |
+
import torch
|
11 |
+
from fairseq2.data import SequenceData, StringLike
|
12 |
+
from fairseq2.data.text import TextTokenizer
|
13 |
+
from fairseq2.generation import (
|
14 |
+
BeamSearchSeq2SeqGenerator,
|
15 |
+
Seq2SeqGenerator,
|
16 |
+
SequenceToTextConverter,
|
17 |
+
StepProcessor,
|
18 |
+
)
|
19 |
+
from fairseq2.nn.padding import (
|
20 |
+
PaddingMask,
|
21 |
+
apply_padding_mask,
|
22 |
+
get_seqs_and_padding_mask,
|
23 |
+
pad_seqs,
|
24 |
+
)
|
25 |
+
from fairseq2.nn.utils.module import infer_device
|
26 |
+
from torch import Tensor
|
27 |
+
|
28 |
+
from seamless_communication.models.unity.model import (
|
29 |
+
UnitYModel,
|
30 |
+
UnitYT2UModel,
|
31 |
+
UnitYX2TModel,
|
32 |
+
)
|
33 |
+
from seamless_communication.models.unity.unit_tokenizer import (
|
34 |
+
UnitTokenDecoder,
|
35 |
+
UnitTokenizer,
|
36 |
+
)
|
37 |
+
|
38 |
+
|
39 |
+
def remove_consecutive_repeated_ngrams(
|
40 |
+
sequence: List[int], min_size: int = 1, max_size: int = 40
|
41 |
+
) -> List[int]:
|
42 |
+
assert 1 <= min_size <= max_size
|
43 |
+
drop_idx = set() # indices that will be dropped from the sequence
|
44 |
+
|
45 |
+
# start from the beginning, check if an ngram of size k (for k=max..min) is
|
46 |
+
# followed by its copy, if so delete the first one, and start over after
|
47 |
+
# the deleted ngram.
|
48 |
+
start = 0
|
49 |
+
while start < len(sequence):
|
50 |
+
for k in range(max_size, min_size - 1, -1):
|
51 |
+
if sequence[start : start + k] == sequence[start + k : start + k + k]:
|
52 |
+
drop_idx |= set(range(start, start + k))
|
53 |
+
start += k - 1 # assumes repeating subsequences don't overlap
|
54 |
+
break
|
55 |
+
start += 1
|
56 |
+
return [token for idx, token in enumerate(sequence) if idx not in drop_idx]
|
57 |
+
|
58 |
+
|
59 |
+
@dataclass
|
60 |
+
class SequenceGeneratorOptions:
|
61 |
+
"""Holds the options to pass to a sequence generator."""
|
62 |
+
|
63 |
+
beam_size: int = 5
|
64 |
+
"""The beam size."""
|
65 |
+
|
66 |
+
soft_max_seq_len: Tuple[int, int] = (1, 200)
|
67 |
+
"""The terms ``a`` and ``b`` of ``ax + b`` where ``x`` is the source
|
68 |
+
sequence length. The generated sequences (including prefix sequence) will
|
69 |
+
have the maximum length of ``min(hard_max_seq_len, ax + b)``. See also
|
70 |
+
``hard_max_seq_len``."""
|
71 |
+
|
72 |
+
hard_max_seq_len: int = 1024
|
73 |
+
"""The hard limit on maximum length of generated sequences."""
|
74 |
+
|
75 |
+
step_processor: Optional[StepProcessor] = None
|
76 |
+
"""The processor called at each generation step."""
|
77 |
+
|
78 |
+
unk_penalty: float = 0.0
|
79 |
+
"""The UNK symbol penalty, where values less than 0 produce more UNKs;
|
80 |
+
values greater than 0 produce fewer UNKs."""
|
81 |
+
|
82 |
+
len_penalty: float = 1.0
|
83 |
+
"""The length penalty, where values less than 1.0 favor shorter
|
84 |
+
sequences; values greater than 1.0 favor longer sequences."""
|
85 |
+
|
86 |
+
|
87 |
+
class UnitYGenerator:
|
88 |
+
"""Generates text translations and speech units from a UnitY model."""
|
89 |
+
|
90 |
+
model: UnitYModel
|
91 |
+
s2t_converter: SequenceToTextConverter
|
92 |
+
t2t_converter: Optional[SequenceToTextConverter]
|
93 |
+
unit_decoder: Optional[UnitTokenDecoder]
|
94 |
+
unit_prefix_indices: Optional[Tensor]
|
95 |
+
unit_generator: Optional[Seq2SeqGenerator]
|
96 |
+
|
97 |
+
def __init__(
|
98 |
+
self,
|
99 |
+
model: UnitYModel,
|
100 |
+
text_tokenizer: TextTokenizer,
|
101 |
+
target_lang: str,
|
102 |
+
unit_tokenizer: Optional[UnitTokenizer] = None,
|
103 |
+
text_opts: Optional[SequenceGeneratorOptions] = None,
|
104 |
+
unit_opts: Optional[SequenceGeneratorOptions] = None,
|
105 |
+
) -> None:
|
106 |
+
"""
|
107 |
+
:param model:
|
108 |
+
The UnitY model to use for generation.
|
109 |
+
:param text_tokenizer:
|
110 |
+
The text tokenizer to use.
|
111 |
+
:param unit_tokenizer:
|
112 |
+
The unit tokenizer to use.
|
113 |
+
:param target_lang:
|
114 |
+
The target language.
|
115 |
+
:param text_generator_opts:
|
116 |
+
The options to pass to the underlying text :class:`Seq2SeqGenerator`.
|
117 |
+
:param unit_generator_opts:
|
118 |
+
The options to pass to the underlying unit :class:`Seq2SeqGenerator`.
|
119 |
+
"""
|
120 |
+
model.eval()
|
121 |
+
|
122 |
+
self.model = model
|
123 |
+
|
124 |
+
if text_opts is None:
|
125 |
+
text_opts = SequenceGeneratorOptions()
|
126 |
+
|
127 |
+
if model.text_decoder is None:
|
128 |
+
raise ValueError(
|
129 |
+
"`UnitYGenerator` requires a text decoder, but the current UnitY model does not have one."
|
130 |
+
)
|
131 |
+
assert model.text_decoder_frontend is not None
|
132 |
+
assert model.final_proj is not None
|
133 |
+
|
134 |
+
s2t_model = UnitYX2TModel(
|
135 |
+
encoder_frontend=model.speech_encoder_frontend,
|
136 |
+
encoder=model.speech_encoder,
|
137 |
+
decoder_frontend=model.text_decoder_frontend,
|
138 |
+
decoder=model.text_decoder,
|
139 |
+
final_proj=model.final_proj,
|
140 |
+
target_vocab_info=model.target_vocab_info,
|
141 |
+
)
|
142 |
+
|
143 |
+
step_processors = []
|
144 |
+
if text_opts.step_processor is not None:
|
145 |
+
step_processors.append(text_opts.step_processor)
|
146 |
+
|
147 |
+
generator = BeamSearchSeq2SeqGenerator(
|
148 |
+
s2t_model,
|
149 |
+
beam_size=text_opts.beam_size,
|
150 |
+
max_gen_len=text_opts.soft_max_seq_len,
|
151 |
+
max_seq_len=text_opts.hard_max_seq_len,
|
152 |
+
echo_prompt=True,
|
153 |
+
step_processors=step_processors,
|
154 |
+
unk_penalty=text_opts.unk_penalty,
|
155 |
+
len_penalty=text_opts.len_penalty,
|
156 |
+
)
|
157 |
+
self.s2t_converter = SequenceToTextConverter(
|
158 |
+
generator, text_tokenizer, "translation", target_lang
|
159 |
+
)
|
160 |
+
|
161 |
+
if model.text_encoder is None:
|
162 |
+
self.t2t_generator = None
|
163 |
+
else:
|
164 |
+
assert model.text_encoder_frontend is not None
|
165 |
+
assert model.text_encoder is not None
|
166 |
+
t2t_model = UnitYX2TModel(
|
167 |
+
encoder_frontend=model.text_encoder_frontend,
|
168 |
+
encoder=model.text_encoder,
|
169 |
+
decoder_frontend=model.text_decoder_frontend,
|
170 |
+
decoder=model.text_decoder,
|
171 |
+
final_proj=model.final_proj,
|
172 |
+
target_vocab_info=model.target_vocab_info,
|
173 |
+
)
|
174 |
+
generator = BeamSearchSeq2SeqGenerator(
|
175 |
+
t2t_model,
|
176 |
+
beam_size=text_opts.beam_size,
|
177 |
+
max_gen_len=text_opts.soft_max_seq_len,
|
178 |
+
max_seq_len=text_opts.hard_max_seq_len,
|
179 |
+
echo_prompt=True,
|
180 |
+
step_processors=step_processors,
|
181 |
+
unk_penalty=text_opts.unk_penalty,
|
182 |
+
len_penalty=text_opts.len_penalty,
|
183 |
+
)
|
184 |
+
self.t2t_converter = SequenceToTextConverter(
|
185 |
+
generator, text_tokenizer, "translation", target_lang
|
186 |
+
)
|
187 |
+
|
188 |
+
self.unit_generator = None
|
189 |
+
self.unit_decoder = None
|
190 |
+
# Set up unit generator.
|
191 |
+
if unit_tokenizer is not None:
|
192 |
+
if model.t2u_model is None:
|
193 |
+
raise ValueError(
|
194 |
+
"`model` does not have a T2U sub-model when `unit_tokenizer` is not None."
|
195 |
+
)
|
196 |
+
|
197 |
+
self.unit_decoder = unit_tokenizer.create_decoder()
|
198 |
+
|
199 |
+
unit_encoder = unit_tokenizer.create_encoder(
|
200 |
+
lang=target_lang, device=infer_device(model.t2u_model)
|
201 |
+
)
|
202 |
+
|
203 |
+
self.unit_prefix_indices = unit_encoder.prefix_indices
|
204 |
+
|
205 |
+
if isinstance(self.model.t2u_model, UnitYT2UModel):
|
206 |
+
if unit_opts is None:
|
207 |
+
# Speech sequences are typically much longer than text sequences.
|
208 |
+
unit_opts = SequenceGeneratorOptions(
|
209 |
+
soft_max_seq_len=(25, 50), hard_max_seq_len=5000
|
210 |
+
)
|
211 |
+
|
212 |
+
step_processors = []
|
213 |
+
if unit_opts.step_processor is not None:
|
214 |
+
step_processors.append(unit_opts.step_processor)
|
215 |
+
|
216 |
+
self.unit_generator = BeamSearchSeq2SeqGenerator(
|
217 |
+
self.model.t2u_model,
|
218 |
+
beam_size=unit_opts.beam_size,
|
219 |
+
max_gen_len=unit_opts.soft_max_seq_len,
|
220 |
+
max_seq_len=unit_opts.hard_max_seq_len,
|
221 |
+
echo_prompt=True,
|
222 |
+
step_processors=step_processors,
|
223 |
+
unk_penalty=unit_opts.unk_penalty,
|
224 |
+
len_penalty=unit_opts.len_penalty,
|
225 |
+
)
|
226 |
+
|
227 |
+
@torch.inference_mode()
|
228 |
+
def __call__(
|
229 |
+
self,
|
230 |
+
source_seqs: Tensor,
|
231 |
+
source_padding_mask: Optional[PaddingMask],
|
232 |
+
input_modality: str = "speech",
|
233 |
+
output_modality: str = "speech",
|
234 |
+
ngram_filtering: bool = False,
|
235 |
+
duration_factor: float = 1.0,
|
236 |
+
prosody_encoder_input: Optional[SequenceData] = None,
|
237 |
+
) -> Tuple[List[StringLike], Optional[Tensor]]:
|
238 |
+
"""
|
239 |
+
:param source_seqs:
|
240 |
+
The source sequences to use for generation. *Shape:* :math:`(N,S,*)`,
|
241 |
+
where :math:`N` is the batch size, :math:`S` is the sequence length,
|
242 |
+
and :math:`*` is any number of sequence-specific dimensions
|
243 |
+
including none.
|
244 |
+
:param source_padding_mask:
|
245 |
+
The padding mask of ``source_seqs``. *Shape:* :math:`(N,S)`, where
|
246 |
+
:math:`N` is the batch size and :math:`S` is the sequence length.
|
247 |
+
:param input_modality:
|
248 |
+
The type of modality to encode.
|
249 |
+
:param output_modality:
|
250 |
+
The type of modality to decode.
|
251 |
+
:param ngram_filtering:
|
252 |
+
If True, removes consecutive repeated ngrams
|
253 |
+
from the decoded unit output.
|
254 |
+
|
255 |
+
:returns:
|
256 |
+
- The output of the text generator.
|
257 |
+
- The output of the unit generator.
|
258 |
+
"""
|
259 |
+
|
260 |
+
if input_modality == "speech":
|
261 |
+
texts, text_gen_output = self.s2t_converter.batch_convert(
|
262 |
+
source_seqs, source_padding_mask
|
263 |
+
)
|
264 |
+
elif input_modality == "text":
|
265 |
+
if self.t2t_converter is None:
|
266 |
+
raise ValueError(
|
267 |
+
"Please set `use_text_encoder` to `True` in your model config to encode text."
|
268 |
+
)
|
269 |
+
texts, text_gen_output = self.t2t_converter.batch_convert(
|
270 |
+
source_seqs, source_padding_mask
|
271 |
+
)
|
272 |
+
else:
|
273 |
+
raise ValueError(f"Unsupported input_modality: {input_modality}")
|
274 |
+
|
275 |
+
# We skip T2U when we only need to output text.
|
276 |
+
if output_modality == "text":
|
277 |
+
return texts, None
|
278 |
+
|
279 |
+
assert self.model.target_vocab_info.pad_idx is not None
|
280 |
+
|
281 |
+
text_seq_list = [h[0].seq for h in text_gen_output.hypotheses]
|
282 |
+
|
283 |
+
text_seqs, text_padding_mask = pad_seqs(
|
284 |
+
text_seq_list, self.model.target_vocab_info.pad_idx
|
285 |
+
)
|
286 |
+
|
287 |
+
# Manually trim the final EOS token to be consistent with fairseq.
|
288 |
+
text_seqs = text_seqs[:, :-1]
|
289 |
+
|
290 |
+
if text_padding_mask is not None:
|
291 |
+
text_padding_mask = text_padding_mask.trim(1)
|
292 |
+
|
293 |
+
# Use the output of the text generator to compute the decoder output.
|
294 |
+
decoder_output, decoder_padding_mask = self.model.decode(
|
295 |
+
text_seqs,
|
296 |
+
text_padding_mask,
|
297 |
+
text_gen_output.encoder_output,
|
298 |
+
text_gen_output.encoder_padding_mask,
|
299 |
+
)
|
300 |
+
|
301 |
+
assert self.model.t2u_model is not None
|
302 |
+
assert self.unit_decoder is not None
|
303 |
+
|
304 |
+
unit_gen_output = None
|
305 |
+
prosody_encoder_out = None
|
306 |
+
if self.model.prosody_encoder_model is not None:
|
307 |
+
assert prosody_encoder_input is not None
|
308 |
+
prosody_input_seqs, prosody_padding_mask = get_seqs_and_padding_mask(
|
309 |
+
prosody_encoder_input
|
310 |
+
)
|
311 |
+
prosody_encoder_out = self.model.prosody_encoder_model(
|
312 |
+
prosody_input_seqs,
|
313 |
+
prosody_padding_mask,
|
314 |
+
).unsqueeze(1)
|
315 |
+
|
316 |
+
if isinstance(self.model.t2u_model, UnitYT2UModel):
|
317 |
+
assert self.unit_generator is not None
|
318 |
+
assert self.unit_prefix_indices is not None
|
319 |
+
|
320 |
+
# (S_pre) -> (N, S_pre)
|
321 |
+
prefix_seqs = self.unit_prefix_indices.expand(decoder_output.size(0), -1)
|
322 |
+
|
323 |
+
unit_gen_output = self.unit_generator(
|
324 |
+
source_seqs=decoder_output,
|
325 |
+
source_padding_mask=decoder_padding_mask,
|
326 |
+
prompt_seqs=prefix_seqs,
|
327 |
+
prompt_padding_mask=None,
|
328 |
+
)
|
329 |
+
|
330 |
+
assert self.model.t2u_model.target_vocab_info.pad_idx is not None
|
331 |
+
|
332 |
+
unit_seq_list = [h[0].seq for h in unit_gen_output.hypotheses]
|
333 |
+
|
334 |
+
unit_seqs, _ = pad_seqs(
|
335 |
+
unit_seq_list, self.model.t2u_model.target_vocab_info.pad_idx
|
336 |
+
)
|
337 |
+
else:
|
338 |
+
t2u_model_output, decoder_padding_mask, _ = self.model.t2u_model(
|
339 |
+
text_decoder_output=decoder_output,
|
340 |
+
text_decoder_padding_mask=decoder_padding_mask,
|
341 |
+
text_seqs=text_seqs,
|
342 |
+
duration_factor=duration_factor,
|
343 |
+
film_cond_emb=prosody_encoder_out,
|
344 |
+
)
|
345 |
+
# (B, S_unit, V_unit)
|
346 |
+
unit_seqs = t2u_model_output.logits.argmax(dim=2)
|
347 |
+
# Apply the padding mask to the generated units.
|
348 |
+
unit_seqs = apply_padding_mask(
|
349 |
+
unit_seqs, decoder_padding_mask, t2u_model_output.vocab_info.pad_idx
|
350 |
+
)
|
351 |
+
|
352 |
+
# Convert to speech units.
|
353 |
+
units = self.unit_decoder(unit_seqs)
|
354 |
+
|
355 |
+
# ngram-filtering doesn't apply to NAR unit decoding.
|
356 |
+
if ngram_filtering and isinstance(self.model.t2u_model, UnitYT2UModel):
|
357 |
+
if units.size(0) > 1:
|
358 |
+
raise NotImplementedError(
|
359 |
+
"unit ngram_filtering is not implemented for batch_size > 1."
|
360 |
+
)
|
361 |
+
arr = remove_consecutive_repeated_ngrams(units[0].tolist())
|
362 |
+
units = torch.tensor(arr).to(units).unsqueeze(0)
|
363 |
+
|
364 |
+
return texts, units
|