Upload 4 files
Browse files- processor_config.json +1 -1
- ultravox_model.py +85 -48
- ultravox_processing.py +188 -75
processor_config.json
CHANGED
@@ -5,7 +5,7 @@
|
|
5 |
"auto_map": {
|
6 |
"AutoProcessor": "ultravox_processing.UltravoxProcessor"
|
7 |
},
|
8 |
-
"encoder_ds_factor":
|
9 |
"processor_class": "UltravoxProcessor",
|
10 |
"stack_factor": 8
|
11 |
}
|
|
|
5 |
"auto_map": {
|
6 |
"AutoProcessor": "ultravox_processing.UltravoxProcessor"
|
7 |
},
|
8 |
+
"encoder_ds_factor": 2,
|
9 |
"processor_class": "UltravoxProcessor",
|
10 |
"stack_factor": 8
|
11 |
}
|
ultravox_model.py
CHANGED
@@ -1,6 +1,6 @@
|
|
1 |
import logging
|
2 |
import re
|
3 |
-
from typing import Any, Dict, Optional, Set, Tuple, Union
|
4 |
|
5 |
import peft
|
6 |
import torch
|
@@ -10,6 +10,7 @@ import transformers
|
|
10 |
import transformers.activations
|
11 |
import transformers.modeling_outputs
|
12 |
import transformers.models
|
|
|
13 |
from transformers.models.whisper import modeling_whisper as whisper
|
14 |
|
15 |
# We must use relative import in this directory to allow uploading to HF Hub
|
@@ -19,7 +20,7 @@ from .ultravox_config import LossFunction
|
|
19 |
from .ultravox_config import UltravoxConfig
|
20 |
|
21 |
|
22 |
-
class UltravoxModel(transformers.LlamaPreTrainedModel):
|
23 |
"""
|
24 |
The Ultravox model which consists of an audio encoder and a language model.
|
25 |
|
@@ -37,6 +38,9 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
|
|
37 |
config: UltravoxConfig # for type hinting
|
38 |
# Usually we load encoder and LLM weights from a pretrained model separately, so they are allowed to be missing
|
39 |
_keys_to_ignore_on_load_missing = ["audio_tower.*", "language_model.*"]
|
|
|
|
|
|
|
40 |
|
41 |
def __init__(self, config: UltravoxConfig):
|
42 |
super().__init__(config)
|
@@ -46,15 +50,16 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
|
|
46 |
self.vocab_size = config.vocab_size
|
47 |
|
48 |
self.audio_tower = self._create_audio_tower(config)
|
|
|
|
|
|
|
49 |
self.multi_modal_projector = self._create_multi_modal_projector(config)
|
50 |
self.language_model = self._create_language_model(config)
|
51 |
|
52 |
# Determine no_split_modules dynamically to use with FSDP auto_wrap policy.
|
53 |
# FSDP throws an error if some of the layer types are not found in the model.
|
54 |
-
# This would be something like ["LlamaDecoderLayer"
|
55 |
-
self._no_split_modules =
|
56 |
-
self.audio_tower._no_split_modules or []
|
57 |
-
)
|
58 |
|
59 |
self.loss_config = LossConfig()
|
60 |
self.post_init()
|
@@ -141,6 +146,24 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
|
|
141 |
)
|
142 |
return {"loss": kl_loss}
|
143 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
144 |
def forward(
|
145 |
self,
|
146 |
input_ids: torch.Tensor,
|
@@ -149,8 +172,9 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
|
|
149 |
labels: Optional[torch.Tensor] = None,
|
150 |
attention_mask: Optional[torch.Tensor] = None,
|
151 |
audio_token_start_idx: Optional[torch.Tensor] = None,
|
152 |
-
|
153 |
audio_token_len: Optional[torch.Tensor] = None,
|
|
|
154 |
past_key_values: Optional[Union[Tuple, transformers.cache_utils.Cache]] = None,
|
155 |
# the alt_* fields are needed for KL divergence loss
|
156 |
alt_input_ids: Optional[torch.Tensor] = None,
|
@@ -181,29 +205,37 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
|
|
181 |
# B x T -> B x T x D
|
182 |
inputs_embeds = self.get_input_embeddings().forward(input_ids)
|
183 |
|
184 |
-
if audio_values is not None:
|
185 |
assert (
|
186 |
-
audio_token_start_idx is not None
|
187 |
-
|
|
|
|
|
|
|
188 |
assert (
|
189 |
-
len(audio_token_start_idx)
|
190 |
-
|
191 |
-
|
192 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
193 |
audio_tower_output = self.audio_tower.forward(
|
194 |
audio_values.to(self.audio_tower.dtype),
|
195 |
-
audio_len=
|
196 |
).last_hidden_state
|
197 |
audio_tower_output = audio_tower_output.to(inputs_embeds.dtype)
|
198 |
-
|
199 |
audio_embeds = self.multi_modal_projector.forward(audio_tower_output)
|
200 |
|
201 |
# combine audio and text embeddings
|
202 |
-
for
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
inputs_embeds[
|
207 |
|
208 |
lm_output = self.language_model.forward(
|
209 |
inputs_embeds=inputs_embeds,
|
@@ -238,7 +270,8 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
|
|
238 |
audio_values: Optional[torch.FloatTensor] = None,
|
239 |
audio_token_start_idx: Optional[torch.Tensor] = None,
|
240 |
audio_token_len: Optional[torch.Tensor] = None,
|
241 |
-
|
|
|
242 |
past_key_values: Optional[Union[Tuple, transformers.cache_utils.Cache]] = None,
|
243 |
attention_mask: Optional[torch.Tensor] = None,
|
244 |
inputs_embeds: Optional[torch.Tensor] = None,
|
@@ -267,7 +300,8 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
|
|
267 |
audio_token_start_idx - prefill_start_idx
|
268 |
)
|
269 |
model_input["audio_token_len"] = audio_token_len
|
270 |
-
model_input["
|
|
|
271 |
|
272 |
return model_input
|
273 |
|
@@ -284,7 +318,7 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
|
|
284 |
cls, config: UltravoxConfig
|
285 |
) -> Union[transformers.Wav2Vec2Model, "ModifiedWhisperEncoder"]:
|
286 |
if config.audio_model_id is not None:
|
287 |
-
if "whisper" in config.audio_model_id
|
288 |
audio_tower = ModifiedWhisperEncoder.from_pretrained(
|
289 |
config.audio_model_id, torch_dtype=config.torch_dtype
|
290 |
)
|
@@ -300,7 +334,7 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
|
|
300 |
config.audio_model_id, torch_dtype=config.torch_dtype
|
301 |
)
|
302 |
else:
|
303 |
-
if "whisper" in config.audio_config._name_or_path:
|
304 |
audio_tower = ModifiedWhisperEncoder(config.audio_config)
|
305 |
audio_tower.init_latency_mask(
|
306 |
config.audio_latency_block_size, dtype=config.torch_dtype
|
@@ -393,13 +427,17 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
|
|
393 |
if state_dict is None:
|
394 |
state_dict = super().state_dict()
|
395 |
|
396 |
-
|
|
|
|
|
|
|
|
|
|
|
397 |
|
398 |
state_dict = {
|
399 |
k: v
|
400 |
for k, v in state_dict.items()
|
401 |
-
if k in self.keep_params
|
402 |
-
or (k in named_params and named_params[k].requires_grad)
|
403 |
}
|
404 |
|
405 |
return state_dict
|
@@ -445,7 +483,7 @@ class UltravoxModel(transformers.LlamaPreTrainedModel):
|
|
445 |
|
446 |
# TODO: refactor common parts to a shared module
|
447 |
def is_cache_empty(
|
448 |
-
past_key_values: Optional[Union[Tuple, transformers.cache_utils.Cache]]
|
449 |
) -> bool:
|
450 |
"""
|
451 |
Check if the cache is empty.
|
@@ -481,12 +519,8 @@ def apply_lora(model: torch.nn.Module, lora_config: dict) -> torch.nn.Module:
|
|
481 |
|
482 |
class StackAudioFrames(nn.Module):
|
483 |
"""
|
484 |
-
Stack the audio embedding frames to reduce the sequence length by a factor
|
485 |
-
|
486 |
-
The number of output frames will be `ceil(T / stack_factor) + 1` where `T` is the number of input frames.
|
487 |
-
NOTE: the extra +1 is intentional: in case the number of audio tokens are over-estimated by the processor,
|
488 |
-
we want to make sure `processor.audio_token_replacement` (i.e. EOS) doesn't get leaked into the middle of embeddings.
|
489 |
-
In most cases this extra padding will get removed in the model's forward function so it has no effect.
|
490 |
"""
|
491 |
|
492 |
def __init__(self, stack_factor: int = 8):
|
@@ -496,7 +530,7 @@ class StackAudioFrames(nn.Module):
|
|
496 |
def forward(self, audio_embeds: torch.Tensor) -> torch.Tensor:
|
497 |
B, T, C = audio_embeds.shape
|
498 |
T_pad = (T + self.stack_factor - 1) // self.stack_factor * self.stack_factor
|
499 |
-
audio_embeds = F.pad(audio_embeds, (0, 0, 0, T_pad - T
|
500 |
B, T, C = audio_embeds.shape
|
501 |
audio_embeds = audio_embeds.view(
|
502 |
B, T // self.stack_factor, C * self.stack_factor
|
@@ -568,17 +602,25 @@ class ModifiedWhisperEncoder(
|
|
568 |
base_model_prefix = "model.encoder"
|
569 |
_no_split_modules = ["WhisperEncoderLayer"]
|
570 |
|
571 |
-
def
|
572 |
-
|
573 |
-
|
574 |
-
return
|
575 |
|
576 |
-
|
577 |
-
|
|
|
578 |
self.config.max_source_positions
|
579 |
* self.conv1.stride[0]
|
580 |
* self.conv2.stride[0]
|
581 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
582 |
assert (
|
583 |
max_seqlen > 0
|
584 |
), f"maximum sequence length must be positive, got {max_seqlen}"
|
@@ -610,11 +652,7 @@ class ModifiedWhisperEncoder(
|
|
610 |
output_hidden_states=None,
|
611 |
return_dict=None,
|
612 |
):
|
613 |
-
expected_seq_length =
|
614 |
-
self.config.max_source_positions
|
615 |
-
* self.conv1.stride[0]
|
616 |
-
* self.conv2.stride[0]
|
617 |
-
)
|
618 |
if input_features.shape[-1] > expected_seq_length:
|
619 |
raise ValueError(
|
620 |
f"Whisper expects the mel input features to be of length {expected_seq_length} or less, but found {input_features.shape[-1]}. Make sure to pad the input mel features to {expected_seq_length}."
|
@@ -665,7 +703,6 @@ class ModifiedWhisperEncoder(
|
|
665 |
attention_mask = self.get_extended_attention_mask(
|
666 |
attention_mask,
|
667 |
None,
|
668 |
-
device=hidden_states.device,
|
669 |
dtype=hidden_states.dtype,
|
670 |
)
|
671 |
|
|
|
1 |
import logging
|
2 |
import re
|
3 |
+
from typing import Any, Dict, Generator, Optional, Set, Tuple, Union
|
4 |
|
5 |
import peft
|
6 |
import torch
|
|
|
10 |
import transformers.activations
|
11 |
import transformers.modeling_outputs
|
12 |
import transformers.models
|
13 |
+
from transformers.generation.utils import GenerationMixin
|
14 |
from transformers.models.whisper import modeling_whisper as whisper
|
15 |
|
16 |
# We must use relative import in this directory to allow uploading to HF Hub
|
|
|
20 |
from .ultravox_config import UltravoxConfig
|
21 |
|
22 |
|
23 |
+
class UltravoxModel(transformers.LlamaPreTrainedModel, GenerationMixin):
|
24 |
"""
|
25 |
The Ultravox model which consists of an audio encoder and a language model.
|
26 |
|
|
|
38 |
config: UltravoxConfig # for type hinting
|
39 |
# Usually we load encoder and LLM weights from a pretrained model separately, so they are allowed to be missing
|
40 |
_keys_to_ignore_on_load_missing = ["audio_tower.*", "language_model.*"]
|
41 |
+
# Since we have kwargs in forward, we need to set this to False, otherwise grad_accum_steps will cause incorrect train loss to be reported
|
42 |
+
# see https://github.com/huggingface/transformers/issues/35856 and https://github.com/huggingface/trl/pull/2615/files
|
43 |
+
accepts_loss_kwargs = False
|
44 |
|
45 |
def __init__(self, config: UltravoxConfig):
|
46 |
super().__init__(config)
|
|
|
50 |
self.vocab_size = config.vocab_size
|
51 |
|
52 |
self.audio_tower = self._create_audio_tower(config)
|
53 |
+
self.audio_tower_context_length: Optional[int] = None
|
54 |
+
self.audio_tower_context_length = self.audio_tower.max_context_length
|
55 |
+
|
56 |
self.multi_modal_projector = self._create_multi_modal_projector(config)
|
57 |
self.language_model = self._create_language_model(config)
|
58 |
|
59 |
# Determine no_split_modules dynamically to use with FSDP auto_wrap policy.
|
60 |
# FSDP throws an error if some of the layer types are not found in the model.
|
61 |
+
# This would be something like ["LlamaDecoderLayer"] as we don't split audio encoder layers.
|
62 |
+
self._no_split_modules = self.language_model._no_split_modules
|
|
|
|
|
63 |
|
64 |
self.loss_config = LossConfig()
|
65 |
self.post_init()
|
|
|
146 |
)
|
147 |
return {"loss": kl_loss}
|
148 |
|
149 |
+
def _audio_iter(
|
150 |
+
self, audio_batch_size: torch.Tensor
|
151 |
+
) -> Generator[Tuple[int, int], None, None]:
|
152 |
+
"""
|
153 |
+
Iterate over the audio batch size and yield the batch index and audio index of each audio item.
|
154 |
+
|
155 |
+
Args:
|
156 |
+
audio_batch_size: A tensor of shape (B,) where B is the batch size.
|
157 |
+
|
158 |
+
Returns:
|
159 |
+
A generator that yields a tuple of (start index, length) for each audio item.
|
160 |
+
"""
|
161 |
+
audio_index = 0
|
162 |
+
for i_b, batch_count in enumerate(audio_batch_size):
|
163 |
+
for _ in range(batch_count):
|
164 |
+
yield i_b, audio_index
|
165 |
+
audio_index += 1
|
166 |
+
|
167 |
def forward(
|
168 |
self,
|
169 |
input_ids: torch.Tensor,
|
|
|
172 |
labels: Optional[torch.Tensor] = None,
|
173 |
attention_mask: Optional[torch.Tensor] = None,
|
174 |
audio_token_start_idx: Optional[torch.Tensor] = None,
|
175 |
+
audio_lens: Optional[torch.Tensor] = None,
|
176 |
audio_token_len: Optional[torch.Tensor] = None,
|
177 |
+
audio_batch_size: Optional[torch.Tensor] = None,
|
178 |
past_key_values: Optional[Union[Tuple, transformers.cache_utils.Cache]] = None,
|
179 |
# the alt_* fields are needed for KL divergence loss
|
180 |
alt_input_ids: Optional[torch.Tensor] = None,
|
|
|
205 |
# B x T -> B x T x D
|
206 |
inputs_embeds = self.get_input_embeddings().forward(input_ids)
|
207 |
|
208 |
+
if audio_values is not None and len(audio_values) > 0:
|
209 |
assert (
|
210 |
+
audio_token_start_idx is not None
|
211 |
+
and audio_token_len is not None
|
212 |
+
and audio_lens is not None
|
213 |
+
and audio_batch_size is not None
|
214 |
+
), "audio_token_start_idx/audio_token_len/audio_lens must be provided if audio_values are provided."
|
215 |
assert (
|
216 |
+
len(audio_token_start_idx)
|
217 |
+
== len(audio_token_len)
|
218 |
+
== len(audio_lens)
|
219 |
+
== len(audio_values)
|
220 |
+
), "audio_token_start_idx/audio_token_len/audio_lens/audio_values must have the same batch size."
|
221 |
+
assert len(audio_batch_size) == len(
|
222 |
+
inputs_embeds
|
223 |
+
), "audio_batch_size and inputs_embeds must have the same batch size."
|
224 |
+
|
225 |
+
# B x A/3200 x (D=max-audio-length-in-batch)
|
226 |
audio_tower_output = self.audio_tower.forward(
|
227 |
audio_values.to(self.audio_tower.dtype),
|
228 |
+
audio_len=audio_lens,
|
229 |
).last_hidden_state
|
230 |
audio_tower_output = audio_tower_output.to(inputs_embeds.dtype)
|
|
|
231 |
audio_embeds = self.multi_modal_projector.forward(audio_tower_output)
|
232 |
|
233 |
# combine audio and text embeddings
|
234 |
+
for i_b, i_a in self._audio_iter(audio_batch_size):
|
235 |
+
start_idx = audio_token_start_idx[i_a]
|
236 |
+
token_len = audio_token_len[i_a]
|
237 |
+
item_embedding = audio_embeds[i_a][:token_len]
|
238 |
+
inputs_embeds[i_b][start_idx : start_idx + token_len] = item_embedding
|
239 |
|
240 |
lm_output = self.language_model.forward(
|
241 |
inputs_embeds=inputs_embeds,
|
|
|
270 |
audio_values: Optional[torch.FloatTensor] = None,
|
271 |
audio_token_start_idx: Optional[torch.Tensor] = None,
|
272 |
audio_token_len: Optional[torch.Tensor] = None,
|
273 |
+
audio_lens: Optional[torch.Tensor] = None,
|
274 |
+
audio_batch_size: Optional[torch.Tensor] = None,
|
275 |
past_key_values: Optional[Union[Tuple, transformers.cache_utils.Cache]] = None,
|
276 |
attention_mask: Optional[torch.Tensor] = None,
|
277 |
inputs_embeds: Optional[torch.Tensor] = None,
|
|
|
300 |
audio_token_start_idx - prefill_start_idx
|
301 |
)
|
302 |
model_input["audio_token_len"] = audio_token_len
|
303 |
+
model_input["audio_batch_size"] = audio_batch_size
|
304 |
+
model_input["audio_lens"] = audio_lens
|
305 |
|
306 |
return model_input
|
307 |
|
|
|
318 |
cls, config: UltravoxConfig
|
319 |
) -> Union[transformers.Wav2Vec2Model, "ModifiedWhisperEncoder"]:
|
320 |
if config.audio_model_id is not None:
|
321 |
+
if "whisper" in config.audio_model_id.lower():
|
322 |
audio_tower = ModifiedWhisperEncoder.from_pretrained(
|
323 |
config.audio_model_id, torch_dtype=config.torch_dtype
|
324 |
)
|
|
|
334 |
config.audio_model_id, torch_dtype=config.torch_dtype
|
335 |
)
|
336 |
else:
|
337 |
+
if "whisper" in config.audio_config._name_or_path.lower():
|
338 |
audio_tower = ModifiedWhisperEncoder(config.audio_config)
|
339 |
audio_tower.init_latency_mask(
|
340 |
config.audio_latency_block_size, dtype=config.torch_dtype
|
|
|
427 |
if state_dict is None:
|
428 |
state_dict = super().state_dict()
|
429 |
|
430 |
+
trainable_params = {k for k, v in self.named_parameters() if v.requires_grad}
|
431 |
+
# normalize the keys to match the original model
|
432 |
+
# Example: audio_tower.base_model.model.layers.0._fsdp_wrapped_module.self_attn.k_proj.lora_B.default.weight
|
433 |
+
trainable_params = {
|
434 |
+
k.replace("_fsdp_wrapped_module.", "") for k in trainable_params
|
435 |
+
}
|
436 |
|
437 |
state_dict = {
|
438 |
k: v
|
439 |
for k, v in state_dict.items()
|
440 |
+
if k in self.keep_params or k in trainable_params
|
|
|
441 |
}
|
442 |
|
443 |
return state_dict
|
|
|
483 |
|
484 |
# TODO: refactor common parts to a shared module
|
485 |
def is_cache_empty(
|
486 |
+
past_key_values: Optional[Union[Tuple, transformers.cache_utils.Cache]],
|
487 |
) -> bool:
|
488 |
"""
|
489 |
Check if the cache is empty.
|
|
|
519 |
|
520 |
class StackAudioFrames(nn.Module):
|
521 |
"""
|
522 |
+
Stack the audio embedding frames to reduce the sequence length by a factor
|
523 |
+
of `stack_factor`.
|
|
|
|
|
|
|
|
|
524 |
"""
|
525 |
|
526 |
def __init__(self, stack_factor: int = 8):
|
|
|
530 |
def forward(self, audio_embeds: torch.Tensor) -> torch.Tensor:
|
531 |
B, T, C = audio_embeds.shape
|
532 |
T_pad = (T + self.stack_factor - 1) // self.stack_factor * self.stack_factor
|
533 |
+
audio_embeds = F.pad(audio_embeds, (0, 0, 0, T_pad - T))
|
534 |
B, T, C = audio_embeds.shape
|
535 |
audio_embeds = audio_embeds.view(
|
536 |
B, T // self.stack_factor, C * self.stack_factor
|
|
|
602 |
base_model_prefix = "model.encoder"
|
603 |
_no_split_modules = ["WhisperEncoderLayer"]
|
604 |
|
605 |
+
def __init__(self, config: transformers.WhisperConfig):
|
606 |
+
super().__init__(config)
|
607 |
+
self.config.is_decoder = False
|
|
|
608 |
|
609 |
+
@property
|
610 |
+
def max_context_length(self):
|
611 |
+
return (
|
612 |
self.config.max_source_positions
|
613 |
* self.conv1.stride[0]
|
614 |
* self.conv2.stride[0]
|
615 |
)
|
616 |
+
|
617 |
+
def init_latency_mask(self, audio_latency_block_size: int, dtype: torch.dtype):
|
618 |
+
if audio_latency_block_size is None:
|
619 |
+
self.audio_streaming_mask = None
|
620 |
+
return
|
621 |
+
|
622 |
+
# Use max_context_length directly in the calculation
|
623 |
+
max_seqlen = self.max_context_length
|
624 |
assert (
|
625 |
max_seqlen > 0
|
626 |
), f"maximum sequence length must be positive, got {max_seqlen}"
|
|
|
652 |
output_hidden_states=None,
|
653 |
return_dict=None,
|
654 |
):
|
655 |
+
expected_seq_length = self.max_context_length
|
|
|
|
|
|
|
|
|
656 |
if input_features.shape[-1] > expected_seq_length:
|
657 |
raise ValueError(
|
658 |
f"Whisper expects the mel input features to be of length {expected_seq_length} or less, but found {input_features.shape[-1]}. Make sure to pad the input mel features to {expected_seq_length}."
|
|
|
703 |
attention_mask = self.get_extended_attention_mask(
|
704 |
attention_mask,
|
705 |
None,
|
|
|
706 |
dtype=hidden_states.dtype,
|
707 |
)
|
708 |
|
ultravox_processing.py
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
import dataclasses
|
2 |
-
from typing import Optional, Union
|
3 |
|
4 |
import numpy as np
|
5 |
import torch
|
@@ -15,7 +15,13 @@ class DataCollatorForSeq2SeqWithAudio(transformers.DataCollatorForSeq2Seq):
|
|
15 |
include_alt_fields: bool = False
|
16 |
|
17 |
def __call__(self, features, *args, **kwargs):
|
18 |
-
audio_values = [f.pop("audio_values",
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
if self.include_alt_fields:
|
20 |
# these fields are hard-coded in the transformer data collator, so they need special handling before calling the super method
|
21 |
alt_features = [
|
@@ -34,8 +40,12 @@ class DataCollatorForSeq2SeqWithAudio(transformers.DataCollatorForSeq2Seq):
|
|
34 |
batch["alt_attention_mask"] = alt_batch["attention_mask"]
|
35 |
batch["alt_labels"] = alt_batch["labels"]
|
36 |
|
|
|
|
|
|
|
|
|
37 |
# Pad the last dimension of all audio_values to the same length, with 0s on the right.
|
38 |
-
if audio_values
|
39 |
max_len = max([x.shape[-1] for x in audio_values])
|
40 |
batch["audio_values"] = torch.stack(
|
41 |
[F.pad(x, (0, max_len - x.shape[-1])) for x in audio_values]
|
@@ -45,10 +55,12 @@ class DataCollatorForSeq2SeqWithAudio(transformers.DataCollatorForSeq2Seq):
|
|
45 |
[f["input_ids"].shape[-1] for f in features]
|
46 |
)
|
47 |
displacement = batch["input_ids"].shape[-1] - input_ids_lens
|
|
|
|
|
|
|
48 |
batch["audio_token_start_idx"] += displacement.to(
|
49 |
batch["audio_token_start_idx"].device
|
50 |
)
|
51 |
-
|
52 |
return batch
|
53 |
|
54 |
|
@@ -62,11 +74,7 @@ class UltravoxProcessor(transformers.ProcessorMixin):
|
|
62 |
"""
|
63 |
|
64 |
attributes = ["audio_processor", "tokenizer"]
|
65 |
-
audio_processor_class = (
|
66 |
-
"Wav2Vec2Processor",
|
67 |
-
"SeamlessM4TFeatureExtractor",
|
68 |
-
"WhisperProcessor",
|
69 |
-
)
|
70 |
tokenizer_class = (
|
71 |
"PreTrainedTokenizer",
|
72 |
"PreTrainedTokenizerFast",
|
@@ -80,27 +88,32 @@ class UltravoxProcessor(transformers.ProcessorMixin):
|
|
80 |
audio_processor=None,
|
81 |
tokenizer=None,
|
82 |
audio_padding: str = "longest",
|
83 |
-
encoder_ds_factor: int =
|
84 |
stack_factor: int = 8,
|
85 |
audio_placeholder: str = "<|audio|>",
|
|
|
|
|
86 |
):
|
87 |
"""
|
88 |
Args:
|
89 |
audio_processor: The audio processor for the audio encoder.
|
90 |
tokenizer: The tokenizer for the language model.
|
91 |
audio_padding: The padding strategy for the audio encoder.
|
92 |
-
encoder_ds_factor: The downsample factor of the audio encoder.
|
93 |
stack_factor: The factor by which the audio encoder output is stacked in the multimodal projector.
|
|
|
94 |
audio_placeholder: The placeholder for the audio in the text.
|
|
|
95 |
"""
|
96 |
self.audio_padding = audio_padding
|
97 |
self.encoder_ds_factor = encoder_ds_factor
|
98 |
self.stack_factor = stack_factor
|
99 |
self.audio_placeholder = audio_placeholder
|
100 |
-
self.
|
101 |
assert (
|
102 |
-
|
103 |
), "The tokenizer has no EOS token. Cannot recover."
|
|
|
|
|
104 |
if tokenizer.pad_token_id is None:
|
105 |
tokenizer.pad_token_id = tokenizer.eos_token_id
|
106 |
|
@@ -114,7 +127,7 @@ class UltravoxProcessor(transformers.ProcessorMixin):
|
|
114 |
audio_processor = transformers.AutoProcessor.from_pretrained(
|
115 |
config.audio_model_id
|
116 |
or config.audio_config._name_or_path
|
117 |
-
or "
|
118 |
)
|
119 |
|
120 |
tokenizer = transformers.AutoTokenizer.from_pretrained(
|
@@ -129,30 +142,100 @@ class UltravoxProcessor(transformers.ProcessorMixin):
|
|
129 |
stack_factor=config.stack_factor,
|
130 |
)
|
131 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
132 |
def __call__(
|
133 |
self,
|
134 |
text: Optional[str] = None,
|
135 |
audio: Optional[Union[np.ndarray, torch.Tensor]] = None,
|
|
|
|
|
|
|
|
|
|
|
136 |
sampling_rate: Optional[int] = None,
|
137 |
return_tensors: Optional[
|
138 |
Union[str, transformers.TensorType]
|
139 |
] = transformers.TensorType.PYTORCH,
|
|
|
140 |
**kwargs,
|
141 |
) -> transformers.BatchFeature:
|
142 |
"""
|
143 |
Main method to prepare for the model one text sequence and audio. This method forwards the `text`
|
144 |
and `kwargs` arguments to PreTrainedTokenizerFast's [`~PreTrainedTokenizerFast.__call__`] if `text` is not `None` to encode
|
145 |
the text. To prepare the audio(s), this method forwards the `audio`, `sampling_rate` and `kwargs` arguments to
|
146 |
-
audio processor's [`~
|
147 |
of the above two methods for more information.
|
148 |
|
149 |
Args:
|
150 |
text (`str`, `List[str]`):
|
151 |
The sequence to be encoded. Sequence can be a string or (pretokenized string).
|
152 |
audio (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, `List[torch.Tensor]`):
|
153 |
-
The audio to be prepared. Audio can be NumPy array or PyTorch tensor.
|
154 |
-
|
155 |
-
|
156 |
sampling_rate (`int`, *optional*, defaults to 16000):
|
157 |
Sampling rate of the input audio. We expect 16kHz audio. Don't change this value unless you know what
|
158 |
you are doing.
|
@@ -176,75 +259,105 @@ class UltravoxProcessor(transformers.ProcessorMixin):
|
|
176 |
Returned when `audio` is not `None`.
|
177 |
- **audio_token_start_idx** -- The index in the tokenized text where the audio starts. Returned when `audio` is not `None`.
|
178 |
"""
|
179 |
-
# TODO: Add support for multiple
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
180 |
data = {}
|
181 |
-
|
182 |
-
if
|
183 |
-
if
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
|
196 |
# Main audio processing. The processor is model-specific.
|
197 |
-
x = self.audio_processor(
|
198 |
-
|
199 |
sampling_rate=sampling_rate,
|
200 |
padding="longest",
|
201 |
-
|
|
|
202 |
return_attention_mask=True,
|
203 |
**kwargs,
|
204 |
)
|
205 |
-
if "input_features" in x:
|
206 |
-
data["audio_values"] = x.input_features
|
207 |
-
else:
|
208 |
-
data["audio_values"] = x.input_values
|
209 |
-
|
210 |
-
# data["audio_len"] is the number of frames in the audio, used for creating attention masks in whisper encoder
|
211 |
-
if (
|
212 |
-
self.audio_padding == "max_length"
|
213 |
-
): # audio is padded to max length, so we rely on the attention mask to determine audio_len
|
214 |
-
data["audio_len"] = (
|
215 |
-
x.attention_mask.sum(-1) - 1
|
216 |
-
) # Whisper attention mask includes an extra 1 at the end that needs to be subtracted
|
217 |
-
else: # audio is not padded, so we can directly use the audio length
|
218 |
-
data["audio_len"] = [torch.as_tensor(data["audio_values"]).shape[-1]]
|
219 |
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
224 |
-
|
225 |
-
|
226 |
-
|
227 |
-
f"audio must be provided when using audio placeholder ({self.audio_placeholder}) in text."
|
228 |
-
)
|
229 |
-
|
230 |
-
start_idx = len(
|
231 |
-
self.tokenizer.encode(
|
232 |
-
text[: text.index(self.audio_placeholder)],
|
233 |
-
add_special_tokens=False,
|
234 |
-
)
|
235 |
-
)
|
236 |
-
data["audio_token_start_idx"] = [start_idx]
|
237 |
-
|
238 |
-
# Replace the audio placeholder with the audio token.
|
239 |
-
# e.g. "Transcribe\n<|audio|>" -> "Transcribe </s></s></s></s></s></s></s></s>"
|
240 |
-
# where the number of </s> is the number of audio frames.
|
241 |
-
text = text.replace(
|
242 |
-
self.audio_placeholder,
|
243 |
-
self.audio_token_replacement * audio_embed_frames,
|
244 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
245 |
|
246 |
# Special tokens like BOS should already have been added by the caller.
|
247 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
248 |
|
249 |
return transformers.BatchFeature(data=data, tensor_type=return_tensors)
|
250 |
|
|
|
1 |
import dataclasses
|
2 |
+
from typing import Any, Dict, List, Optional, Union
|
3 |
|
4 |
import numpy as np
|
5 |
import torch
|
|
|
15 |
include_alt_fields: bool = False
|
16 |
|
17 |
def __call__(self, features, *args, **kwargs):
|
18 |
+
audio_values = [x for f in features for x in f.pop("audio_values", [])]
|
19 |
+
audio_lens = [x for f in features for x in f.pop("audio_lens", [])]
|
20 |
+
audio_token_len = [x for f in features for x in f.pop("audio_token_len", [])]
|
21 |
+
audio_token_start_idx = [
|
22 |
+
x for f in features for x in f.pop("audio_token_start_idx", [])
|
23 |
+
]
|
24 |
+
|
25 |
if self.include_alt_fields:
|
26 |
# these fields are hard-coded in the transformer data collator, so they need special handling before calling the super method
|
27 |
alt_features = [
|
|
|
40 |
batch["alt_attention_mask"] = alt_batch["attention_mask"]
|
41 |
batch["alt_labels"] = alt_batch["labels"]
|
42 |
|
43 |
+
batch["audio_token_start_idx"] = torch.stack(audio_token_start_idx)
|
44 |
+
batch["audio_lens"] = torch.stack(audio_lens)
|
45 |
+
batch["audio_token_len"] = torch.stack(audio_token_len)
|
46 |
+
|
47 |
# Pad the last dimension of all audio_values to the same length, with 0s on the right.
|
48 |
+
if audio_values:
|
49 |
max_len = max([x.shape[-1] for x in audio_values])
|
50 |
batch["audio_values"] = torch.stack(
|
51 |
[F.pad(x, (0, max_len - x.shape[-1])) for x in audio_values]
|
|
|
55 |
[f["input_ids"].shape[-1] for f in features]
|
56 |
)
|
57 |
displacement = batch["input_ids"].shape[-1] - input_ids_lens
|
58 |
+
displacement = displacement.repeat_interleave(
|
59 |
+
batch["audio_batch_size"].squeeze(-1)
|
60 |
+
)
|
61 |
batch["audio_token_start_idx"] += displacement.to(
|
62 |
batch["audio_token_start_idx"].device
|
63 |
)
|
|
|
64 |
return batch
|
65 |
|
66 |
|
|
|
74 |
"""
|
75 |
|
76 |
attributes = ["audio_processor", "tokenizer"]
|
77 |
+
audio_processor_class = ("WhisperProcessor",)
|
|
|
|
|
|
|
|
|
78 |
tokenizer_class = (
|
79 |
"PreTrainedTokenizer",
|
80 |
"PreTrainedTokenizerFast",
|
|
|
88 |
audio_processor=None,
|
89 |
tokenizer=None,
|
90 |
audio_padding: str = "longest",
|
91 |
+
encoder_ds_factor: int = 2,
|
92 |
stack_factor: int = 8,
|
93 |
audio_placeholder: str = "<|audio|>",
|
94 |
+
# Defaults to whisper encoder context size
|
95 |
+
audio_context_size: Optional[int] = 3000,
|
96 |
):
|
97 |
"""
|
98 |
Args:
|
99 |
audio_processor: The audio processor for the audio encoder.
|
100 |
tokenizer: The tokenizer for the language model.
|
101 |
audio_padding: The padding strategy for the audio encoder.
|
|
|
102 |
stack_factor: The factor by which the audio encoder output is stacked in the multimodal projector.
|
103 |
+
encoder_ds_factor: The downsampling factor of the audio encoder.
|
104 |
audio_placeholder: The placeholder for the audio in the text.
|
105 |
+
audio_context_size: The maximum number of frames that the audio encoder can handle.
|
106 |
"""
|
107 |
self.audio_padding = audio_padding
|
108 |
self.encoder_ds_factor = encoder_ds_factor
|
109 |
self.stack_factor = stack_factor
|
110 |
self.audio_placeholder = audio_placeholder
|
111 |
+
self.audio_context_size = audio_context_size
|
112 |
assert (
|
113 |
+
tokenizer.eos_token is not None
|
114 |
), "The tokenizer has no EOS token. Cannot recover."
|
115 |
+
self.vocab = tokenizer.get_vocab()
|
116 |
+
self.audio_token_replacement = tokenizer.eos_token
|
117 |
if tokenizer.pad_token_id is None:
|
118 |
tokenizer.pad_token_id = tokenizer.eos_token_id
|
119 |
|
|
|
127 |
audio_processor = transformers.AutoProcessor.from_pretrained(
|
128 |
config.audio_model_id
|
129 |
or config.audio_config._name_or_path
|
130 |
+
or "openai/whisper-tiny"
|
131 |
)
|
132 |
|
133 |
tokenizer = transformers.AutoTokenizer.from_pretrained(
|
|
|
142 |
stack_factor=config.stack_factor,
|
143 |
)
|
144 |
|
145 |
+
def _chunk_and_pad_audio(
|
146 |
+
self,
|
147 |
+
audio_values: torch.Tensor,
|
148 |
+
audio_lens: torch.Tensor,
|
149 |
+
include_audio_num_chunks: bool = False,
|
150 |
+
) -> Dict[str, Any]:
|
151 |
+
"""
|
152 |
+
Processes the audio batch by chunking any items in the batch according to the audio_context_size,
|
153 |
+
padding the last chunk if needed, and returns a dictionary with updated audio data.
|
154 |
+
|
155 |
+
Args:
|
156 |
+
audio_values (torch.Tensor): A tensor of audio values (e.g., in B, D, T format).
|
157 |
+
audio_lens (torch.Tensor): A tensor of audio lengths.
|
158 |
+
|
159 |
+
Returns:
|
160 |
+
Dict[str, Any]: Dictionary with the following keys:
|
161 |
+
- "audio_values": The concatenated audio tensor after chunking and padding.
|
162 |
+
- "audio_lens": Tensor of lengths for each chunk.
|
163 |
+
- "audio_is_continuation": Tensor of booleans indicating if the chunk is a continuation of the previous chunk.
|
164 |
+
- "audio_batch_size": A Tensor with one integer representing the number of chunks.
|
165 |
+
|
166 |
+
"""
|
167 |
+
chunked_audio_values: List[torch.Tensor] = []
|
168 |
+
chunked_audio_lens: List[int] = []
|
169 |
+
is_continuation_list: List[bool] = []
|
170 |
+
num_chunks: List[int] = []
|
171 |
+
context_size = self.audio_context_size or audio_values.shape[-1]
|
172 |
+
|
173 |
+
for i in range(audio_values.shape[0]): # iterate over the batch
|
174 |
+
num_chunks.append(int(np.ceil(audio_lens[i] / context_size)))
|
175 |
+
for offset in range(0, audio_lens[i], context_size):
|
176 |
+
is_continuation = offset > 0
|
177 |
+
chunk = audio_values[i, :, offset : offset + context_size]
|
178 |
+
if is_continuation and chunk.shape[-1] < context_size:
|
179 |
+
# N.B. We only need to pad continuation chunks. If none of the samples require chunking, the
|
180 |
+
# batch might not (need to) be padded all the way to the audio_context_size, in which case
|
181 |
+
# we've already included the padding above. On the other hand, if we have any continuation
|
182 |
+
# chunks we know that the batch needs to be padded to audio_context_size because that's what
|
183 |
+
# we're slicing to.
|
184 |
+
chunk = F.pad(chunk, (0, context_size - chunk.shape[-1]))
|
185 |
+
chunked_audio_values.append(chunk)
|
186 |
+
chunked_audio_lens.append(
|
187 |
+
min(int(audio_lens[i].item()) - offset, context_size)
|
188 |
+
)
|
189 |
+
is_continuation_list.append(is_continuation)
|
190 |
+
|
191 |
+
data = {
|
192 |
+
"audio_values": torch.stack(chunked_audio_values, dim=0),
|
193 |
+
"audio_lens": torch.tensor(
|
194 |
+
chunked_audio_lens, dtype=torch.int64, device=audio_values.device
|
195 |
+
),
|
196 |
+
"audio_is_continuation": torch.tensor(
|
197 |
+
is_continuation_list, dtype=torch.bool, device=audio_values.device
|
198 |
+
),
|
199 |
+
"audio_batch_size": torch.tensor(
|
200 |
+
[len(chunked_audio_values)], device=audio_values.device
|
201 |
+
),
|
202 |
+
}
|
203 |
+
if include_audio_num_chunks:
|
204 |
+
data["audio_num_chunks"] = torch.tensor(
|
205 |
+
num_chunks, dtype=torch.int64, device=audio_values.device
|
206 |
+
)
|
207 |
+
return data
|
208 |
+
|
209 |
def __call__(
|
210 |
self,
|
211 |
text: Optional[str] = None,
|
212 |
audio: Optional[Union[np.ndarray, torch.Tensor]] = None,
|
213 |
+
audios: Optional[
|
214 |
+
Union[
|
215 |
+
List[Union[np.ndarray, torch.Tensor]], Union[np.ndarray, torch.Tensor]
|
216 |
+
]
|
217 |
+
] = None,
|
218 |
sampling_rate: Optional[int] = None,
|
219 |
return_tensors: Optional[
|
220 |
Union[str, transformers.TensorType]
|
221 |
] = transformers.TensorType.PYTORCH,
|
222 |
+
include_audio_num_chunks: bool = False,
|
223 |
**kwargs,
|
224 |
) -> transformers.BatchFeature:
|
225 |
"""
|
226 |
Main method to prepare for the model one text sequence and audio. This method forwards the `text`
|
227 |
and `kwargs` arguments to PreTrainedTokenizerFast's [`~PreTrainedTokenizerFast.__call__`] if `text` is not `None` to encode
|
228 |
the text. To prepare the audio(s), this method forwards the `audio`, `sampling_rate` and `kwargs` arguments to
|
229 |
+
audio processor's [`~WhisperProcessor.__call__`] if `audio` is not `None`. Please refer to the docstring
|
230 |
of the above two methods for more information.
|
231 |
|
232 |
Args:
|
233 |
text (`str`, `List[str]`):
|
234 |
The sequence to be encoded. Sequence can be a string or (pretokenized string).
|
235 |
audio (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, `List[torch.Tensor]`):
|
236 |
+
The audio to be prepared. Audio can be a single-channel (1-dimensional) NumPy array or PyTorch tensor.
|
237 |
+
audios (`np.ndarray`, `torch.Tensor`, `List[np.ndarray]`, `List[torch.Tensor]`):
|
238 |
+
A list or two dimensional array of audio to be prepared.
|
239 |
sampling_rate (`int`, *optional*, defaults to 16000):
|
240 |
Sampling rate of the input audio. We expect 16kHz audio. Don't change this value unless you know what
|
241 |
you are doing.
|
|
|
259 |
Returned when `audio` is not `None`.
|
260 |
- **audio_token_start_idx** -- The index in the tokenized text where the audio starts. Returned when `audio` is not `None`.
|
261 |
"""
|
262 |
+
# TODO: Add support for multiple text inputs.
|
263 |
+
if audio is not None and audios is not None:
|
264 |
+
raise ValueError("Only one of `audio` or `audios` should be provided.")
|
265 |
+
elif audio is not None:
|
266 |
+
audios = audio if isinstance(audio, list) or audio.ndim == 2 else [audio]
|
267 |
+
elif audios is None:
|
268 |
+
audios = []
|
269 |
+
|
270 |
data = {}
|
271 |
+
audio_is_continuation = []
|
272 |
+
if len(audios) > 0:
|
273 |
+
audios = [x.numpy() if isinstance(x, torch.Tensor) else x for x in audios]
|
274 |
+
|
275 |
+
# Pad out each audio to at least 2 hops (the minimum required by the processor).
|
276 |
+
hop_length = self.audio_processor.feature_extractor.hop_length
|
277 |
+
audios = [
|
278 |
+
(
|
279 |
+
np.pad(x, (0, 2 * hop_length - len(x)), mode="constant")
|
280 |
+
if len(x) < 2 * hop_length
|
281 |
+
else x
|
282 |
+
)
|
283 |
+
for x in audios
|
284 |
+
]
|
285 |
|
286 |
# Main audio processing. The processor is model-specific.
|
287 |
+
x: transformers.BatchFeature = self.audio_processor(
|
288 |
+
audios,
|
289 |
sampling_rate=sampling_rate,
|
290 |
padding="longest",
|
291 |
+
pad_to_multiple_of=hop_length, # The attention mask effectively gets padded to the hop length, so pad the audio to be consistent.
|
292 |
+
truncation=False,
|
293 |
return_attention_mask=True,
|
294 |
**kwargs,
|
295 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
296 |
|
297 |
+
data.update(
|
298 |
+
self._chunk_and_pad_audio(
|
299 |
+
audio_values=torch.as_tensor(
|
300 |
+
x.input_features if "input_features" in x else x.input_values
|
301 |
+
),
|
302 |
+
audio_lens=torch.as_tensor(x.attention_mask).sum(-1),
|
303 |
+
include_audio_num_chunks=include_audio_num_chunks,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
304 |
)
|
305 |
+
)
|
306 |
+
|
307 |
+
audio_is_continuation = data.pop("audio_is_continuation")
|
308 |
+
data["audio_token_len"] = torch.ceil(
|
309 |
+
data["audio_lens"] / (self.encoder_ds_factor * self.stack_factor)
|
310 |
+
).to(dtype=torch.int)
|
311 |
+
|
312 |
+
if text is not None:
|
313 |
+
if not isinstance(text, str):
|
314 |
+
raise ValueError("Text must be a string. Batch mode not supported yet.")
|
315 |
|
316 |
# Special tokens like BOS should already have been added by the caller.
|
317 |
+
tokenized_parts = self.tokenizer(
|
318 |
+
text.split(
|
319 |
+
"<|audio|>" # The placeholder isn't part of the vocabulary, so split the text around it.
|
320 |
+
),
|
321 |
+
add_special_tokens=False,
|
322 |
+
**kwargs,
|
323 |
+
)
|
324 |
+
|
325 |
+
audio_token_start_idx = []
|
326 |
+
placeholder_index = -1
|
327 |
+
split_input_ids = tokenized_parts["input_ids"]
|
328 |
+
input_ids: List[int] = []
|
329 |
+
|
330 |
+
audio_token_replacement_token_id = self.vocab[self.audio_token_replacement]
|
331 |
+
|
332 |
+
for i, token_len in enumerate(data.get("audio_token_len", [])):
|
333 |
+
if not audio_is_continuation[i]:
|
334 |
+
placeholder_index += 1
|
335 |
+
if placeholder_index >= len(split_input_ids):
|
336 |
+
raise ValueError(
|
337 |
+
f"Text contains too few audio placeholders. (Expected {len(audios)} placeholders)"
|
338 |
+
)
|
339 |
+
|
340 |
+
input_ids.extend(split_input_ids[placeholder_index])
|
341 |
+
|
342 |
+
audio_token_start_idx.append(len(input_ids))
|
343 |
+
|
344 |
+
input_ids.extend([audio_token_replacement_token_id] * token_len)
|
345 |
+
|
346 |
+
# Include any tokens after the last audio.
|
347 |
+
placeholder_index += 1
|
348 |
+
if placeholder_index != len(split_input_ids) - 1:
|
349 |
+
raise ValueError(
|
350 |
+
f"Text contains too many audio placeholders. (Expected {len(audios)} placeholders)"
|
351 |
+
)
|
352 |
+
input_ids.extend(split_input_ids[placeholder_index])
|
353 |
+
|
354 |
+
if "audio_token_len" in data:
|
355 |
+
data["audio_token_start_idx"] = torch.as_tensor(audio_token_start_idx)
|
356 |
+
|
357 |
+
data["input_ids"] = [input_ids]
|
358 |
+
data["attention_mask"] = [[1] * len(input_ids)]
|
359 |
+
|
360 |
+
# Ensure that there are no audio placeholders after the last audio.
|
361 |
|
362 |
return transformers.BatchFeature(data=data, tensor_type=return_tensors)
|
363 |
|