DionTimmer
commited on
Upload 2 files
Browse files- README.md +77 -1
- model_class.py +202 -0
README.md
CHANGED
@@ -1,3 +1,79 @@
|
|
1 |
---
|
2 |
-
license:
|
|
|
|
|
|
|
3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
---
|
2 |
+
license: cc-by-nc-4.0
|
3 |
+
language:
|
4 |
+
- en
|
5 |
+
library_name: transformers
|
6 |
---
|
7 |
+
# Whisper Multitask Analyzer
|
8 |
+
|
9 |
+
A transformer encoder-decoder model for automatic audio captioning. As opposed to speech-to-text, captioning describes the content and features of audio clips.
|
10 |
+
|
11 |
+
- **Model, codebase & card adapted from:** MU-NLPC/whisper-small-audio-captioning
|
12 |
+
- **Model type:** Whisper encoder-decoder transformer
|
13 |
+
- **Language(s) (NLP):** en
|
14 |
+
- **License:** cc-by-4.0
|
15 |
+
- **Parent Model:** openai/whisper-small
|
16 |
+
|
17 |
+
## Usage
|
18 |
+
|
19 |
+
The model expects an audio clip (up to 30s) to the encoder as an input and information about caption style as forced prefix to the decoder.
|
20 |
+
The forced prefix is an integer which is mapped to various tasks. This mapping is defined in the model config and can be retrieved with a function.
|
21 |
+
|
22 |
+
The tag mapping of the current model is:
|
23 |
+
|
24 |
+
| Task | ID | Description |
|
25 |
+
| -------- | -- | ------------------------------------------------------ |
|
26 |
+
| tags | 0 | General descriptions, can include genres and features. |
|
27 |
+
| genre | 1 | Estimated musical genres. |
|
28 |
+
| mood | 2 | Estimated emotional feeling. |
|
29 |
+
| movement | 3 | Estimated audio pace and expression. |
|
30 |
+
| theme | 4 | Estimated audio usage (not very accurate) |
|
31 |
+
|
32 |
+
```
|
33 |
+
|
34 |
+
Minimal example:
|
35 |
+
|
36 |
+
```python
|
37 |
+
# Load model
|
38 |
+
checkpoint = "DionTimmer/whisper-small-multitask-analyzer"
|
39 |
+
model = WhisperForAudioCaptioning.from_pretrained(checkpoint)
|
40 |
+
tokenizer = transformers.WhisperTokenizer.from_pretrained(checkpoint, language="en", task="transcribe")
|
41 |
+
feature_extractor = transformers.WhisperFeatureExtractor.from_pretrained(checkpoint)
|
42 |
+
|
43 |
+
# Load and preprocess audio
|
44 |
+
input_file = "..."
|
45 |
+
audio, sampling_rate = librosa.load(input_file, sr=feature_extractor.sampling_rate)
|
46 |
+
features = feature_extractor(audio, sampling_rate=sampling_rate, return_tensors="pt").input_features
|
47 |
+
|
48 |
+
# Mappings by ID
|
49 |
+
print(model.task_mapping) # {0: 'tags', 1: 'genre', 2: 'mood', 3: 'movement', 4: 'theme'}
|
50 |
+
|
51 |
+
# Inverted
|
52 |
+
print(model.named_task_mapping) # {'tags': 0, 'genre': 1, 'mood': 2, 'movement': 3, 'theme': 4}
|
53 |
+
|
54 |
+
# Prepare caption style
|
55 |
+
style_prefix = f"{model.named_task_mapping['tags']}: "
|
56 |
+
style_prefix_tokens = tokenizer("", text_target=style_prefix, return_tensors="pt", add_special_tokens=False).labels
|
57 |
+
|
58 |
+
# Generate caption
|
59 |
+
model.eval()
|
60 |
+
outputs = model.generate(
|
61 |
+
inputs=features.to(model.device),
|
62 |
+
forced_ac_decoder_ids=style_prefix_tokens,
|
63 |
+
max_length=100,
|
64 |
+
)
|
65 |
+
|
66 |
+
print(tokenizer.batch_decode(outputs, skip_special_tokens=True)[0])
|
67 |
+
```
|
68 |
+
|
69 |
+
Example output:
|
70 |
+
*0: advertising, beautiful, beauty, bright, cinematic, commercial, corporate, emotional, epic, film, heroic, hopeful, inspiration, inspirational, inspiring, love, love story, movie, orchestra, orchestral, piano, positive, presentation, romantic, sentimental*
|
71 |
+
|
72 |
+
WhisperTokenizer must be initialized with `language="en"` and `task="transcribe"`.
|
73 |
+
|
74 |
+
The model class `WhisperForAudioCaptioning` can be found in the git repository or here on the HuggingFace Hub in the model repository. The class overrides default Whisper `generate` method to support forcing decoder prefix.
|
75 |
+
|
76 |
+
|
77 |
+
## Licence
|
78 |
+
|
79 |
+
The model weights are published under non-commercial license CC BY-NC 4.0 as the model was finetuned on a dataset for non-commercial use.
|
model_class.py
ADDED
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import transformers
|
2 |
+
import torch
|
3 |
+
from typing import Optional, Tuple, Union
|
4 |
+
from transformers.modeling_outputs import Seq2SeqLMOutput
|
5 |
+
from transformers.generation.logits_process import WhisperTimeStampLogitsProcessor
|
6 |
+
from transformers.models.whisper.tokenization_whisper import TASK_IDS, TO_LANGUAGE_CODE
|
7 |
+
|
8 |
+
|
9 |
+
class WhisperForAudioCaptioning(transformers.WhisperForConditionalGeneration):
|
10 |
+
def __init__(self, config):
|
11 |
+
super().__init__(config)
|
12 |
+
|
13 |
+
@property
|
14 |
+
def task_mapping(self):
|
15 |
+
return {v: k for k, v in self.config.task_mapping.items()}
|
16 |
+
|
17 |
+
@property
|
18 |
+
def named_task_mapping(self):
|
19 |
+
return self.config.task_mapping
|
20 |
+
|
21 |
+
def forward(
|
22 |
+
self,
|
23 |
+
input_features: Optional[torch.FloatTensor] = None,
|
24 |
+
attention_mask: Optional[torch.LongTensor] = None,
|
25 |
+
decoder_input_ids: Optional[torch.LongTensor] = None,
|
26 |
+
decoder_position_ids: Optional[torch.LongTensor] = None,
|
27 |
+
decoder_attention_mask: Optional[torch.LongTensor] = None,
|
28 |
+
head_mask: Optional[torch.Tensor] = None,
|
29 |
+
decoder_head_mask: Optional[torch.Tensor] = None,
|
30 |
+
cross_attn_head_mask: Optional[torch.Tensor] = None,
|
31 |
+
encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
32 |
+
past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
|
33 |
+
decoder_inputs_embeds: Optional[Tuple[torch.FloatTensor]] = None,
|
34 |
+
labels: Optional[torch.LongTensor] = None,
|
35 |
+
use_cache: Optional[bool] = None,
|
36 |
+
output_attentions: Optional[bool] = None,
|
37 |
+
output_hidden_states: Optional[bool] = None,
|
38 |
+
return_dict: Optional[bool] = None,
|
39 |
+
forced_ac_decoder_ids: Optional[
|
40 |
+
torch.LongTensor
|
41 |
+
] = None, # added to be ignored when passed from trainer
|
42 |
+
) -> Union[Tuple[torch.Tensor], Seq2SeqLMOutput]:
|
43 |
+
return super().forward(
|
44 |
+
input_features=input_features,
|
45 |
+
attention_mask=attention_mask,
|
46 |
+
decoder_input_ids=decoder_input_ids,
|
47 |
+
decoder_position_ids=decoder_position_ids,
|
48 |
+
decoder_attention_mask=decoder_attention_mask,
|
49 |
+
head_mask=head_mask,
|
50 |
+
decoder_head_mask=decoder_head_mask,
|
51 |
+
cross_attn_head_mask=cross_attn_head_mask,
|
52 |
+
encoder_outputs=encoder_outputs,
|
53 |
+
past_key_values=past_key_values,
|
54 |
+
decoder_inputs_embeds=decoder_inputs_embeds,
|
55 |
+
labels=labels,
|
56 |
+
use_cache=use_cache,
|
57 |
+
output_attentions=output_attentions,
|
58 |
+
output_hidden_states=output_hidden_states,
|
59 |
+
return_dict=return_dict,
|
60 |
+
)
|
61 |
+
|
62 |
+
# copy-pasted and adapted from transformers.WhisperForConditionalGeneration.generate
|
63 |
+
def generate(
|
64 |
+
self,
|
65 |
+
inputs: Optional[torch.Tensor] = None,
|
66 |
+
forced_ac_decoder_ids: Optional[torch.Tensor] = None,
|
67 |
+
generation_config=None,
|
68 |
+
logits_processor=None,
|
69 |
+
stopping_criteria=None,
|
70 |
+
prefix_allowed_tokens_fn=None,
|
71 |
+
synced_gpus=False,
|
72 |
+
return_timestamps=None,
|
73 |
+
task="transcribe",
|
74 |
+
language="english",
|
75 |
+
**kwargs,
|
76 |
+
):
|
77 |
+
if generation_config is None:
|
78 |
+
generation_config = self.generation_config
|
79 |
+
|
80 |
+
if return_timestamps is not None:
|
81 |
+
if not hasattr(generation_config, "no_timestamps_token_id"):
|
82 |
+
raise ValueError(
|
83 |
+
"You are trying to return timestamps, but the generation config is not properly set."
|
84 |
+
"Make sure to initialize the generation config with the correct attributes that are needed such as `no_timestamps_token_id`."
|
85 |
+
"For more details on how to generate the approtiate config, refer to https://github.com/huggingface/transformers/issues/21878#issuecomment-1451902363"
|
86 |
+
)
|
87 |
+
|
88 |
+
generation_config.return_timestamps = return_timestamps
|
89 |
+
else:
|
90 |
+
generation_config.return_timestamps = False
|
91 |
+
|
92 |
+
if language is not None:
|
93 |
+
generation_config.language = language
|
94 |
+
if task is not None:
|
95 |
+
generation_config.task = task
|
96 |
+
|
97 |
+
forced_decoder_ids = []
|
98 |
+
if task is not None or language is not None:
|
99 |
+
if hasattr(generation_config, "language"):
|
100 |
+
if generation_config.language in generation_config.lang_to_id.keys():
|
101 |
+
language_token = generation_config.language
|
102 |
+
elif generation_config.language in TO_LANGUAGE_CODE.keys():
|
103 |
+
language_token = (
|
104 |
+
f"<|{TO_LANGUAGE_CODE[generation_config.language]}|>"
|
105 |
+
)
|
106 |
+
else:
|
107 |
+
raise ValueError(
|
108 |
+
f"Unsupported language: {language}. Language should be one of:"
|
109 |
+
f" {list(TO_LANGUAGE_CODE.keys()) if generation_config.language in TO_LANGUAGE_CODE.keys() else list(TO_LANGUAGE_CODE.values())}."
|
110 |
+
)
|
111 |
+
forced_decoder_ids.append(
|
112 |
+
(1, generation_config.lang_to_id[language_token])
|
113 |
+
)
|
114 |
+
else:
|
115 |
+
forced_decoder_ids.append(
|
116 |
+
(1, None)
|
117 |
+
) # automatically detect the language
|
118 |
+
|
119 |
+
if hasattr(generation_config, "task"):
|
120 |
+
if generation_config.task in TASK_IDS:
|
121 |
+
forced_decoder_ids.append(
|
122 |
+
(2, generation_config.task_to_id[generation_config.task])
|
123 |
+
)
|
124 |
+
else:
|
125 |
+
raise ValueError(
|
126 |
+
f"The `{generation_config.task}`task is not supported. The task should be one of `{TASK_IDS}`"
|
127 |
+
)
|
128 |
+
else:
|
129 |
+
forced_decoder_ids.append(
|
130 |
+
(2, generation_config.task_to_id["transcribe"])
|
131 |
+
) # defaults to transcribe
|
132 |
+
if (
|
133 |
+
hasattr(generation_config, "no_timestamps_token_id")
|
134 |
+
and not generation_config.return_timestamps
|
135 |
+
):
|
136 |
+
idx = forced_decoder_ids[-1][0] + 1 if forced_decoder_ids else 1
|
137 |
+
forced_decoder_ids.append(
|
138 |
+
(idx, generation_config.no_timestamps_token_id)
|
139 |
+
)
|
140 |
+
|
141 |
+
# Legacy code for backward compatibility
|
142 |
+
elif (
|
143 |
+
hasattr(self.config, "forced_decoder_ids")
|
144 |
+
and self.config.forced_decoder_ids is not None
|
145 |
+
):
|
146 |
+
forced_decoder_ids = self.config.forced_decoder_ids
|
147 |
+
elif (
|
148 |
+
hasattr(self.generation_config, "forced_decoder_ids")
|
149 |
+
and self.generation_config.forced_decoder_ids is not None
|
150 |
+
):
|
151 |
+
forced_decoder_ids = self.generation_config.forced_decoder_ids
|
152 |
+
|
153 |
+
if generation_config.return_timestamps:
|
154 |
+
logits_processor = [WhisperTimeStampLogitsProcessor(generation_config)]
|
155 |
+
|
156 |
+
decoder_input_ids = None
|
157 |
+
|
158 |
+
if len(forced_decoder_ids) > 0:
|
159 |
+
# get the token sequence coded in forced_decoder_ids
|
160 |
+
forced_decoder_ids.sort()
|
161 |
+
if min(forced_decoder_ids)[0] != 0:
|
162 |
+
forced_decoder_ids = [
|
163 |
+
(0, self.config.decoder_start_token_id)
|
164 |
+
] + forced_decoder_ids
|
165 |
+
|
166 |
+
position_indices, decoder_input_ids = zip(*forced_decoder_ids)
|
167 |
+
assert tuple(position_indices) == tuple(
|
168 |
+
range(len(position_indices))
|
169 |
+
), "forced_decoder_ids is not a (continuous) prefix, we can't handle that"
|
170 |
+
|
171 |
+
device = self.get_decoder().device
|
172 |
+
|
173 |
+
if forced_ac_decoder_ids is None:
|
174 |
+
forced_ac_decoder_ids = torch.tensor(
|
175 |
+
[[]], device=device, dtype=torch.long
|
176 |
+
)
|
177 |
+
|
178 |
+
# enrich every sample's forced_ac_decoder_ids with Whisper's forced_decoder_ids
|
179 |
+
batch_size = forced_ac_decoder_ids.shape[0]
|
180 |
+
fluff_len = len(decoder_input_ids)
|
181 |
+
decoder_input_ids = torch.tensor(
|
182 |
+
decoder_input_ids, device=device, dtype=torch.long
|
183 |
+
)
|
184 |
+
decoder_input_ids = decoder_input_ids.expand((batch_size, fluff_len))
|
185 |
+
decoder_input_ids = torch.cat(
|
186 |
+
[decoder_input_ids, forced_ac_decoder_ids], dim=1
|
187 |
+
)
|
188 |
+
|
189 |
+
generation_config.forced_decoder_ids = forced_decoder_ids
|
190 |
+
|
191 |
+
return super(
|
192 |
+
transformers.WhisperPreTrainedModel, self
|
193 |
+
).generate( # changed by adam (calling grandparent)
|
194 |
+
inputs,
|
195 |
+
generation_config,
|
196 |
+
logits_processor,
|
197 |
+
stopping_criteria,
|
198 |
+
prefix_allowed_tokens_fn,
|
199 |
+
synced_gpus,
|
200 |
+
decoder_input_ids=decoder_input_ids,
|
201 |
+
**kwargs,
|
202 |
+
)
|