Hugging Face's logo

nvidia
/
audio-flamingo

English
audio understanding
audio captioning
audio question answering
audio classification
audio dialogues
retrieval augmented generation
in-context learning
Model card Files Community
1
audio-flamingo
File size: 5,233 Bytes 
9f748f8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
 
train_config:
  expdir: YOUR_DATA_ROOT_DIR/audio-flamingo-data/checkpoint
  run_name: foundation_pretrain
  delete_previous_checkpoint: true 
  batch_size: 6
  gradient_accumulation_steps: 8  # global batchsize = 384
  seed: 42
  learning_rate: 0.0001
  lr_scheduler: constant
  loss_multiplier: 1.0
  warmup_steps: 1875
  weight_decay: 0.1
  precision: amp_bf16
  gradient_checkpointing: False 
  num_epochs: 100  # num_epochs * dataset_blending_global_weight = 1
  offline: false
  freeze_lm_embeddings: true
  logging_steps: 10
  dist_backend: nccl
  dist_url: env://
  no_set_device_rank: false 
  fsdp: true 
  fsdp_use_orig_params: false  # Passed into the FSDP constructor. Enables param_groups and gradient masking for weight_decay. Does not work with OPT.
  fsdp_sharding_strategy: full  # full, hybrid
  horovod: false

data_config:
  dataset_blending_global_weight: 0.01

  dataset_blending_config:

    # Audio QA
    
    OpenAQA-AQA/train:
      weight: 1.0
      prefix_prob: 0.0
      augmentations:
        do_nothing: 1.0

    # Audio Captioning 

    BBCSoundEffects-AudioDescription/train: 
      weight: 5.0
      prefix_prob: 0.5
      augmentations:
        do_nothing: 1.0

    CLAP_freesound-AudioCaptioning/train: 
      weight: 1.0
      prefix_prob: 0.5
      augmentations:
        do_nothing: 1.0

    SoundDescs-AudioDescription/train:
      weight: 1.0
      prefix_prob: 0.5
      augmentations:
        do_nothing: 1.0

    WavCaps-AudioSet_SL-AudioCaptioning/train: 
      weight: 1.0
      prefix_prob: 0.5
      augmentations:
        do_nothing: 1.0

    WavCaps-BBC_Sound_Effects-AudioCaptioning/train: 
      weight: 2.0
      prefix_prob: 0.5
      augmentations:
        do_nothing: 1.0

    WavCaps-FreeSound-AudioCaptioning/train: 
      weight: 2.0
      prefix_prob: 0.5
      augmentations:
        do_nothing: 1.0
    
    WavCaps-SoundBible-AudioCaptioning/train: 
      weight: 5.0
      prefix_prob: 0.5
      augmentations:
        do_nothing: 1.0

    # Audio Classification

    AudioSetFullwoAudioMusicCaps-EventClassification/train:
      weight: 1.0
      prefix_prob: 0.5
      augmentations:
        num_words: 0.8
        do_nothing: 0.2
    
    WavText5K-Tagging/train: 
      weight: 3.0
      prefix_prob: 0.5
      augmentations:
        num_words: 0.8
        do_nothing: 0.2

    # Speech Emotion Classification

    MSP-PODCAST-Publish-1.9-EmotionClassification/train:
      weight: 1.2
      prefix_prob: 0.5
      augmentations:
        provide_all_labels: 0.9
        do_nothing: 0.1
    
    MELD-EmotionClassification/train:
      weight: 1.2
      prefix_prob: 0.5
      augmentations:
        provide_all_labels: 0.9
        do_nothing: 0.1

    MELD-SentimentClassification/train:
      weight: 1.2
      prefix_prob: 0.5
      augmentations:
        provide_all_labels: 0.9
        do_nothing: 0.1
    
    # Music QA
    
    Music-AVQA-AVQA_All/train:
      weight: 3.0
      prefix_prob: 0.5
      augmentations:
        AQA_binary_instruction: 1.0
    
    MU-LLAMA-AQA/train:
      weight: 1.2
      prefix_prob: 0.5
      augmentations:
        do_nothing: 1.0
    
    # Music Captioning

    LP-MusicCaps-MSD-AudioCaptioning/train:
      weight: 1.0
      prefix_prob: 0.5
      augmentations:
        do_nothing: 1.0

    # Music Understanding 

    NSynth-MIR/train:
      weight: 0.4
      prefix_prob: 0.5
      augmentations:
        do_nothing: 1.0
    
    mtg-jamendo-MusicTagging/train:
      weight: 1.0
      prefix_prob: 0.5
      augmentations:
        do_nothing: 1.0
    
  dataset_file_root: YOUR_DATA_ROOT_DIR/audio-flamingo-data/dataset_files
  data_root: YOUR_DATA_ROOT_DIR/datasets
  dataset_blending_output: dataset_blending.json
  max_tokens: 512
  num_workers: 4

  valid_dataset_config: 
    CLAP_freesound-AudioCaptioning/test: true 
    SoundDescs-AudioDescription/val: true
    Clotho-AQA-EventClassification/val: true
    MSP-PODCAST-Publish-1.9-EmotionClassification/val: true 
    MELD-EmotionClassification/val: true
    MELD-SentimentClassification/val: true
    MU-LLAMA-AQA/test: true 
    LP-MusicCaps-MSD-AudioCaptioning/val: true 
    NSynth-MIR/val: true
    mtg-jamendo-MusicTagging/val: true

clap_config:
  # method: laion-clap
  # audio_embed_dim: 512
  # model_name: 630k-fusion-best
  # checkpoint: YOUR_DATA_ROOT_DIR/audio-flamingo-data/laion-clap-pretrained/laion_clap/630k-fusion-best.pt
  
  method: microsoft-clap
  audio_embed_dim: 1024
  config_root: YOUR_REPO_ROOT_DIR/foundation/my_ms_clap/src/configs
  model_name: 'clapcap'
  checkpoint: YOUR_DATA_ROOT_DIR/audio-flamingo-data/clap/clapcap_weights_2023.pth

  window_length: 7.0  # seconds
  window_overlap: 5.25  # seconds
  max_num_window: 16  # total = 33.25 seconds
  max_num_fewshot: 8  # number of fewshot samples

model_config:
  cache_dir: YOUR_DATA_ROOT_DIR/audio-flamingo-data/LLM_pretrained/.cache

  lang_encoder_path: facebook/opt-iml-max-1.3b
  tokenizer_path: facebook/opt-iml-max-1.3b
  cross_attn_every_n_layers: 1
  audio_transformer_kwargs: {
    n_head: 8,
    n_layers: 3,
    d_inner: 2048,
    max_num_media: 128,  # must >= max_num_window * num_fewshot_samples
    max_window_per_audio: 16,  # must = max_num_window
  }