| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788899091929394959697 |
- eval_data:
- audio:
- audio_root_dir: /fsx-ust/data/audio_zips/
- fbanks_num_mel_bins: 80
- fbanks_standardize_audio: true
- fbanks_waveform_scale: 32768
- fbank_feats_pad_idx: 0
- manifest_list: dev_asr_only_aggregated_adapted
- manifest_list_path: null
- manifest_path_prefix: /data/home/mavlyutov/s2t_ondevice/
- max_seconds_per_input_audio: 15
- fixed_batch_size: 40
- max_tgt_text_tokens_per_batch: 1000
- max_tgt_text_tokens_per_sample: 300
- max_units_per_sample: 1500
- num_threads: 5
- prefech_batches: null
- prepend_tgt_lang_tag: true
- shuffle_window: 1000
- text_tokenization:
- from_model: null
- langtoks:
- - eng
- - rus
- - hin
- - por
- - spa
- spm_path: /data/home/mavlyutov/s2t_ondevice/vocab20k/5_5_20k.model
- unit_tokenization:
- from_model: seamlessM4T_large
- langtoks: null
- num_units: null
- unit_tokenizer_name: seamlessM4T_large
- model:
- custom_params:
- model_embed_dim: 768
- nllb_decoder_layers: 3
- nllb_encoder_layers: 1
- nllb_vocabulary_size: 256102
- t2u_decoder_layers: 1
- t2u_encoder_layers: 1
- unit_vocabulary_size: 10082
- w2v2_encoder_layers: 6
- w2v2_encoder_layers_layernorm_features: false
- w2v2_encoder_layers_use_conformer: true
- w2v2_num_pos_conv_groups: 0
- w2v2_pos_conv_kernel_size: 0
- w2v2_pos_encoder_depth: 0
- w2v2_pos_encoder_type: relative
- from_model: null
- from_model_config: null
- pretrained_s2t_decoder_path: null
- pretrained_t2u_path: null
- pretrained_w2v2_path: null
- train_data:
- audio:
- audio_root_dir: /fsx-ust/data/audio_zips/
- fbanks_num_mel_bins: 80
- fbanks_standardize_audio: true
- fbanks_waveform_scale: 32768
- fbank_feats_pad_idx: 0
- manifest_list: train_asr_only_aggregated_5_dial_filtered_adapted_wh_transc
- manifest_list_path: null
- manifest_path_prefix: /data/home/mavlyutov/s2t_ondevice/
- max_seconds_per_input_audio: 15
- fixed_batch_size: 40
- max_tgt_text_tokens_per_batch: 600
- max_tgt_text_tokens_per_sample: 300
- max_units_per_sample: 1500
- num_threads: 4
- prefech_batches: null
- prepend_tgt_lang_tag: true
- shuffle_window: 1000
- text_tokenization:
- from_model: null
- langtoks:
- - eng
- - rus
- - hin
- - por
- - spa
- spm_path: /data/home/mavlyutov/s2t_ondevice/vocab20k/5_5_20k.model
- unit_tokenization:
- from_model: seamlessM4T_large
- langtoks: null
- num_units: null
- unit_tokenizer_name: seamlessM4T_large
- training:
- eval_steps: 1000
- float_dtype: fp32
- label_smoothing: 0.2
- learning_rate: 0.0001
- log_steps: 50
- max_epochs: 100
- patience: 10
- start_learning_rate: 1.0e-07
- warmup_steps: 1000
|
