| 12345678910111213141516171819202122232425262728293031323334353637383940414243444546474849505152535455565758596061626364656667686970717273747576777879808182838485868788 |
- 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_fleurs_arb-eng,dev_fleurs_ben-eng,dev_fleurs_hin-eng,dev_fleurs_ind-eng,dev_fleurs_ita-eng,dev_fleurs_jpn-eng,dev_fleurs_por-eng,dev_fleurs_rus-eng,dev_fleurs_swh-eng,dev_fleurs_tha-eng,dev_fleurs_tur-eng,dev_fleurs_urd-eng,dev_fleurs_vie-eng,dev_fleurs_spa-eng,dev_fleurs_eng-arb,dev_fleurs_eng-ben,dev_fleurs_eng-hin,dev_fleurs_eng-ind,dev_fleurs_eng-ita,dev_fleurs_eng-jpn,dev_fleurs_eng-por,dev_fleurs_eng-rus,dev_fleurs_eng-swh,dev_fleurs_eng-tha,dev_fleurs_eng-tur,dev_fleurs_eng-urd,dev_fleurs_eng-vie,dev_fleurs_eng-spa
- manifest_list_path: null
- manifest_path_prefix: /fsx-ust/spopuri/datasets/S2ST/V1/M4T_V1_phase2/primary/
- max_seconds_per_input_audio: 150
- fixed_batch_size: 10
- max_tgt_text_tokens_per_batch: null
- max_tgt_text_tokens_per_sample: 3000
- max_units_per_sample: 1500
- num_threads: 10
- prefech_batches: 10
- prepend_tgt_lang_tag: true
- shuffle_window: 1000
- text_tokenization:
- from_model: seamlessM4T_large
- spm_path: null
- langtoks: null
- unit_tokenization:
- from_model: seamlessM4T_large
- langtoks: null
- num_units: null
- unit_tokenizer_name: seamlessM4T_large
- model:
- custom_params:
- model_embed_dim: 1024
- nllb_vocabulary_size: 256103
- w2v2_encoder_layers: 24
- w2v2_encoder_layers_use_conformer: true
- w2v2_encoder_layers_layernorm_features: false
- w2v2_pos_encoder_type: "relative"
- w2v2_pos_encoder_depth: 0
- w2v2_pos_conv_kernel_size: 0
- w2v2_num_pos_conv_groups: 0
- nllb_encoder_layers: 24
- nllb_decoder_layers: 24
- t2u_encoder_layers: 6
- t2u_decoder_layers: 6
- unit_vocabulary_size: 10082
- from_model: null
- from_model_config: null
- pretrained_s2t_decoder_path: /fsx-ust/spopuri/datasets/PT_CKPT/S2T/S2T_M4T_V1_V1_cleaned.pt
- pretrained_t2u_path: /fsx-ust/spopuri/datasets/PT_CKPT/T2U/V5_10K_p2_14_80K.pt
- pretrained_w2v2_path: /fsx-ust/spopuri/datasets/PT_CKPT/w2v2/w2vbert2rpq_600m_al5.pt
- 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: null
- manifest_list_path: /data/home/mavlyutov/train_configs/m4t_v1_train_manifests.txt
- manifest_path_prefix: /fsx-ust/spopuri/datasets/S2ST/V1/M4T_V1_phase2/primary
- max_seconds_per_input_audio: 15
- fixed_batch_size: null
- max_batch_size: 25
- max_tgt_text_tokens_per_batch: 300
- max_tgt_text_tokens_per_sample: 150
- max_units_per_sample: 1200
- num_threads: 10
- prefech_batches: 10
- prepend_tgt_lang_tag: true
- shuffle_window: 1000
- text_tokenization:
- from_model: seamlessM4T_large
- spm_path: null
- langtoks: null
- unit_tokenization:
- from_model: seamlessM4T_large
- langtoks: null
- num_units: null
- unit_tokenizer_name: seamlessM4T_large
- training:
- eval_steps: 5000
- float_dtype: bf16
- label_smoothing: 0.2
- learning_rate: 0.0001
- log_steps: 200
- max_epochs: 100
- patience: 10
- start_learning_rate: 1.0e-07
- warmup_steps: 1000
|
