egs/aishell/paraformer/conf/train_asr_paraformer_conformer_12e_6d_2048_256.yaml
@@ -84,7 +84,7 @@ - 40 num_time_mask: 2 predictor: cif_predictor_v2 predictor: cif_predictor predictor_conf: idim: 256 threshold: 1.0 egs/aishell/paraformerbert/conf/train_asr_paraformerbert_conformer_12e_6d_2048_256.yaml
@@ -29,6 +29,17 @@ self_attention_dropout_rate: 0.0 src_attention_dropout_rate: 0.0 # frontend related frontend: wav_frontend frontend_conf: fs: 16000 window: hamming n_mels: 80 frame_length: 25 frame_shift: 10 lfr_m: 1 lfr_n: 1 # hybrid CTC/attention model: paraformer_bert model_conf: @@ -41,19 +52,10 @@ embed_dims: 768 embeds_loss_weight: 2.0 # minibatch related #batch_type: length #batch_bins: 40000 batch_type: numel batch_bins: 2000000 num_workers: 16 # optimization related accum_grad: 4 accum_grad: 1 grad_clip: 5 max_epoch: 50 max_epoch: 150 val_scheduler_criterion: - valid - acc @@ -92,8 +94,17 @@ threshold: 1.0 l_order: 1 r_order: 1 tail_threshold: 0.45 dataset_conf: shuffle: True shuffle_conf: shuffle_size: 2048 sort_size: 500 batch_conf: batch_type: token batch_size: 25000 num_workers: 8 log_interval: 50 normalize: None allow_variable_data_keys: true normalize: None egs/aishell/paraformerbert/run.sh
@@ -111,12 +111,12 @@ world_size=$gpu_num # run on one machine if [ ${stage} -le 3 ] && [ ${stop_stage} -ge 3 ]; then echo "stage 3: Training" if ! "${skip_extract_embed}"; then echo "extract embeddings..." local/extract_embeds.sh \ --bert_model_name ${bert_model_name} \ --raw_dataset_path ${feats_dir} fi # if ! "${skip_extract_embed}"; then # echo "extract embeddings..." # local/extract_embeds.sh \ # --bert_model_name ${bert_model_name} \ # --raw_dataset_path ${feats_dir} # fi mkdir -p ${exp_dir}/exp/${model_dir} mkdir -p ${exp_dir}/exp/${model_dir}/log INIT_FILE=${exp_dir}/exp/${model_dir}/ddp_init
