Create params.yaml

This parameters are used while training wav2vec model on NMSQA dataset.

params.yaml

@@ -0,0 +1,44 @@
+mode: "train"
+pretrained_model_name: "facebook/wav2vec2-base-960h"
+freeze_pretrained: True
+output_dir: "out/model_asr_base_alldata"
+dataset_script: "dataset/asr_dataset.py"
+part_name: "nmsqa_all_asr"
+window_size: 321
+learning_rate: 0.0002
+group_by_length: True
+per_device_train_batch_size: 2
+num_train_epochs: 10
+fp16: True
+save_steps: 1000
+logging_steps: 500
+warmup_steps: 500
+weight_decay: 0.0005
+load_best_model_at_end: True
+metric_for_best_model: wer
+greater_is_better: True
+save_total_limit: 2
+eval: True
+per_device_eval_batch_size: 2
+eval_steps: 1000
+evaluation_strategy: "steps"
+per_device_test_batch_size: 2
+shuffle: False
+attention_dropout: 0.0
+hidden_dropout: 0.0
+feat_proj_dropout: 0.0
+mask_time_prob: 0.05
+layer_dropout: 0.0
+ctc_loss_reduction: "mean"
+unk: "[UNK]"
+pad: "[PAD]"
+word_delimited: "|"
+sampling_rate: 16000
+padding_value: 0.0
+feature_size: 1
+normalize: True
+gradient_accumulation_steps: 2
+gradient_checkpointing: True
+max_answer_length : 500
+test_weight: 0.6
+eval_part_name: validation