efrainmain/draft-chunkformer-finetune-encoder

1encoder: chunkformer2is_json_cmvn: true3cmvn_file: chunkformer-large-vie/global_cmvn4input_dim: 805output_dim: 69926 7encoder_conf:8    output_size: 512    # dimension of attention9    attention_heads: 810    linear_units: 2048  # the number of units of position-wise feed forward11    num_blocks: 17      # the number of encoder blocks12    dropout_rate: 0.113    positional_dropout_rate: 0.114    attention_dropout_rate: 0.115    input_layer: 'dw_striding' # encoder input type, you can chose conv2d, conv2d6 and conv2d816    normalize_before: true17    cnn_module_kernel: 1518    use_cnn_module: true19    activation_type: 'swish'20    pos_enc_layer_type: 'chunk_rel_pos'21    selfattention_layer_type: 'chunk_rel_seflattn'22    causal: false23    use_dynamic_chunk: false24    dynamic_chunk_sizes: [64, 128, 256]25    dynamic_right_context_sizes: [128, 128, 128]26    dynamic_left_context_sizes: [128, 256, 128]27    cnn_module_norm: 'layer_norm' # using nn.LayerNorm makes model converge faster28    use_dynamic_left_chunk: false29    dynamic_conv: true30