python/FunASR-XL.git

			@@ -2,47 +2,52 @@
			# encoder related
			encoder: data2vec_encoder
			encoder_conf:
			extractor_mode: layer_norm
			encoder_layerdrop: 0.05
			dropout_input: 0.0
			dropout_features: 0.0
			feature_grad_mult: 1.0
			encoder_embed_dim: 768
			extractor_mode: layer_norm
			encoder_layerdrop: 0.05
			dropout_input: 0.0
			dropout_features: 0.0
			feature_grad_mult: 1.0
			encoder_embed_dim: 768

			mask_prob: 0.65
			mask_length: 10
			mask_prob: 0.65
			mask_length: 10

			loss_beta: 0
			loss_scale: null
			loss_beta: 0
			loss_scale: null

			instance_norm_target_layer: true
			average_top_k_layers: 8
			instance_norm_target_layer: true
			average_top_k_layers: 8

			pos_conv_depth: 5
			conv_pos: 95
			pos_conv_depth: 5
			conv_pos: 95

			ema_decay: 0.999
			ema_end_decay: 0.9999
			ema_anneal_end_step: 30000
			ema_transformer_only: true
			ema_layers_only: true
			ema_decay: 0.999
			ema_end_decay: 0.9999
			ema_anneal_end_step: 30000
			ema_transformer_only: true
			ema_layers_only: true

			require_same_masks: true
			mask_dropout: 0
			require_same_masks: true
			mask_dropout: 0

			log_interval: 50
			normalize: None
			# 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

			# minibatch related
			batch_type: length
			batch_bins: 64000
			num_workers: 16
			model: data2vec

			# optimization related
			accum_grad: 1
			grad_clip: 5
			patience: none
			max_epoch: 600
			max_epoch: 1800
			val_scheduler_criterion:
			- valid
			- acc
			@@ -67,8 +72,8 @@
			# for dataset
			dataset_conf:
			batch_mode: clipping
			data_names: speech,none
			data_types: kaldi_ark,none
			data_names: speech
			data_types: sound
			shuffle: true
			shuffle_conf:
			shuffle_size: 12800
			@@ -76,4 +81,7 @@
			batch_conf:
			batch_type: token
			batch_size: 64000
			num_workers: 8
			num_workers: 8

			log_interval: 50
			normalize: None