python/FunASR-XL.git

			@@ -1,11 +1,5 @@
			# This is an example that demonstrates how to configure a model file.
			# You can modify the configuration according to your own requirements.

			# to print the register_table:
			# from funasr.register import tables
			# tables.print()
			# network architecture
			model: SenseVoice
			model: SenseVoiceRWKV
			model_conf:
			lsm_weight: 0.1
			length_normalized_loss: true
			@@ -27,6 +21,26 @@
			n_text_head: 20
			n_text_layer: 32


			# decoder
			decoder: SenseVoiceDecoder
			decoder_conf:
			rwkv_cfg:
			n_embd: 1280
			dropout: 0
			head_size_a: 64
			ctx_len: 1280
			dim_att: 1280 #${model_conf.rwkv_cfg.n_embd}
			dim_ffn: null
			head_size_divisor: 8
			n_layer: 32
			pre_ffn: 0
			ln0: false
			ln1: false
			init_rwkv: false
			datatype: bf16


			# frontend related
			frontend: WhisperFrontend
			frontend_conf:
			@@ -42,12 +56,17 @@

			dataset: SenseVoiceDataset
			dataset_conf:
			index_ds: IndexDSJsonl
			index_ds: IndexDSJsonlRankSplit
			batch_sampler: EspnetStyleBatchSampler
			batch_type: length # example or length
			batch_size: 7000 # if batch_type is example, batch_size is the numbers of samples; if length, batch_size is source_token_len+target_token_len;
			max_token_length: 2000 # filter samples if source_token_len+target_token_len > max_token_length,
			rank_split: true
			batch_type: token # example or length
			batch_size: 3500 # if batch_type is example, batch_size is the numbers of samples; if length, batch_size is source_token_len+target_token_len;
			max_token_length: 2200
			min_token_length: 60
			max_source_length: 2000
			min_source_length: 60
			max_target_length: 150
			min_target_length: 0
			shuffle: True
			num_workers: 4
			sos: ${model_conf.sos}
			@@ -60,6 +79,7 @@
			keep_nbest_models: 20
			avg_nbest_model: ${train_conf.keep_nbest_models}
			log_interval: 50
			reset_gpu_cache: true

			optim: adamw
			optim_conf: