Upload hyperparams.yaml

hyperparams.yaml

@@ -0,0 +1,241 @@
+# ############################################################################
+# Model: Streaming E2E Conformer-Transducer ASR
+# Encoder: Conformer
+# Decoder: LSTM + greedy search
+# Tokens: BPE with unigram
+# losses: Transducer + CTC (optional) + CE (optional)
+# Training: Librispeech 960h
+# Authors:  Sylvain de Langen 2023, Titouan Parcollet 2023, Abdel HEBA, Mirco Ravanelli, Sung-Lin Yeh 2020
+# ############################################################################
+
+save_folder: !ref librispeech-streaming-conformer-transducer
+
+# Training parameters
+# To make Transformers converge, the global bath size should be large enough.
+# The global batch size is computed as batch_size * n_gpus * grad_accumulation_factor.
+# Empirically, we found that this value should be >= 128.
+# Please, set your parameters accordingly.
+number_of_epochs: 50
+warmup_steps: 25000
+num_workers: 4
+batch_size_valid: 4
+lr: 0.0008
+weight_decay: 0.01
+number_of_ctc_epochs: 40
+ctc_weight: 0.3 # Multitask with CTC for the encoder (0.0 = disabled)
+ce_weight: 0.0 # Multitask with CE for the decoder (0.0 = disabled)
+max_grad_norm: 5.0
+loss_reduction: 'batchmean'
+precision: fp32 # bf16, fp16 or fp32
+
+# The batch size is used if and only if dynamic batching is set to False
+# Validation and testing are done with fixed batches and not dynamic batching.
+batch_size: 8
+grad_accumulation_factor: 4
+sorting: ascending
+avg_checkpoints: 10 # Number of checkpoints to average for evaluation
+
+# Feature parameters
+sample_rate: 16000
+n_fft: 512
+n_mels: 80
+win_length: 32
+
+# Streaming
+streaming: True  # controls all Dynamic Chunk Training & chunk size & left context mechanisms
+
+# This setup works well for 3090 24GB GPU, adapt it to your needs.
+# Adjust grad_accumulation_factor depending on the DDP node count (here 3)
+# Or turn it off (but training speed will decrease)
+dynamic_batching: True
+max_batch_len: 250
+max_batch_len_val: 50 # we reduce it as the beam is much wider (VRAM)
+num_bucket: 200
+
+dynamic_batch_sampler:
+   max_batch_len: !ref <max_batch_len>
+   max_batch_len_val: !ref <max_batch_len_val>
+   num_buckets: !ref <num_bucket>
+   shuffle_ex: True # if true re-creates batches at each epoch shuffling examples.
+   batch_ordering: random
+   max_batch_ex: 256
+
+# Model parameters
+# Transformer
+d_model: 512
+joint_dim: 640
+nhead: 8
+num_encoder_layers: 12
+num_decoder_layers: 0
+d_ffn: 2048
+transformer_dropout: 0.1
+activation: !name:torch.nn.GELU
+output_neurons: 1000
+dec_dim: 512
+dec_emb_dropout: 0.2
+dec_dropout: 0.1
+
+# Decoding parameters
+blank_index: 0
+bos_index: 0
+eos_index: 0
+pad_index: 0
+beam_size: 10
+nbest: 1
+# by default {state,expand}_beam = 2.3 as mention in paper
+# https://arxiv.org/abs/1904.02619
+state_beam: 2.3
+expand_beam: 2.3
+lm_weight: 0.50
+
+epoch_counter: !new:speechbrain.utils.epoch_loop.EpochCounter
+   limit: !ref <number_of_epochs>
+
+normalize: !new:speechbrain.processing.features.InputNormalization
+   norm_type: global
+   update_until_epoch: 4
+
+compute_features: !new:speechbrain.lobes.features.Fbank
+   sample_rate: !ref <sample_rate>
+   n_fft: !ref <n_fft>
+   n_mels: !ref <n_mels>
+   win_length: !ref <win_length>
+
+CNN: !new:speechbrain.lobes.models.convolution.ConvolutionFrontEnd
+   input_shape: (8, 10, 80)
+   num_blocks: 2
+   num_layers_per_block: 1
+   out_channels: (64, 32)
+   kernel_sizes: (3, 3)
+   strides: (2, 2)
+   residuals: (False, False)
+
+Transformer: !new:speechbrain.lobes.models.transformer.TransformerASR.TransformerASR # yamllint disable-line rule:line-length
+   input_size: 640
+   tgt_vocab: !ref <output_neurons>
+   d_model: !ref <d_model>
+   nhead: !ref <nhead>
+   num_encoder_layers: !ref <num_encoder_layers>
+   num_decoder_layers: !ref <num_decoder_layers>
+   d_ffn: !ref <d_ffn>
+   dropout: !ref <transformer_dropout>
+   activation: !ref <activation>
+   encoder_module: conformer
+   attention_type: RelPosMHAXL
+   normalize_before: True
+   causal: False
+
+# We must call an encoder wrapper so the decoder isn't run (we don't have any)
+enc: !new:speechbrain.lobes.models.transformer.TransformerASR.EncoderWrapper
+   transformer: !ref <Transformer>
+
+# For MTL CTC over the encoder
+proj_ctc: !new:speechbrain.nnet.linear.Linear
+   input_size: !ref <joint_dim>
+   n_neurons: !ref <output_neurons>
+
+# Define some projection layers to make sure that enc and dec
+# output dim are the same before joining
+proj_enc: !new:speechbrain.nnet.linear.Linear
+   input_size: !ref <d_model>
+   n_neurons: !ref <joint_dim>
+   bias: False
+
+proj_dec: !new:speechbrain.nnet.linear.Linear
+   input_size: !ref <dec_dim>
+   n_neurons: !ref <joint_dim>
+   bias: False
+
+# Uncomment for MTL with CTC
+ctc_cost: !name:speechbrain.nnet.losses.ctc_loss
+   blank_index: !ref <blank_index>
+   reduction: !ref <loss_reduction>
+
+emb: !new:speechbrain.nnet.embedding.Embedding
+   num_embeddings: !ref <output_neurons>
+   consider_as_one_hot: True
+   blank_id: !ref <blank_index>
+
+dec: !new:speechbrain.nnet.RNN.LSTM
+   input_shape: [null, null, !ref <output_neurons> - 1]
+   hidden_size: !ref <dec_dim>
+   num_layers: 1
+   re_init: True
+
+Tjoint: !new:speechbrain.nnet.transducer.transducer_joint.Transducer_joint
+   joint: sum # joint [sum | concat]
+   nonlinearity: !ref <activation>
+
+transducer_lin: !new:speechbrain.nnet.linear.Linear
+   input_size: !ref <joint_dim>
+   n_neurons: !ref <output_neurons>
+   bias: False
+
+log_softmax: !new:speechbrain.nnet.activations.Softmax
+   apply_log: True
+
+# for MTL
+# update model if any HEAD module is added
+modules:
+   CNN: !ref <CNN>
+   enc: !ref <enc>
+   emb: !ref <emb>
+   dec: !ref <dec>
+   Tjoint: !ref <Tjoint>
+   transducer_lin: !ref <transducer_lin>
+   normalize: !ref <normalize>
+   proj_ctc: !ref <proj_ctc>
+   proj_dec: !ref <proj_dec>
+   proj_enc: !ref <proj_enc>
+#   dec_lin: !ref <dec_lin>
+
+# for MTL
+# update model if any HEAD module is added
+model: !new:torch.nn.ModuleList
+   - [!ref <CNN>, !ref <enc>, !ref <emb>, !ref <dec>, !ref <proj_enc>, !ref <proj_dec>, !ref <proj_ctc>, !ref <transducer_lin>]
+
+# Tokenizer initialization
+tokenizer: !new:sentencepiece.SentencePieceProcessor
+
+Greedysearcher: !new:speechbrain.decoders.transducer.TransducerBeamSearcher
+   decode_network_lst: [!ref <emb>, !ref <dec>, !ref <proj_dec>]
+   tjoint: !ref <Tjoint>
+   classifier_network: [!ref <transducer_lin>]
+   blank_id: !ref <blank_index>
+   beam_size: 1
+   nbest: 1
+
+Beamsearcher: !new:speechbrain.decoders.transducer.TransducerBeamSearcher
+   decode_network_lst: [!ref <emb>, !ref <dec>, !ref <proj_dec>]
+   tjoint: !ref <Tjoint>
+   classifier_network: [!ref <transducer_lin>]
+   blank_id: !ref <blank_index>
+   beam_size: !ref <beam_size>
+   nbest: !ref <nbest>
+   # FIXME: when lm pretrained, use this
+   # lm_module: !ref <lm_model>
+   # lm_weight: !ref <lm_weight>
+   state_beam: !ref <state_beam>
+   expand_beam: !ref <expand_beam>
+
+pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
+   collect_in: !ref <save_folder>
+   loadables:
+      model: !ref <model>
+      normalizer: !ref <normalize>
+      tokenizer: !ref <tokenizer>
+
+# inference stuff
+
+make_tokenizer_streaming_context: !name:speechbrain.tokenizers.SentencePiece.SentencePieceDecoderStreamingContext
+tokenizer_decode_streaming: !name:speechbrain.tokenizers.SentencePiece.spm_decode_preserve_leading_space
+
+fea_streaming_extractor: !new:speechbrain.lobes.features.StreamingFeatureWrapper
+   module: !new:speechbrain.lobes.models.convolution.ConformerFeatureExtractorWrapper
+      - !ref <compute_features>
+      - !ref <normalize>
+      - !ref <CNN>
+   # don't consider normalization as part of the input filter chain
+   properties: !!python/object/apply:speechbrain.utils.filter_analysis.stack_filter_properties
+      - [!ref <compute_features>, !ref <CNN>]
+