Update hyperparams.yaml

hyperparams.yaml

@@ -1,114 +1,26 @@
-###################################
-# Experiment Parameters and setup #
-###################################
-seed: 888
-__set_seed: !apply:torch.manual_seed [888]
-output_folder: results/s2ut/888
-save_folder: results/s2ut/888/save
-train_log: results/s2ut/888/train_log.txt
-epochs: 100
-use_tensorboard: true
-
-progress_samples: true
-progress_sample_path: results/s2ut/888/samples
-progress_samples_interval: 1
-progress_batch_sample_size: 4
-
-evaluation_interval: 10
-
-#################################
-# Data files and pre-processing #
-#################################
-src_data_folder: /workspace/speechbrain/common_voice # e.g, /corpus/CommonVoice/fr (French Data)
-tgt_data_folder: /workspace/speechbrain/cvss # e.g, /corpus/CV4/fr (English Data)
+pretrained_path: sinarashidi/s2st_fa-en_cvss
+
 sample_rate: 16000
 
-train_json: results/s2ut/888/save/train.json
-valid_json: results/s2ut/888/save/valid.json
-valid_small_json: results/s2ut/888/save/valid_small.json
-test_json: results/s2ut/888/save/test.json
-splits: [train, valid_small, valid, test]
-skip_prep: false
-
-# SSL model used to encode target features
-encoder_source: facebook/hubert-base-ls960
-layer: 6
-kmeans_source: speechbrain/tts-hifigan-unit-hubert-l6-k100-ljspeech
-codes_folder: results/s2ut/888/save/codes
-skip_extract: false
-
-# Vocoder model used for evaluation
-vocoder_source: speechbrain/tts-hifigan-unit-hubert-l6-k100-ljspeech
-vocoder_download_path: results/s2ut/888/save/pretrained_models/vocoder
-
-# ASR model used for evaluation
-asr_source: speechbrain/asr-wav2vec2-librispeech
-asr_download_path: results/s2ut/888/save/pretrained_models/asr
-
-# Wav2vec2 encoder
-wav2vec2_source: m3hrdadfi/wav2vec2-large-xlsr-persian-v3
-wav2vec2_download_path: results/s2ut/888/save/pretrained_models
-
-# wav2vec2 encoder specific parameters
-wav2vec2_frozen: false
-wav2vec2_freeze_steps: 10000
-
-####################### Training Parameters ####################################
-lr: 0.0005
-lr_wav2vec: 0.00001
-loss_reduction: batchmean
+# URL for the HuggingFace model we want to load as encoder
+wav2vec2_hub: m3hrdadfi/wav2vec2-large-xlsr-persian-v3
 
 # Outputs
-# blank_index: 102
+vocab_size: 103
+blank_index: 102
 bos_index: 100
 eos_index: 101
 pad_index: 102
-label_smoothing: 0.2
-
-# Dynamic batching
-sorting: random
-num_workers: 4
-dynamic_batching: true
-max_batch_len: 80 # 40 GB GPU
-num_bucket: 200
-
-train_batch_size: 32 # if not using dynamic batching
-valid_batch_size: 1
-
-dynamic_batch_sampler:
-  max_batch_len: 80
-  num_buckets: 200
-  shuffle_ex: true   # if true re-creates batches at each epoch shuffling examples.
-  batch_ordering: random
-  max_batch_ex: 128
-
-train_dataloader_opts:
-  batch_size: 32
-  drop_last: false
-  num_workers: 4
-  collate_fn: !name:speechbrain.dataio.batch.PaddedBatch
-    padding_kwargs:
-      value: 102
-
-valid_dataloader_opts:
-  batch_size: 1
-  num_workers: 4
-  collate_fn: !name:speechbrain.dataio.batch.PaddedBatch
-    padding_kwargs:
-      value: 102
-
-################################
-# Model Parameters and model   #
-################################
-
-# Feature parameters (W2V2 etc)
-features_dim: 1024 # large wav2vec output dimension, for base replace by 768
+label_smoothing: 0.0
+
+# Encoder
+features_dim: 1024
 
 # Length Regulator
 enc_kernel_size: 3
 enc_stride: 2
 
-# Transformer
+# Transformer decoder
 embedding_size: 512
 d_model: 512
 nhead: 8
@@ -116,113 +28,73 @@ num_encoder_layers: 0
 num_decoder_layers: 6
 d_ffn: 2048
 transformer_dropout: 0.1
-activation: &id001 !name:torch.nn.GELU
-output_neurons: 103 # /!\ needs to be changed accordingly to the vocabulary
-attention_type: RelPosMHAXL   # "RelPosMHAXL" or "regularMHA"
+activation: !name:torch.nn.GELU
+output_neurons: !ref <vocab_size>
+attention_type: "RelPosMHAXL"
 
 # Decoding parameters
-test_bs: 10
 min_decode_ratio: 0.0
 max_decode_ratio: 1.0
 
-############################## models ################################
-wav2vec2: &id002 !new:speechbrain.lobes.models.huggingface_transformers.wav2vec2.Wav2Vec2
-  source: m3hrdadfi/wav2vec2-large-xlsr-persian-v3
-  output_norm: true   ### Test in baseline_v2
-  freeze: false
-  freeze_feature_extractor: false
-  save_path: results/s2ut/888/save/pretrained_models
-  apply_spec_augment: true
-
-enc: &id003 !new:speechbrain.nnet.CNN.Conv1d
-  input_shape: [null, null, 1024]
-  out_channels: 512
-  kernel_size: 3
-  stride: 2
-
-transformer: &id004 !new:speechbrain.lobes.models.transformer.TransformerST.TransformerST
-                                                                                   # yamllint disable-line rule:line-length
-  input_size: 512
-  tgt_vocab: 103
-  d_model: 512
-  nhead: 8
-  num_encoder_layers: 0
-  num_decoder_layers: 6
-  d_ffn: 2048
-  dropout: 0.1
-  activation: *id001
-  attention_type: RelPosMHAXL
-  normalize_before: true
-  causal: true
+wav2vec2: !new:speechbrain.lobes.models.huggingface_transformers.wav2vec2.Wav2Vec2
+    source: !ref <wav2vec2_hub>
+    output_norm: True
+    freeze: True
+    freeze_feature_extractor: True
+    apply_spec_augment : True
+    save_path: wav2vec2_checkpoints
+
+length_regulator: !new:speechbrain.nnet.CNN.Conv1d
+  input_shape: [null, null, !ref <features_dim>]
+  out_channels: !ref <embedding_size>
+  kernel_size: !ref <enc_kernel_size>
+  stride: !ref <enc_stride>
+
+transformer_decoder: !new:speechbrain.lobes.models.transformer.TransformerST.TransformerST # yamllint disable-line rule:line-length
+    input_size: !ref <embedding_size>
+    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>
+    attention_type: !ref <attention_type>
+    normalize_before: True
+    causal: False
 
 log_softmax: !new:speechbrain.nnet.activations.Softmax
-  apply_log: true
+    apply_log: True
+
+seq_lin: !new:speechbrain.nnet.linear.Linear
+    input_size: !ref <d_model>
+    n_neurons: !ref <output_neurons>
+
+model: !new:torch.nn.ModuleList
+    - [!ref <length_regulator>, !ref <transformer_decoder>, !ref <seq_lin>]
 
-seq_lin: &id005 !new:speechbrain.nnet.linear.Linear
+encoder: !new:speechbrain.nnet.containers.LengthsCapableSequential
+    wav2vec2: !ref <wav2vec2>
+    length_regulator: !ref <length_regulator>
 
-  input_size: 512
-  n_neurons: 103
+decoder_beamsearch: !new:speechbrain.decoders.seq2seq.S2STransformerBeamSearcher
+    modules: [!ref <transformer_decoder>, !ref <seq_lin>]
+    bos_index: !ref <bos_index>
+    eos_index: !ref <eos_index>
+    min_decode_ratio: !ref <min_decode_ratio>
+    max_decode_ratio: !ref <max_decode_ratio>
+    beam_size: 10
+    temperature: 1.0
 
 modules:
-  wav2vec2: *id002
-  enc: *id003
-  transformer: *id004
-  seq_lin: *id005
-model: &id006 !new:torch.nn.ModuleList
-- [*id003, *id004, *id005]
-opt_class: !name:torch.optim.AdamW
-  lr: 0.0005
-  betas: (0.9, 0.98)
-
-wav2vec_opt_class: !name:torch.optim.AdamW
-  lr: 0.00001
-
-seq_cost: !name:speechbrain.nnet.losses.nll_loss
-  label_smoothing: 0.2
-  reduction: batchmean
-
-noam_annealing: &id008 !new:speechbrain.nnet.schedulers.NoamScheduler
-  lr_initial: 0.0005
-  n_warmup_steps: 5000
-
-wav2vec_annealing: &id009 !new:speechbrain.nnet.schedulers.NewBobScheduler
-  initial_value: 0.00001
-  improvement_threshold: 0.0025
-  annealing_factor: 0.98
-
-#epoch object
-epoch_counter: &id007 !new:speechbrain.utils.epoch_loop.EpochCounter
-  limit: 100
-
-train_logger: !new:speechbrain.utils.train_logger.FileTrainLogger
-  save_file: results/s2ut/888/train_log.txt
-
-valid_search: !new:speechbrain.decoders.seq2seq.S2STransformerGreedySearcher
-  modules: [*id004, *id005, null]
-  bos_index: 100
-  eos_index: 101
-  min_decode_ratio: 0.0
-  max_decode_ratio: 1.0
-  temperature: 1.0
-
-test_search: !new:speechbrain.decoders.seq2seq.S2STransformerBeamSearcher
-  modules: [*id004, *id005]
-  bos_index: 100
-  eos_index: 101
-  min_decode_ratio: 0.0
-  max_decode_ratio: 1.0
-  beam_size: 10
-
-acc_computer: !name:speechbrain.utils.Accuracy.AccuracyStats
-bleu_computer: !name:speechbrain.utils.bleu.BLEUStats
-  merge_words: false
-
-#checkpointer
-checkpointer: !new:speechbrain.utils.checkpoints.Checkpointer
-  checkpoints_dir: results/s2ut/888/save
-  recoverables:
-    model: *id006
-    wav2vec2: *id002
-    counter: *id007
-    noam_scheduler: *id008
-    wav2vec_scheduler: *id009
+    encoder: !ref <encoder>
+    decoder: !ref <decoder_beamsearch>
+
+pretrainer: !new:speechbrain.utils.parameter_transfer.Pretrainer
+    loadables:
+      model: !ref <model>
+      wav2vec2: !ref <wav2vec2>
+    paths:
+        wav2vec2: !ref <pretrained_path>/wav2vec2.ckpt
+        model: !ref <pretrained_path>/model.ckpt