config.json

config.json

@@ -0,0 +1,80 @@
+{
+  "_name_or_path": "naver/mHuBERT-147-ASR-fr",
+  "activation_dropout": 0.1,
+  "add_interface": true,
+  "apply_spec_augment": true,
+  "architectures": [
+    "MHubertForCTC"
+  ],
+  "attention_dropout": 0.1,
+  "bos_token_id": 1,
+  "classifier_proj_size": 256,
+  "conv_bias": false,
+  "conv_dim": [
+    512,
+    512,
+    512,
+    512,
+    512,
+    512,
+    512
+  ],
+  "conv_kernel": [
+    10,
+    3,
+    3,
+    3,
+    3,
+    2,
+    2
+  ],
+  "conv_stride": [
+    5,
+    2,
+    2,
+    2,
+    2,
+    2,
+    2
+  ],
+  "ctc_loss_reduction": "mean",
+  "ctc_token_id": 174,
+  "ctc_zero_infinity": false,
+  "do_stable_layer_norm": false,
+  "eos_token_id": 2,
+  "feat_extract_activation": "gelu",
+  "feat_extract_dropout": 0.0,
+  "feat_extract_norm": "group",
+  "feat_proj_dropout": 0.1,
+  "feat_proj_layer_norm": true,
+  "final_dropout": 0.3,
+  "fine_tuning_strategy": "all",
+  "gradient_checkpointing": false,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.1,
+  "hidden_dropout_prob": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.1,
+  "mask_feature_length": 10,
+  "mask_feature_min_masks": 0,
+  "mask_feature_prob": 0.0,
+  "mask_time_length": 10,
+  "mask_time_min_masks": 2,
+  "mask_time_prob": 0.05,
+  "model_type": "hubert",
+  "num_attention_heads": 12,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 12,
+  "num_interface_layers": 3,
+  "pad_token_id": 173,
+  "tokenizer_class": "Wav2Vec2CTCTokenizer",
+  "torch_dtype": "float32",
+  "transformers_version": "4.32.0",
+  "use_weighted_layer_sum": false,
+  "vocab_size": 175
+}

config.yaml

@@ -1,24 +0,0 @@
-group_by_length: True
-evaluation_strategy: "steps"
-num_train_epochs: 100
-fp16: False
-gradient_checkpointing: True
-eval_steps: 10000
-save_steps: 10000
-logging_steps: 10000
-learning_rate: 1e-4
-adam_beta1: 0.9
-adam_beta2: 0.98
-adam_epsilon: 1e-08
-warmup_ratio: 0.2
-save_total_limit: 4
-load_best_model_at_end: True
-per_device_train_batch_size: 8
-per_device_eval_batch_size: 2
-metric_for_best_model: "cer"
-greater_is_better: False
-gradient_accumulation_steps: 8
-final_dropout: 0.3
-seed: 3452
-add_interface_layer: True
-num_interface_layers: 3

inference_code/run_inference.py

@@ -16,23 +16,12 @@ fbk_test_id = 'FBK-MT/Speech-MASSIVE-test'
 mhubert_id = 'utter-project/mHuBERT-147'
 
 def load_asr_model():
-	def init_config():
-		config = HubertConfig.from_pretrained(mhubert_id)
-		config.pad_token_id = processor.tokenizer.pad_token_id
-		config.ctc_token_id = processor.tokenizer.convert_tokens_to_ids('[CTC]')
-		config.vocab_size = len(processor.tokenizer)
-
-		config.output_hidden_states = False
-		config.add_interface = True
-		config.num_interface_layers = 3
-		return config
-
 	# Load the ASR model
 	tokenizer = Wav2Vec2CTCTokenizer('vocab.json', unk_token="[UNK]", pad_token="[PAD]", word_delimiter_token="|")
 	feature_extractor = Wav2Vec2FeatureExtractor.from_pretrained(mhubert_id)
 	processor = Wav2Vec2Processor(feature_extractor=feature_extractor, tokenizer=tokenizer)
 
-	config = init_config()
+	config = HubertConfig.from_pretrained('config.json')
 	model = mHubertForCTC.from_pretrained("naver/mHuBERT-147-ASR-fr", config=config)
 	model.eval()
 	return model, processor