Prepare the runner

config.json

@@ -0,0 +1,85 @@
+{
+  "activation_dropout": 0.0,
+  "apply_spec_augment": true,
+  "architectures": [
+    "Wav2Vec2ForPreTraining"
+  ],
+  "attention_dropout": 0.1,
+  "bos_token_id": 1,
+  "codevector_dim": 256,
+  "contrastive_logits_temperature": 0.1,
+  "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": "sum",
+  "ctc_zero_infinity": false,
+  "diversity_loss_weight": 0.1,
+  "do_stable_layer_norm": true,
+  "eos_token_id": 2,
+  "feat_extract_activation": "gelu",
+  "feat_extract_norm": "layer",
+  "feat_proj_dropout": 0.1,
+  "feat_quantizer_dropout": 0.0,
+  "final_dropout": 0.0,
+  "freeze_feat_extract_train": true,
+  "gradient_checkpointing": true,
+  "hidden_act": "gelu",
+  "hidden_dropout": 0.1,
+  "hidden_size": 768,
+  "initializer_range": 0.02,
+  "intermediate_size": 3072,
+  "layer_norm_eps": 1e-05,
+  "layerdrop": 0.05,
+  "mask_channel_length": 10,
+  "mask_channel_min_space": 1,
+  "mask_channel_other": 0.0,
+  "mask_channel_prob": 0.0,
+  "mask_channel_selection": "static",
+  "mask_feature_length": 10,
+  "mask_feature_prob": 0.0,
+  "mask_time_length": 10,
+  "mask_time_min_space": 1,
+  "mask_time_other": 0.0,
+  "mask_time_prob": 0.05,
+  "mask_time_selection": "static",
+  "model_type": "wav2vec2",
+  "no_mask_channel_overlap": false,
+  "no_mask_time_overlap": false,
+  "num_attention_heads": 12,
+  "num_codevector_groups": 2,
+  "num_codevectors_per_group": 320,
+  "num_conv_pos_embedding_groups": 16,
+  "num_conv_pos_embeddings": 128,
+  "num_feat_extract_layers": 7,
+  "num_hidden_layers": 12,
+  "num_negatives": 100,
+  "pad_token_id": 0,
+  "proj_codevector_dim": 256,
+  "transformers_version": "4.9.0.dev0",
+  "vocab_size": 40
+}

preprocessor_config.json

@@ -0,0 +1,9 @@
+{
+  "do_normalize": true,
+  "feature_extractor_type": "Wav2Vec2FeatureExtractor",
+  "feature_size": 1,
+  "padding_side": "right",
+  "padding_value": 0.0,
+  "return_attention_mask": true,
+  "sampling_rate": 16000
+}

src/run_config.sh

@@ -0,0 +1,13 @@
+#!/bin/bash
+export LC_ALL=C.UTF-8
+export LANG=C.UTF-8
+
+
+# export OUTPUT_DIR=./
+export OUTPUT_DIR=/home/m3hrdadfi/code/wav2vec2-base-persian
+export NAME_OR_PATH=facebook/wav2vec2-base
+
+
+python src/run_config.py \
+    --output_dir="$OUTPUT_DIR"  \
+    --name_or_path="$NAME_OR_PATH"

src/run_persian.sh

@@ -4,22 +4,23 @@ export LC_ALL=C.UTF-8
 export LANG=C.UTF-8
 
 export OUTPUT_DIR=/home/m3hrdadfi/code/wav2vec2-base-persian
+export OUTPUT_DIR=/home/m3hrdadfi/data_cache/
 export MODEL_NAME_OR_PATH=/home/m3hrdadfi/code/wav2vec2-base-persian
 
 
-export TRAIN_FILE=/home/m3hrdadfi/code/data/train.csv
-export VALIDATION_FILE=/home/m3hrdadfi/code/data/test.csv
+export TRAIN_FILE=/home/m3hrdadfi/data/fa/train.csv
+export VALIDATION_FILE=/home/m3hrdadfi/data/fa/test.csv
 export SPEECH_FILE_COLUMN=path
 
 
 #export MAX_EVAL_SAMPLES=5000
-export PER_DEVICE_TRAIN_BATCH_SIZE=32
-export PER_DEVICE_EVAL_BATCH_SIZE=32
+export PER_DEVICE_TRAIN_BATCH_SIZE=8
+export PER_DEVICE_EVAL_BATCH_SIZE=8
 #export GRADIENT_ACCUMULATION_STEPS=2
 export NUM_TRAIN_EPOCHS=5.0
 export LEARNING_RATE=5e-4
 export WARMUP_STEPS=1000
-#export LOGGING_STEPS=500
+export LOGGING_STEPS=500
 #export EVAL_STEPS=2500
 #export SAVE_STEPS=2500
 export PREPROCESSING_NUM_WORKERS=4
@@ -32,6 +33,7 @@ export PAD_TO_MULTIPLE_OF=16384
 
 python src/run_wav2vec2_pretrain_flax.py \
     --output_dir="$OUTPUT_DIR"  \
+    --cache_dir="$CACHE_DIR" \
     --train_file="$TRAIN_FILE" \
     --validation_file="$VALIDATION_FILE" \
     --speech_file_column="$SPEECH_FILE_COLUMN" \
@@ -48,4 +50,5 @@ python src/run_wav2vec2_pretrain_flax.py \
     --adam_beta2=$ADAM_BETA_2 \
     --dtype="$D_TYPE" \
     --pad_to_multiple_of=$PAD_TO_MULTIPLE_OF \
+    --logging_steps=$LOGGING_STEPS \
     --push_to_hub

src/run_wav2vec2_pretrain_flax.py

@@ -349,9 +349,14 @@ def main():
         do_normalize=True
     )
 
+    target_sampling_rate = 16_000
     def prepare_dataset(batch):
         # check that all files have the correct sampling rate
-        batch["speech"], _ = librosa.load(batch[data_args.speech_file_column], sr=feature_extractor.sampling_rate)
+        # batch["speech"], _ = librosa.load(batch[data_args.speech_file_column], sr=feature_extractor.sampling_rate)
+        speech_array, sampling_rate = torchaudio.load(batch["path"])
+        resampler = torchaudio.transforms.Resample(sampling_rate, target_sampling_rate)
+        batch["speech"] = resampler(speech_array).squeeze().numpy()
+
         return batch
 
     # load audio files into numpy arrays
@@ -377,6 +382,7 @@ def main():
         load_from_cache_file=not data_args.overwrite_cache,
         remove_columns=vectorized_datasets["train"].column_names,
     )
+    vectorized_datasets.save_to_disk(model_args.cache_dir)
 
     # pretraining is only supported for "newer" stable layer norm architecture
     # apply_spec_augment has to be True, mask_feature_prob has to be 0.0