up

run.sh

@@ -0,0 +1,35 @@
+#!/usr/bin/env bash
+# python -m torch.distributed.launch \
+#        --nproc_per_node 2 run_speech_recognition_ctc.py \
+CUDA_VISIBLE_DEVICES="0" python run_speech_recognition_ctc.py \
+--dataset_name="multilingual_librispeech" \
+--model_name_or_path="facebook/wav2vec2-xls-r-300m" \
+--train_split_name="train.9h" \
+--dataset_config_name="german" \
+--output_dir="./wav2vec2-300m-mls-german-ft" \
+--overwrite_output_dir \
+--num_train_epochs="100" \
+--per_device_train_batch_size="32" \
+--gradient_accumulation_steps="1" \
+--learning_rate="1e-4" \
+--activation_dropout="0.1" \
+--warmup_steps="1000" \
+--evaluation_strategy="steps" \
+--text_column_name="text" \
+--save_steps="1000" \
+--eval_steps="500" \
+--logging_steps="10" \
+--layerdrop="0.1" \
+--hidden_dropout="0.0" \
+--save_total_limit="1" \
+--mask_time_prob="0.75" \
+--mask_feature_prob="0.25" \
+--mask_feature_length="64" \
+--freeze_feature_extractor \
+--fp16 \
+--group_by_length \
+--do_eval \
+--gradient_checkpointing \
+--use_auth_token \
+--do_train --do_eval \
+--push_to_hub \