File size: 1,261 Bytes
031abfe |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 |
python kyrgiz/run_speech_recognition_ctc.py \
--dataset_name="mozilla-foundation/common_voice_8_0" \
--model_name_or_path="facebook/wav2vec2-xls-r-300m" \
--dataset_config_name="ky" \
--train_split_name="train+validation[:50%]" \
--eval_split_name="validation[50%:]" \
--output_dir="./xls-r-kyrgiz-cv8" \
--overwrite_output_dir \
--num_train_epochs="50" \
--per_device_train_batch_size="16" \
--per_device_eval_batch_size="8" \
--gradient_accumulation_steps="4" \
--learning_rate="1e-4" \
--warmup_steps="250" \
--length_column_name="input_length" \
--evaluation_strategy="steps" \
--text_column_name="sentence" \
--chars_to_ignore , ? . ! \- \; \: \\ _ \| ‒ ☺ ♂ © « ¬ » \" „ “ % ” � — ’ ، ؛ ؟ ‹ › − … – \
--eval_metrics="wer" \
--save_steps="500" \
--eval_steps="500" \
--logging_steps="100" \
--min_duration_in_seconds="0.2" \
--layerdrop="0.01" \
--activation_dropout="0.1" \
--save_total_limit="3" \
--freeze_feature_encoder \
--feat_proj_dropout="0.01" \
--mask_time_prob="0.50" \
--mask_time_length="10" \
--mask_feature_prob="0.25" \
--mask_feature_length="64" \
--gradient_checkpointing \
--use_auth_token \
--fp16 \
--group_by_length \
--do_train --do_eval \
--push_to_hub
|