librispeech 960h

run_librispeech.sh

@@ -2,23 +2,25 @@
 python run_flax_speech_recognition_seq2seq.py \
         --dataset_name="librispeech_asr" \
         --model_name_or_path="./" \
-        --dataset_config_name="clean" \
-        --train_split_name="train.100" \
-        --eval_split_name="validation" \
-	--dataset_cache_dir="~/cache/huggingface/datasets" \
-        --output_dir="./" \
+        --dataset_config_name="all" \
+        --train_split_name="train.clean.100+train.clean.360+train.other.500" \
+        --eval_split_name="validation.clean" \
+	--dataset_cache_dir="/home/sanchitgandhi/cache/huggingface/datasets" \
+        --output_dir="./output_dir" \
         --preprocessing_num_workers="16" \
         --length_column_name="input_length" \
         --overwrite_output_dir \
-        --num_train_epochs="10" \
-        --per_device_train_batch_size="4" \
-        --per_device_eval_batch_size="4" \
-	--gradient_accumulation_steps="1" \
+        --num_train_epochs="3" \
+        --per_device_train_batch_size="8" \
+        --per_device_eval_batch_size="2" \
         --logging_steps="25" \
-        --max_duration_in_seconds="10" \
-        --max_target_length="64" \
+	--gradient_checkpointing \
+        --max_duration_in_seconds="20" \
+        --max_target_length="128" \
         --generation_max_length="40" \
         --generation_num_beams="1" \
+	--final_generation_max_length="200" \
+	--final_generation_num_beams="5" \
         --learning_rate="3e-4" \
         --warmup_steps="500" \
         --text_column_name="text" \
@@ -28,7 +30,5 @@ python run_flax_speech_recognition_seq2seq.py \
         --do_lower_case \
         --do_eval \
         --do_train \
-        --push_to_hub \
-        --use_auth_token \
-        --wandb_project="flax-wav2vec2-2-bart-large"	
+        --wandb_project="flax-wav2vec2-2-bart-large-960h"