up

run_tedlium.sh

@@ -0,0 +1,40 @@
+#!/usr/bin/env bash
+python run_flax_speech_recognition_seq2seq.py \
+        --dataset_name="LIUM/tedlium" \
+        --model_name_or_path="sanchit-gandhi/flax-wav2vec2-2-bart-large-scan" \
+        --dataset_config_name="release3" \
+        --train_split_name="train" \
+        --eval_split_name="validation" \
+        --test_split_name="test" \
+        --text_column_name="text" \
+        --id_column_name="id" \
+        --output_dir="./flax-wav2vec2-2-bart-large-tedlium-hidden-activation-featproj-dropout-0.3" \
+        --wandb_project="tedlium" \
+        --wandb_name="flax-wav2vec2-2-bart-large-tedlium-hidden-activation-featproj-dropout-0.3" \
+        --dataset_cache_dir="/home/sanchitgandhi/cache/huggingface/datasets" \
+        --per_device_train_batch_size="8" \
+        --per_device_eval_batch_size="4" \
+        --logging_steps="25" \
+        --max_steps="50000" \
+        --eval_steps="10000" \
+        --save_steps="10000" \
+        --generation_max_length="40" \
+        --generation_num_beams="1" \
+        --generation_length_penalty="1.2" \
+        --final_generation_max_length="200" \
+        --final_generation_num_beams="5" \
+        --learning_rate="1e-4" \
+        --warmup_steps="500" \
+        --overwrite_output_dir \
+        --gradient_checkpointing \
+        --freeze_feature_encoder \
+        --predict_with_generate \
+        --do_lower_case \
+        --do_eval \
+        --do_train \
+        --do_predict \
+        --push_to_hub \
+        --use_auth_token \
+	--hidden_dropout="0.3" \
+	--activation_dropout="0.3" \
+	--feat_proj_dropout="0.3"