fix language hard coding

run_speech_recognition_seq2seq_streaming.py

@@ -49,7 +49,7 @@ from transformers import (
     set_seed,
 )
 from transformers.models.whisper.english_normalizer import BasicTextNormalizer
-from transformers.models.whisper.tokenization_whisper import TO_LANGUAGE_CODE
+from transformers.models.whisper.tokenization_whisper import TO_LANGUAGE_CODE, LANGUAGES
 from transformers.trainer_pt_utils import IterableDatasetShard
 from transformers.trainer_utils import get_last_checkpoint, is_main_process
 from transformers.utils import check_min_version, send_example_telemetry
@@ -843,8 +843,9 @@ def main():
     pred_labels = [f"Prediction: {pred}\nLabel: {label}\n" for pred, label in zip(preds, labels)]
     logger.info("Before setting language and task")
     logger.info(f"{pred_labels}")
+    language_name = LANGUAGES[data_args.language_eval]
     trainer.model.config.forced_decoder_ids = \
-        tokenizer.get_decoder_prompt_ids(language=data_args.language_eval, task=data_args.task, no_timestamps=True)
+        tokenizer.get_decoder_prompt_ids(language_name, task=data_args.task, no_timestamps=True)
     preds = tokenizer.batch_decode(predictions.predictions)
     labels = tokenizer.batch_decode(predictions.label_ids)
     pred_labels = [f"Prediction: {pred}\nLabel: {label}\n" for pred, label in zip(preds, labels)]

test_run_nordic.sh

@@ -9,19 +9,17 @@ python $1run_speech_recognition_seq2seq_streaming.py \
 	--language_eval="sv" \
 	--eval_split_name="test" \
 	--model_index_name="Whisper Tiny Nordic" \
-	--max_train_samples="64" \
-	--max_eval_samples="32" \
-	--max_steps="1" \
+	--max_steps="500" \
 	--output_dir="./" \
-	--per_device_train_batch_size="1" \
-	--per_device_eval_batch_size="1" \
+	--per_device_train_batch_size="128" \
+	--per_device_eval_batch_size="64" \
 	--logging_steps="25" \
 	--learning_rate="1e-5" \
-	--warmup_steps="500" \
+	--warmup_steps="50" \
 	--evaluation_strategy="steps" \
-	--eval_steps="1000" \
+	--eval_steps="100" \
 	--save_strategy="steps" \
-	--save_steps="1000" \
+	--save_steps="100" \
 	--generation_max_length="225" \
 	--length_column_name="input_length" \
 	--max_duration_in_seconds="30" \