update

run_whisper_finetuning.py

@@ -92,7 +92,7 @@ class Seq2SeqTrainingArguments(TrainingArguments):
             )
         },
     )
-
+    xla: bool = field(default=False, metadata={"help": "Whether to activate the XLA compilation or not"})
 
 @dataclass
 class ModelArguments:

run_xla.sh

@@ -1,7 +1,7 @@
 # Whisper Finetuning script for the NST dataset
 # This is a test script for XLA on TPU
 
-python xla_spawn.py --num_cores=4 run_whisper_finetuining.py 
+python xla_spawn.py --num_cores=4 run_whisper_finetuning.py\
 	--model_name_or_path="openai/whisper-small" \
 	--output_dir="../whisper-NST-TPU" \
 	--overwrite_output_dir=True \
@@ -14,7 +14,7 @@ python xla_spawn.py --num_cores=4 run_whisper_finetuining.py
 	--audio_column_name="audio" \
 	--text_column_name="text" \
 	--per_device_train_batch_size=16 \
-    --per_device_train_batch_size=16 \
+    	--per_device_train_batch_size=16 \
 	--learning_rate=2e-5 \
 	--warmup_steps=500 \
 	--max_steps=5000 \
@@ -26,7 +26,6 @@ python xla_spawn.py --num_cores=4 run_whisper_finetuining.py
 	--eval_steps=1000 \
 	--max_eval_samples=100 \
 	--logging_steps=250 \
-	--fp16=True \
 	--load_best_model_at_end=True \
 	--metric_for_best_model="wer" \
 	--greater_is_better=False \
@@ -34,6 +33,7 @@ python xla_spawn.py --num_cores=4 run_whisper_finetuining.py
 	--predict_with_generate=True \
 	--generation_max_length=225 \
 	--print_training_arguments=True \
+	--xla=True \
 	--push_to_hub=True
 
 
@@ -44,4 +44,4 @@ python xla_spawn.py --num_cores=4 run_whisper_finetuining.py
     #     The zone of the TPU the process is running on. If not specified, we will attempt to automatically detect
     #     from metadata.
     # xla (:obj:`bool`, `optional`):
-    #     Whether to activate the XLA compilation or not.
+    #     Whether to activate the XLA compilation or not.

run_xla_test.sh

@@ -0,0 +1,47 @@
+# Whisper Finetuning script for the NST dataset
+# This is a test script for XLA on TPU
+
+PT_XLA_DEBUG=1 python xla_spawn.py --num_cores=4 run_whisper_finetuning.py\
+	--model_name_or_path="openai/whisper-small" \
+	--output_dir="../whisper-NST-TPU-test" \
+	--overwrite_output_dir=True \
+	--language="Norwegian" \
+	--task="transcribe" \
+	--dataset_name="NbAiLab/NST" \
+	--dataset_config="no-close" \
+	--do_train=True \
+	--do_eval=True \
+	--audio_column_name="audio" \
+	--text_column_name="text" \
+	--per_device_train_batch_size=16 \
+    	--per_device_train_batch_size=16 \
+	--learning_rate=2e-5 \
+	--warmup_steps=0 \
+	--max_steps=10 \
+	--gradient_checkpointing=True \
+	--gradient_accumulation_steps=1 \
+	--group_by_length=False \
+	--evaluation_strategy="steps" \
+	--save_steps=10 \
+	--eval_steps=10 \
+	--max_eval_samples=2 \
+	--logging_steps=10 \
+	--load_best_model_at_end=True \
+	--metric_for_best_model="wer" \
+	--greater_is_better=False \
+	--report_to="tensorboard" \
+	--predict_with_generate=True \
+	--generation_max_length=225 \
+	--print_training_arguments=True \
+	--xla=True \
+	--push_to_hub=True
+
+
+    # Very likely that some of this parameters needs to be added
+    # tpu_name (:obj:`str`, `optional`):
+    #     The name of the TPU the process is running on.
+    # tpu_zone (:obj:`str`, `optional`):
+    #     The zone of the TPU the process is running on. If not specified, we will attempt to automatically detect
+    #     from metadata.
+    # xla (:obj:`bool`, `optional`):
+    #     Whether to activate the XLA compilation or not.