marinone94
commited on
Commit
•
4123391
1
Parent(s):
3644968
delete checkpoint at the end
Browse files- huggingface_training.py +8 -1
huggingface_training.py
CHANGED
@@ -282,8 +282,11 @@ training_args = Seq2SeqTrainingArguments(
|
|
282 |
logging_steps=logging_steps,
|
283 |
save_strategy="steps",
|
284 |
save_steps=eval_steps,
|
285 |
-
save_total_limit=2,
|
286 |
learning_rate=1e-5,
|
|
|
|
|
|
|
287 |
warmup_ratio=0.5 if test_script is True else 0.2,
|
288 |
per_device_train_batch_size=train_bs,
|
289 |
per_device_eval_batch_size=eval_bs,
|
@@ -370,4 +373,8 @@ print(final_metrics)
|
|
370 |
|
371 |
# Pushing to hub during training slows down training
|
372 |
# so we push it only in the end.
|
|
|
|
|
|
|
|
|
373 |
trainer.push_to_hub()
|
|
|
282 |
logging_steps=logging_steps,
|
283 |
save_strategy="steps",
|
284 |
save_steps=eval_steps,
|
285 |
+
# save_total_limit=2,
|
286 |
learning_rate=1e-5,
|
287 |
+
load_best_model_at_end=True,
|
288 |
+
metric_for_best_model="wer",
|
289 |
+
greater_is_better=False,
|
290 |
warmup_ratio=0.5 if test_script is True else 0.2,
|
291 |
per_device_train_batch_size=train_bs,
|
292 |
per_device_eval_batch_size=eval_bs,
|
|
|
373 |
|
374 |
# Pushing to hub during training slows down training
|
375 |
# so we push it only in the end.
|
376 |
+
# Since training is completed and best model has been saved, we first delete the checkpoints
|
377 |
+
for filename in os.listdir("."):
|
378 |
+
if filename.startswith("checkpoint-"):
|
379 |
+
os.remove(f"./{filename}")
|
380 |
trainer.push_to_hub()
|