Use return_tensors="np" instead of "tf" (#21266)
Return NP instead of TF tensors for our data loading pipeline
This commit is contained in:
@@ -499,7 +499,7 @@ def main():
|
||||
model=model,
|
||||
label_pad_token_id=label_pad_token_id,
|
||||
pad_to_multiple_of=64, # Reduce the number of unique shapes for XLA, especially for generation
|
||||
return_tensors="tf",
|
||||
return_tensors="np",
|
||||
)
|
||||
num_replicas = training_args.strategy.num_replicas_in_sync
|
||||
total_train_batch_size = training_args.per_device_train_batch_size * num_replicas
|
||||
|
||||
Reference in New Issue
Block a user