Is it possible to decode file on GPU?

TensorFlow General Discussion
, ,
1

Is it possible to decode file on GPU during training a model? Resizing, Rescaling etc can be done as part of a model.

with tf.device('/GPU:0'):
   tf.io.decode_*

model = Sequential(
     [
        ImageReader(), 
        ImageResizer(), 
        ImageNetModel()
        ...
     ]
)

Reference. NVIDIA Developer Data Loading Library (DALI) | NVIDIA Developer

2

I don’t think that currently we have a GPU decoding.

We had a thread with preprocessing + decoding at:
https://github.com/keras-team/keras-cv/pull/146#issuecomment-1048128659

We have also discussed something for Video:

https://github.com/tensorflow/io/issues/840

Another emerging approach is:
RGB no more: Minimally-decoded JPEG Vision Transformers

2 Likes
4

Does nvidia dali suit your usecase? I have not used it but could be something.

Also if using distributed strategies see these experimental options:

@tf_export("distribute.InputOptions", v1=[])
class InputOptions(
    collections.namedtuple("InputOptions", [
        "experimental_fetch_to_device",
        "experimental_replication_mode",
        "experimental_place_dataset_on_device",
        "experimental_per_replica_buffer_size",
    ])):
...

Perhaps experimental_place_dataset_on_device does what you are looking for