Keras: Write ModelCheckpoint callback to GCS (Google Cloud Storage)

[sakinaljana]

Dear Keras Folks,

I'm understand that we cannot directly write checkpoint model to GCS because the inability of h5py library to support write to GCS. Is there any workaroud to mitigate this problem? i suggest to implement this:

• first, write the checkpoint model locally,
• then read the saved model using tensorflow.python.lib.io.file_io.FileIO from local path, then write again using same FileIO directly to GCS.

What do you think guys?

[fchollet]

Yes, we can add such a feature to model.save(), save_weights, load_model
and load_weights. Unfortunately there's nothing better than this workaround
at this point. Feel free to open a PR

On Sep 19, 2017 19:38, "dyngts" [e-mail] wrote:

Dear Keras Folks,

I'm understand that we cannot directly write checkpoint model to GCS
because the inability of h5py library to support write to GCS. Is there any
workaroud to mitigate this problem? i suggest to implement this:

• first, write the checkpoint model locally,
• then read the saved model using tensorflow.python.lib.io.file_io.FileIO
  from local path, then write again using same FileIO directly to GCS.

What do you think guys?

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
https://github.com/fchollet/keras/issues/7935, or mute the thread
https://github.com/notifications/unsubscribe-auth/AArWb6-E1KQta7r3WvmO3cR-XSlG2VUdks5skHqagaJpZM4PdQ30
.

[sakinaljana]

i see. thanks @fchollet for the confirmation.

[Harshini-Gadige]

@dyngts Is there any PR opened for this ?

[andhus]

@harshini-gadige (@dyngts) I ran into the same problem, please see: https://github.com/keras-team/keras/pull/11636