3

Using Keras on Google cloud ml:

Saving model from training:

model.save('model.h5')

if cfg.cloud:
    # Copy model.h5 over to Google Cloud Storage
    with file_io.FileIO('model.h5', mode='rb') as input_f:
        with file_io.FileIO(data_folder + 'model.h5', mode='wb+') as output_f:
            output_f.write(input_f.read())

Note: not saving to job_folder, I need to read this later, not wanting to keep track of the latest job (even if that's a good way to keep models apart).

Now I want to read from my next run:

f = file_io.FileIO(model_file, mode='rb')
model = load_model(f)
model.load_weights(f)

Where 'model_file' is given as input in my submission, pointing at 

--model-file gs://$BUCKET_NAME/resources/model.h5

Complains from google cloud ml jobs:

TypeError: expected str, bytes or os.PathLike object, not FileIO

I tried a number of things, but my basic question is: what's the best practice writing and especially reading in models from gcp buckets?

VictorGGl
  • 1,848
  • 10
  • 15
user2427317
  • 111
  • 1
  • 9
  • Please clarify where in code this exception occurs, is it in your code at your load_model(f) or is it in sdk code? – Oliver Apr 22 '18 at 22:00
  • Thanks Oliver, I did some debugging to answer your question in more detail, and while doing that I found the solution below. – user2427317 Apr 23 '18 at 14:11

1 Answers1

4

Finally I got it working, using the solution here:

loading saved keras model from gs to pydatalab

Thanks Tíarnán McGrath (I don't have enough points to do a plus 1)

user2427317
  • 111
  • 1
  • 9