loading saved keras model from gs to pydatalab

问题:

My keras model is saved in google storage with model.save(model_name)

I cannot load the model on pydatalab. When I save the model on my local machine, I can just open it with load_model(filepath). Also I did import keras.backend as K, based on NameError when opening Keras model that uses Tensorflow Backend

I have tried the following:

1. model = load_model(tf.gfile.Open(model_file))
   

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

2. load_model('gs://mybucket/model.h5')
   

Error: IOError: Unable to open file (unable to open file: name = 'gs://mybucket/model.h5', errno = 2, error message = 'No such file or directory', flags = 0, o_flags = 0)

3. with file_io.FileIO(model_file, 'r') as f:
   modl = load_model(f)
   

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

回答1:

Load the file from gs storage

from tensorflow.python.lib.io import file_io
model_file = file_io.FileIO('gs://mybucket/model.h5', mode='rb')

Save a temporary copy of the model locally

temp_model_location = './temp_model.h5'
temp_model_file = open(temp_model_location, 'wb')
temp_model_file.write(model_file.read())
temp_model_file.close()
model_file.close()

Load model saved locally

model = load_model(temp_model_location)

回答2:

I don't think Keras supports the TensorFlow file system which in turn knows how to read from GCS.

You could try downloading from GCS to a local path, and then reading from that to load the model.

回答3:

The following function works for retraining an already trained keras model (with new data) on gcloud machine learning platform (Thanks to Tíarnán McGrath).

def load_models(model_file):

    model = conv2d_model() #the architecture of my model, not compiled yet
    file_stream = file_io.FileIO(model_file, mode='r')
    temp_model_location = './temp_model.h5'
    temp_model_file = open(temp_model_location, 'wb')
    temp_model_file.write(file_stream.read())
    temp_model_file.close()
    file_stream.close()
    model.load_weights(temp_model_location)

    return model

For some reason, load_model from keras.models does not work for me anymore, so I have to build the model each time.