Convert a Keras weight checkpoint train on multigpu to a regular Keras weight checkpoint

weights_converter.py

import os
from collections import defaultdict

import h5py
import numpy as np

pjoin = os.path.join


def convert_multi_to_single(multipath, fp):
    """
    Convert a h5py weight file trained on a multigpu to a single-gpu.
    Notes:
        Tested for Xception
    Args:
        multipath: h5 filepath of the multigpu model
        fp: h5 filepath of the output
    """
    f = h5py.File(multipath, 'r')
    if 'model_1' not in f:
        raise ValueError("`multipath` doesn't have a subnetwork called `'model_1'`.")
    ref = f['model_1']
    out = h5py.File(fp, 'w')

    layer_name = []
    weight_names = defaultdict(list)
    # Find all the weights name
    for w in ref.attrs['weight_names']:
        weight_names[w.decode('utf').split('/')[0]].append(w)

    # Assign the weight and the weights name to all layers.
    for k, v in ref.items():
        name = k.split('/')[-1]
        layer_name.append(str.encode(name))
        for k1, v1 in v.items():
            out[pjoin(k, k, k1)] = v1.value
        out[name].attrs['weight_names'] = weight_names[name]

    # Set metadata
    out.attrs['keras_version'] = f.attrs['keras_version']
    out.attrs['backend'] = str.encode('tensorflow')
    out.attrs['layer_names'] = np.array(layer_name)
    f.close()
    out.close()