ResNeXt50, ResNeXt101

create_env.sh

git clone https://github.com/ocjosen/keras
cd keras/
git checkout 48875c680a1f22dc26ece8acb37178557af67a38
cd ../

git clone https://github.com/keras-team/keras-applications
cd keras-applications/
git checkout eaf8aed7e42ec568ac5760382f313df026826b1d
cd ../

pip install numpy==1.16 tensorflow==1.13.1 ./keras/ ./keras-applications/

save_converted_weights.py

import json

import keras
import numpy as np
from keras.applications import ResNeXt50, ResNeXt101


def save_converted_model(model_name, include_top, verify=True):
    keras.backend.clear_session()

    target_layer_suffix = '_2_reduce'
    channels_index = (
        3 if keras.backend.image_data_format() == 'channels_last' else 1)

    if model_name == 'resnext50':
        def build_fn():
            return ResNeXt50(weights='imagenet', include_top=include_top)
    elif model_name == 'resnext101':
        def build_fn():
            return ResNeXt101(weights='imagenet', include_top=include_top)
    else:
        raise ValueError(
            '`model_name` can be `resnext50` or `resnext101` only.')

    old_model = build_fn()
    config = json.loads(old_model.to_json())
    old_weights_list = [layer.get_weights() for layer in old_model.layers]
    del old_model
    keras.backend.clear_session()

    for layer_config in config['config']['layers']:
        if target_layer_suffix in layer_config['name']:
            layer_config['class_name'] = 'Conv2D'
            filters = np.array(
                layer_config['config']['function'][2][0]).shape[channels_index]
            layer_config['config'] = keras.layers.Conv2D(
                filters, 1, use_bias=False, trainable=False,
                name=layer_config['name']).get_config()
    new_model = keras.models.model_from_config(config)
    del config
    assert len(old_weights_list) == len(new_model.layers)

    for old_weights, new_layer in zip(old_weights_list, new_model.layers):
        if target_layer_suffix not in new_layer.name:
            new_layer.set_weights(old_weights)
        else:
            _, _, in_channels, out_channels = new_layer.get_weights()[0].shape
            filters = out_channels
            c = in_channels // out_channels
            kernel = np.zeros((1, 1, filters * c, filters), dtype='f')
            for i in range(filters):
                start = (i // c) * c * c + i % c
                end = start + c * c
                kernel[:, :, start:end:c, i] = 1
            new_layer.set_weights([kernel])
    del old_weights_list

    if include_top:
        file_name = model_name + '_weights_tf_dim_ordering_tf_kernels.h5'
    else:
        file_name = model_name + '_weights_tf_dim_ordering_tf_kernels_notop.h5'
    new_model.save_weights(file_name)

    if verify:
        np.random.seed(42)
        x = np.random.rand(1, 224, 224, 3).astype('f')

        new_outputs = new_model.predict(x)
        keras.backend.clear_session()

        old_outputs = build_fn().predict(x)
        keras.backend.clear_session()
        assert np.array_equal(old_outputs, new_outputs)


if __name__ == '__main__':
    for model_name in ['resnext50', 'resnext101']:
        for include_top in [True, False]:
            save_converted_model(model_name, include_top)