TensorFlow/Keras: How to get missing models (ResNet101, ResNeXt, etc.) from Keras' applications module?

A number of documented Keras applications are missing from my (up-to-date) Keras installation and TensorFlow 1.10 Keras API installation. I import Keras' applications module as suggested and use it as follows:

from keras import applications
resnet = applications.ResNeXt101(include_top=False, weights='imagenet', input_shape=(SCALED_HEIGHT, SCALED_WIDTH, 3), pooling=None)

I also tried

resnet = applications.resnext.ResNeXt101(include_top=False, weights='imagenet', input_shape=(SCALED_HEIGHT, SCALED_WIDTH, 3), pooling=None)

But in both cases I get the same type of error:

AttributeError: module 'keras.applications' has no attribute 'ResNeXt101'

Printing help(applications) yields:

Help on package keras.applications in keras:

NAME
    keras.applications

PACKAGE CONTENTS
    densenet
    imagenet_utils
    inception_resnet_v2
    inception_v3
    mobilenet
    mobilenet_v2
    mobilenetv2
    nasnet
    resnet50
    vgg16
    vgg19
    xception

FUNCTIONS
    DenseNet121 = wrapper(*args, **kwargs)
    DenseNet169 = wrapper(*args, **kwargs)
    DenseNet201 = wrapper(*args, **kwargs)
    InceptionResNetV2 = wrapper(*args, **kwargs)
    InceptionV3 = wrapper(*args, **kwargs)
    MobileNet = wrapper(*args, **kwargs)
    MobileNetV2 = wrapper(*args, **kwargs)
    NASNetLarge = wrapper(*args, **kwargs)
    NASNetMobile = wrapper(*args, **kwargs)
    ResNet50 = wrapper(*args, **kwargs)
    VGG16 = wrapper(*args, **kwargs)
    VGG19 = wrapper(*args, **kwargs)
    Xception = wrapper(*args, **kwargs)
    keras_modules_injection(base_fun)

which shows that the models initially aren't present in my installation. Why not? They're also not packaged in the TensorFlow's Keras API.

I tried copying the files from the Keras applications GitHub repository and pasting them in site-packages/keras/applications/, but this results in the following stacktrace:

File "myscript.py", line 517, in get_fpn
    resnet = applications.resnext.ResNeXt101(include_top=False, weights='imagenet', input_shape=(SCALED_HEIGHT, SCALED_WIDTH, 3), pooling=None)
  File "site-packages/keras/applications/resnet_common.py", line 575, in ResNeXt101
    **kwargs)
  File "site-packages/keras/applications/resnet_common.py", line 348, in ResNet
    data_format=backend.image_data_format(),
AttributeError: 'NoneType' object has no attribute 'image_data_format'

Any ideas on how to fix this? Why aren't these included and working in default installations of either Keras or TensorFlow? Why does the documentation not explain this?

Solution

Reason for the problem:

The backend object is None at line 348. My guess is you tried something like this:

>>> from keras_applications import resnext
>>> resnext.ResNeXt101(weights=None)

The backend information is injected from keras.applications to keras_applications via the keras_modules_injection decorator.

https://github.com/keras-team/keras/blob/c658993cf596fbd39cf800873bc457e69cfb0cdb/keras/applications/resnext.py#L17

Procedure to fix the problem:

Make sure the keras & keras applications versions are as follows:

>>pip list |grep Keras
Keras                  2.2.4
Keras-Applications     1.0.8

If they're not, upgrade using

>>pip install --upgrade keras keras-applications

Update the changes from this pull request, https://github.com/keras-team/keras/pull/11203/files into site-packages/keras/applications

from keras import applications
resnext = applications.resnext.ResNeXt101(include_top=False, weights=None, input_shape=(299,299,3))
print(type(resnext))