The Inception-ResNet v2 model using Keras (with weight files)

Tested with tensorflow-gpu==1.2.1 and Keras==2.0.6

Layers and namings follows the TF-slim implementation: https://github.com/tensorflow/models/blob/master/slim/nets/inception_resnet_v2.py

Basically the same with the keras.applications.InceptionV3 model.

from inception_resnet_v2 import InceptionResNetV2

# ImageNet classification
model = InceptionResNetV2()
model.predict(...)

# Finetuning on another 100-class dataset
base_model = InceptionResNetV2(include_top=False, pooling='avg')
outputs = Dense(100, activation='softmax')(base_model.output)
model = Model(base_model.inputs, outputs)
model.compile(...)
model.fit(...)

python extract_weights.py

By default, the TF checkpoint file will be downloaded to ./models folder, and the layer weights (.npy files) will be saved to ./weights folder.

python load_weights.py

The following weight files:

• models/inception_resnet_v2_weights_tf_dim_ordering_tf_kernels.h5
• models/inception_resnet_v2_weights_tf_dim_ordering_tf_kernels_notop.h5

will be generated.

To test whether this implementation gives the same prediction as TF-slim implementation:

PYTHONPATH=../tensorflow-models/slim python test_inception_resnet_v2.py

PYTHONPATH should point to the slim folder under the https://github.com/tensorflow/models repo.

The image file elephant.jpg (and basically the entire idea of converting weights from TF-slim to Keras) comes from: https://github.com/kentsommer/keras-inception-resnetV2

• Extract weights from TF-slim
• Convert weights to HDF5 files
• Test weight loading and image prediction (elephant.jpg)
• Release weight files
• Test accuracy on benchmark datasets