pip install Keras-Applications-3D

### Prepare your 3D data [Ex. (64,64,64,1) shape]
# ...
X = np.zeros((128, 64, 64, 64, 1))
y = np.zeros((128, 10))
y[:,0] = 1
# ...

### Oriinal channel of model is '64->128->256->...'
### But it takes a lot of memory, So we reduce first channel 64 to 16
model = vgg19.VGG19(
  input_shape=(64,64,64,1), classes=10,
  base_channel=16
)
# model.summary()

### Training
model.compile(
    loss='categorical_crossentropy', 
    optimizer=Adam(learning_rate=1e-4),
    metrics=['acc']
)

history = model.fit(
    x=X, y=y, 
)

### Inference
pred = model.predict(X).squeeze()
real = y

pred = pred.argmax(axis=1)
real = real.argmax(axis=1)

accr = (pred == real).sum() / len(real)
print('Accuracy: {:04f}'.format(accr))

• Major 2D CNN architecture for 3D ([V]: Complete, []: In progress)
  • VGG (16, 19)
  • ResNet (50, 101, 152)
  • ResNetV2 (50, 101, 152)
  • DenseNet (121, 169, 201)
  • ResNext
  • InceptionV3
  • Inception_Resnet_V2
  • Xception
  • EfficientNet (B0, B1, ..., B7)
  • Mobilenet (V1, V2)
  • SE-ResNet
  • NFNet
• Convolution function for 3D (keras_applications_3d/custom_layers.py)
  • DepthwiseConv3D
  • SeparableConv3D
• Documentation
  • Documentation
• Exmaple
  • Classification
  • Regression
  • Visualize trained model
• Visualization
  • Saliency map (Simple gradient)
  • Class Activation Map (GradCAM)
  • Activation Maximization
• Pretrained weight
  • ModelNet10
  • ModelNet40
  • (Please recommand any 3D dataset)

• ModelNet10 classificaition using VGG19(16)

• So We prepare some custom model to handle this.
• Please check base_channel or growth_rate option.

• So We also prepare some option to exclude batch-norm.

We are migrating for TF (>= 2.5)
If you interested in this project, feel free and suggest anything.