keras-applications-3D

'keras-applications-3D' is 3D-image deep learning models based on popular 2D models. (Based on the Keras)

---

Install

pip install Keras-Applications-3D

---

Usage

### 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))

---

This repository provide below things

• 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)

---

Model benchmark

Model	ModelNet10 accuracy	Number of parameters
VGG16 (16)	0.9001	23,780,058
VGG19 (16)	0.9075	24,775,706
ResNet50 (16)	0.8062	2,923,818
ResNet101 (16)	0.7467	5,393,578
ResNet152 (16)	0.7588	7,429,418
ResNet50V2 (16)	0.8194	2,918,090
ResNet101V2 (16)	0.8128	5,385,674
ResNet152V2 (16)	0.8007	7,419,594
DenseNet121 (16)	0.9042	2,884,010
DenseNet169 (16)	0.8855	4,768,298
DenseNet201 (16)	0.8998	6,519,594

---

Example

• ModelNet10 classificaition using VGG19(16)

---

FYI

If you want to use 3D CNN, you'd better reduce number of parameter because of curse of dimension.

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

When you use 3D CNN, BatchNormalization may not work well in the scarce data.

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

---

NOTICE

If you interested in this project, feel free and suggest anything.