You are here: Start » Supported Keras layers and features
Supported Keras layers and features
The table below lists the supported layers with possible restrictions. For parameters not explicitly mentioned, WEAVER can accept any value valid from the Keras’ point of view.
Keras layer | Restrictions |
---|---|
Activation | activation=relu/relu6/sigmoid/softmax/tanh |
Add | Only two inputs |
BatchNormalization | axis=-1 |
Conv2D | activation=None/tanh/relu/sigmoid/softmax/elu |
Conv2DTranspose | activation=None/tanh/relu/sigmoid/softmax/elu output_padding=None |
Concatenate | axis=-1 |
Cropping2D | |
Dense | activation=None/tanh/relu/sigmoid/softmax/elu |
DepthwiseConv2D | activation=None/tanh/relu/sigmoid/softmax/elu depth_multiplier=1 |
Dropout | |
ELU | |
GlobalAveragePooling2D | |
InputLayer | |
LeakyReLU | |
MaxPooling2D | padding=valid |
Multiply | Only two inputs |
ReLU | negative_slope=0 threshold=0 |
Reshape | target_shape have to count 1, 2 or 3 values |
PReLU | shared_axes=None/[1, 2] |
TimeDistributed | |
UpSampling2D | interpolation=nearest Both values in size must be the same in case of running a model on CUDA. |
ZeroPadding2D |
Additionally:
- For layers that have "data_format" parameter, only "channels_last" is supported.
- Nested models are also supported as long as they contain only supported layers.
- WEAVER works with float32 data type only.
- WEAVER works with Keras Functional models only. Python script train_mnist_model.py located in "Trained MNIST Model" example shows definition and training of an example model.
Previous: SDK Usage |