Keras Issues

---

Up to now, we have generally supported ALL the layers of Keras 1.2.2 to be loaded into BigDL.

Self-defined Keras layers or Lambda layers are not supported for now. Weight sharing is not supported for now.

This page lists the so-far unsupported arguments for specific layers and some known issues.

Unsupported Layer Arguments

For the following arguments, currently only the default values are expected and supported.

• For initializations, lecun_uniform, orthogonal, glorot_normal, he_normal and he_uniform are not supported for all layers. For those unsupported initialization methods, we will default values from BigDL instead. Thus in this case, you are strongly recommended to load the weights from Keras together with its model definition.
• Constraints (W_constraint, b_constraint, etc.) are not supported for all layers.
• activity_regularizer is not supported for all layers.
• For Dropout, noise_shape is not supported.
• For Merge layers, node_indices, tensor_indices, output_mask are not supported. Lambda/function as merge mode is not supported.
• For Merge layers with mode 'dot' or 'cosine', only 2D input with dot_axes=1 is supported.
• For AtrousConvolution1D, only bias=True is supported.
• For AtrousConvolution2D, only dim_ordering='th' and bias=True are supported.
• For DeConvolution2D, only dim_ordering='th' is supported. For border_mode='same', only symmetric padding on both sides is supported.
• For Convolution3D, only dim_ordering='th' is supported.
• For UpSampling3D, only dim_ordering='th' is supported.
• For MaxPooling3D, only dim_ordering='th' and border_mode='valid' are supported.
• For AveragePooling3D, only dim_ordering='th' and border_mode='valid' are supported.
• For GlobalMaxPooling3D, only dim_ordering='th' is supported.
• For GlobalAveragePooling3D, only dim_ordering='th' is supported.
• For all Recurrent layers (SimpleRNN, GRU, LSTM and ConvLSTM2D), stateful, consume_less, dropout_W and dropout_U are not supported. If RNNs are wrapped with Bidirectional, only return_sequences=True is supported.
• For ConvLSTM2D, only border_mode='valid', nb_row==nb_col(square kernel) and subsample[0]==subsample[1](equal strides) are supported.
• For Embedding, mask_zero and dropout are not supported.
• For PReLU, init, weights and shared_axes are not supported.
• For ParametricSoftplus, weights and shared_axes are not supported. Only alpha_init = 1/beta_init is supported.
• For BatchNormalization, only 4D input with mode=0 is supported. Only channel_first (dim_ordering='th' with axis=1) and channel_last (dim_ordering='tf' with axis=-1) is supported. gamma_regularizer and beta_regularizer are not supported.

Known Issues

• For some layers such as ZeroPadding2D, ZeroPadding3D, Cropping2D, Cropping3D, etc., Keras doesn't save dim_ordering into config. In this case, the default dim_ordering found in ~/.keras/keras.json will be used instead.
• For ConvLSTM2D, Keras doesn't save W_regularizer, U_regularizer and b_regularizer into config. In this case, the default None will be used instead.