Random data for training.

Return keys of a dictionary

Serialize all weights of a single normal layer

Include size of data to generate into shape.

Change dimension order of 2d filter weights to the one used in fdeep

Load and construct an fdeep::model from an istream providing the exported json content. Throws an exception if a problem occurs.

Replace every None with a fixed value.

Measure time of one single forward pass using dummy input data.

Test data can only be generated if all Embedding layers and CategoryEncoding layers are positioned directly behind the input nodes.

Leave lists untouched, convert non-list types to a singleton list

Convert any (h5-)stored Keras model to the frugally-deep model format.

Convert a sequential model to the underlying functional format

Special version for convolution with strides_x == 1 and strides_y == 1. Reduces the forward-pass runtime of VGG19 about 15%, by using fewer but larger

Dilated variant: filter values are at positions (y_filt * dil_y, x_filt * dil_x) in the effective receptive field. We loop over the un-dilated filter

Special version for convolution with strides 1×1×1. Uses fewer but larger GEMMs by collapsing the spatial output loops into one big input mapping.

Generate data for model verification test.

Random input data for training.

Random data for training.

Serialize all weights of the models layers

Return .name or ._name

Check if a string only contains ascii characters

Convert a keras shape to an fdeep shape

Parse command line and convert model.

Generate different test models and save them to the given directory.

Save famous example models in Keras-h5 and fdeep-json format.

Convert any Keras model to the frugally-deep model format.

Change dimension order of 1d filter weights to the one used in fdeep