moritzwag / lpdn-1 Goto Github PK

This repository contains an unofficial implementation of Lightweight Probabilistic Deep Networks using Keras (and assuming Tensorflow backend).

Notice that this library is at en early stage of development and proper unity tests are not there yet. Moreover, only ReLU and Linear are supported as layer activations.

In your Python 3 (virtual) environment:

pip install git+https://github.com/marcoancona/LPDN.git

Notice that this will not install any additional dependencies but LPDN assumes numpy, tensorflow(-gpu) and keras are available.

A Keras probabilistic model can be built from scratch or converted from an existing model.

Let assume we want to build the equivalent of the following model:

from keras.layers import Dense, Flatten, Activation, Conv2D, MaxPooling2D

model = Sequential()
model.add(Conv2D(32, kernel_size=(3, 3),
                 activation='relu',
                 input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
model.add(Flatten())
model.add(Dense(num_classes))

We can either use the conversion utility

from lpdn import convert_to_lpdn

lp_model = convert_to_lpdn(model)

or build the model from scratch by replacing the original layers with the Lightweight Propabilistic (LP-) equivalent.

from lpdn import LPDense, LPFlatten, LPActivation, LPConv2D, LPMaxPooling2D

lp_model = Sequential()
lp_model.add(LPConv2D(32, kernel_size=(3, 3),
                 activation='relu',
                 input_shape=input_shape))
lp_model.add(LPConv2D(64, (3, 3), activation='relu'))
lp_model.add(LPMaxPooling2D(pool_size=(2, 2)))
lp_model.add(LPFlatten())
lp_model.add(LPDense(num_classes))

Notice that, in both cases, the probabilistic model is initialized with random weights. You can easily transfer the weights of the original model:

model.save_weights('w.h5')
lp_model.load_weights('w.h5')

If model takes an input of shape [batch, n_features], lp_model requires an input of shape [batch, n_features, 2] where mean and variance of the input features are stacked along the last dimension. Similarly, the output will also have one additional dimension to account for mean and variance.