def _model_3(_dataset):
if not is_module_installed('keras_rewiring'):
return
from keras_rewiring import Sparse, SparseConv2D, SparseDepthwiseConv2D
x_train, y_train, x_test, y_test = _dataset
axis = 1 if keras.backend.image_data_format() == 'channels_first' else -1
input_shape = x_train.shape[1:]
input_layer = Input(input_shape)
layer = SparseConv2D(filters=16, kernel_size=(5, 5),
strides=(2, 2))(input_layer)
layer = BatchNormalization(axis=axis)(layer)
layer = Activation('relu')(layer)
layer = AveragePooling2D()(layer)
branch1 = SparseConv2D(filters=32,
kernel_size=(3, 3),
padding='same',
activation='relu')(layer)
branch2 = SparseDepthwiseConv2D(kernel_size=(1, 1),
activation='relu')(layer)
layer = Concatenate(axis=axis)([branch1, branch2])
layer = SparseConv2D(filters=10, kernel_size=(3, 3),
activation='relu')(layer)
layer = Flatten()(layer)
layer = Dropout(1e-5)(layer)
layer = Sparse(units=10, activation='softmax')(layer)
model = Model(input_layer, layer)
model.compile('adam', 'categorical_crossentropy', ['accuracy'])
# Train model with backprop.
history = model.fit(x_train,
y_train,
batch_size=64,
epochs=1,
verbose=2,
validation_data=(x_test, y_test))
assert history.history['val_accuracy'][-1] > 0.96
return model
##############
# This section creates a simple CNN using Keras, and trains it
# with backpropagation. There are no spikes involved at this point.
input_shape = x_train.shape[1:]
input_layer = Input(input_shape)
built_in_sparsity = [0.5] * 4
layer = SparseConv2D(
filters=16,
kernel_size=(
5,
5),
strides=(
2,
2),
activation='relu',
use_bias=False,
connectivity_level=built_in_sparsity.pop(0) or None)(input_layer)
layer = SparseConv2D(
filters=32,
kernel_size=(
3,
3),
activation='relu',
use_bias=False,
connectivity_level=built_in_sparsity.pop(0) or None)(layer)
layer = AveragePooling2D()(layer)
layer = SparseConv2D(