def genSfc(input, num_outputs, learning_parameters):
# A function to generate the lfc network topology which matches the overlay for the Pynq board.
# WARNING: If you change this file, it's likely the resultant weights will not fit on the Pynq overlay.
if num_outputs < 1 or num_outputs > 64:
error("num_outputs should be in the range of 1 to 64.")
num_units = 256
n_hidden_layers = 3
if learning_parameters.activation_bits == 1:
act_quant = q.QuantizationBinary()
else:
act_quant = q.QuantizationFixed(
learning_parameters.activation_bits,
learning_parameters.activation_bits - 2)
activation = qn.FixedHardTanH(act_quant)
if learning_parameters.weight_bits == 1:
weight_quant = q.QuantizationBinary()
else:
weight_quant = q.QuantizationFixed(learning_parameters.weight_bits,
learning_parameters.weight_bits - 2)
W_LR_scale = learning_parameters.W_LR_scale
epsilon = learning_parameters.epsilon
alpha = learning_parameters.alpha
dropout_in = learning_parameters.dropout_in
dropout_hidden = learning_parameters.dropout_hidden
mlp = lasagne.layers.InputLayer(shape=(None, 1, 28, 28), input_var=input)
mlp = lasagne.layers.DropoutLayer(mlp, p=dropout_in)
for k in range(n_hidden_layers):
mlp = qn.DenseLayer(mlp,
quantization=weight_quant,
W_LR_scale=W_LR_scale,
nonlinearity=lasagne.nonlinearities.identity,
num_units=num_units)
mlp = lasagne.layers.BatchNormLayer(mlp, epsilon=epsilon, alpha=alpha)
mlp = lasagne.layers.NonlinearityLayer(mlp, nonlinearity=activation)
mlp = lasagne.layers.DropoutLayer(mlp, p=dropout_hidden)
mlp = qn.DenseLayer(mlp,
quantization=weight_quant,
W_LR_scale=W_LR_scale,
nonlinearity=lasagne.nonlinearities.identity,
num_units=num_outputs)
mlp = lasagne.layers.BatchNormLayer(mlp, epsilon=epsilon, alpha=alpha)
return mlp
def genCnv(input, num_outputs, learning_parameters):
# A function to generate the cnv network topology which matches the overlay for the Pynq board.
# WARNING: If you change this file, it's likely the resultant weights will not fit on the Pynq overlay.
if num_outputs < 1 or num_outputs > 64:
error("num_outputs should be in the range of 1 to 64.")
if learning_parameters.activation_bits == 1:
act_quant = q.QuantizationBinary()
else:
act_quant = q.QuantizationFixed(learning_parameters.activation_bits,
learning_parameters.activation_bits - 2)
activation = qn.FixedHardTanH(act_quant)
if learning_parameters.weight_bits == 1:
weight_quant = q.QuantizationBinary()
else:
weight_quant = q.QuantizationFixed(learning_parameters.weight_bits,
learning_parameters.weight_bits - 2)
input_quant = q.QuantizationFixed(8, 7, narrow_range=False)
W_LR_scale = learning_parameters.W_LR_scale
epsilon = learning_parameters.epsilon
alpha = learning_parameters.alpha
cnn = lasagne.layers.InputLayer(
shape=(None, 3, 32, 32),
input_var=input)
# Quantize the inputs.
# Precondition: inputs have already been scaled to [-1, 1]
cnn = lasagne.layers.NonlinearityLayer(
cnn,
nonlinearity=qn.FixedHardTanH(input_quant))
# 64C3-64C3-P2
cnn = qn.Conv2DLayer(
cnn,
quantization=weight_quant,
W_LR_scale=W_LR_scale,
num_filters=64,
filter_size=(3, 3),
pad='valid',
flip_filters=False,
nonlinearity=lasagne.nonlinearities.identity)
cnn = lasagne.layers.BatchNormLayer(
cnn,
epsilon=epsilon,
alpha=alpha)
cnn = lasagne.layers.NonlinearityLayer(
cnn,
nonlinearity=activation)
cnn = qn.Conv2DLayer(
cnn,
quantization=weight_quant,
W_LR_scale=W_LR_scale,
num_filters=64,
filter_size=(3, 3),
pad='valid',
flip_filters=False,
nonlinearity=lasagne.nonlinearities.identity)
cnn = lasagne.layers.MaxPool2DLayer(cnn, pool_size=(2, 2), ignore_border=False)
cnn = lasagne.layers.BatchNormLayer(
cnn,
epsilon=epsilon,
alpha=alpha)
cnn = lasagne.layers.NonlinearityLayer(
cnn,
nonlinearity=activation)
# 256C3-256C3-P2
cnn = qn.Conv2DLayer(
cnn,
quantization=weight_quant,
W_LR_scale=W_LR_scale,
num_filters=128,
filter_size=(3, 3),
pad='valid',
flip_filters=False,
nonlinearity=lasagne.nonlinearities.identity)
cnn = lasagne.layers.BatchNormLayer(
cnn,
epsilon=epsilon,
alpha=alpha)
cnn = lasagne.layers.NonlinearityLayer(
cnn,
nonlinearity=activation)
cnn = qn.Conv2DLayer(
cnn,
quantization=weight_quant,
W_LR_scale=W_LR_scale,
num_filters=128,
filter_size=(3, 3),
pad='valid',
flip_filters=False,
nonlinearity=lasagne.nonlinearities.identity)
cnn = lasagne.layers.MaxPool2DLayer(cnn, pool_size=(2, 2), ignore_border=False)
cnn = lasagne.layers.BatchNormLayer(
cnn,
epsilon=epsilon,
alpha=alpha)
cnn = lasagne.layers.NonlinearityLayer(
cnn,
nonlinearity=activation)
# 256C3-256C3
cnn = qn.Conv2DLayer(
cnn,
quantization=weight_quant,
W_LR_scale=W_LR_scale,
num_filters=256,
filter_size=(3, 3),
pad='valid',
flip_filters=False,
nonlinearity=lasagne.nonlinearities.identity)
cnn = lasagne.layers.BatchNormLayer(
cnn,
epsilon=epsilon,
alpha=alpha)
cnn = lasagne.layers.NonlinearityLayer(
cnn,
nonlinearity=activation)
cnn = qn.Conv2DLayer(
cnn,
quantization=weight_quant,
W_LR_scale=W_LR_scale,
num_filters=256,
filter_size=(3, 3),
pad='valid',
flip_filters=False,
nonlinearity=lasagne.nonlinearities.identity)
cnn = lasagne.layers.BatchNormLayer(
cnn,
epsilon=epsilon,
alpha=alpha)
cnn = lasagne.layers.NonlinearityLayer(
cnn,
nonlinearity=activation)
# print(cnn.output_shape)
# 1024FP-1024FP-10FP
cnn = qn.DenseLayer(
cnn,
quantization=weight_quant,
W_LR_scale=W_LR_scale,
nonlinearity=lasagne.nonlinearities.identity,
num_units=512)
cnn = lasagne.layers.BatchNormLayer(
cnn,
epsilon=epsilon,
alpha=alpha)
cnn = lasagne.layers.NonlinearityLayer(
cnn,
nonlinearity=activation)
cnn = qn.DenseLayer(
cnn,
quantization=weight_quant,
W_LR_scale=W_LR_scale,
nonlinearity=lasagne.nonlinearities.identity,
num_units=512)
cnn = lasagne.layers.BatchNormLayer(
cnn,
epsilon=epsilon,
alpha=alpha)
cnn = lasagne.layers.NonlinearityLayer(
cnn,
nonlinearity=activation)
cnn = qn.DenseLayer(
cnn,
quantization=weight_quant,
W_LR_scale=W_LR_scale,
nonlinearity=lasagne.nonlinearities.identity,
num_units=num_outputs,
b=None)
cnn = lasagne.layers.BatchNormLayer(
cnn,
epsilon=epsilon,
alpha=alpha,
axes=(0,1))
return cnn