def build_nature_network(input_width, input_height, output_dim, num_frames,
batch_size, human_net, input_var):
"""
Build a large network consistent with the DeepMind Nature paper.
"""
from lasagne.layers import cuda_convnet
l_in = None
if human_net:
l_in = lasagne.layers.InputLayer(shape=(None, num_frames, input_width,
input_height),
input_var=input_var)
else:
l_in = lasagne.layers.InputLayer(shape=(None, num_frames, input_width,
input_height))
l_conv1 = cuda_convnet.Conv2DCCLayer(
l_in,
num_filters=32,
filter_size=(8, 8),
stride=(4, 4),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(), # Defaults to Glorot
b=lasagne.init.Constant(.1),
dimshuffle=True)
l_conv2 = cuda_convnet.Conv2DCCLayer(
l_conv1,
num_filters=64,
filter_size=(4, 4),
stride=(2, 2),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1),
dimshuffle=True)
l_conv3 = cuda_convnet.Conv2DCCLayer(
l_conv2,
num_filters=64,
filter_size=(3, 3),
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1),
dimshuffle=True)
l_hidden1 = lasagne.layers.DenseLayer(
l_conv3,
num_units=512,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
l_out = lasagne.layers.DenseLayer(l_hidden1,
num_units=output_dim,
nonlinearity=None,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
return l_out
def build_nips_network(input_width, input_height, output_dim, num_frames,
batch_size, human_net, input_var):
"""
Build a network consistent with the 2013 NIPS paper.
"""
from lasagne.layers import cuda_convnet
l_in = None
if human_net:
l_in = lasagne.layers.InputLayer(shape=(None, num_frames, input_width,
input_height),
input_var=input_var)
else:
l_in = lasagne.layers.InputLayer(shape=(None, num_frames, input_width,
input_height))
l_conv1 = cuda_convnet.Conv2DCCLayer(
l_in,
num_filters=16,
filter_size=(8, 8),
stride=(4, 4),
nonlinearity=lasagne.nonlinearities.rectify,
#W=lasagne.init.HeUniform(c01b=True),
W=lasagne.init.Normal(.01),
b=lasagne.init.Constant(.1),
dimshuffle=True)
l_conv2 = cuda_convnet.Conv2DCCLayer(
l_conv1,
num_filters=32,
filter_size=(4, 4),
stride=(2, 2),
nonlinearity=lasagne.nonlinearities.rectify,
#W=lasagne.init.HeUniform(c01b=True),
W=lasagne.init.Normal(.01),
b=lasagne.init.Constant(.1),
dimshuffle=True)
l_hidden1 = lasagne.layers.DenseLayer(
l_conv2,
num_units=256,
nonlinearity=lasagne.nonlinearities.rectify,
#W=lasagne.init.HeUniform(),
W=lasagne.init.Normal(.01),
b=lasagne.init.Constant(.1))
l_out = lasagne.layers.DenseLayer(
l_hidden1,
num_units=output_dim,
nonlinearity=None,
#W=lasagne.init.HeUniform(),
W=lasagne.init.Normal(.01),
b=lasagne.init.Constant(.1))
return l_out
def build_network(self, input_width, input_height, output_dim, num_frames,
batch_size):
l_in = lasagne.layers.InputLayer(shape=(batch_size, num_frames,
input_width, input_height))
l_in = cuda_convnet.bc01_to_c01b(l_in)
l_conv1 = cuda_convnet.Conv2DCCLayer(
l_in,
num_filters=32,
filter_size=(8, 8),
strides=(4, 4),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.Uniform(0.01),
b=lasagne.init.Constant(0.1),
dimshuffle=False)
l_conv2 = cuda_convnet.Conv2DCCLayer(
l_conv1,
num_filters=64,
filter_size=(4, 4),
strides=(2, 2),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.Uniform(0.01),
b=lasagne.init.Constant(0.1),
dimshuffle=False)
l_conv3 = cuda_convnet.Conv2DCCLayer(
l_conv2,
num_filters=64,
filter_size=(3, 3),
strides=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.Uniform(0.01),
b=lasagne.init.Constant(0.1))
l_conv3 = cuda_convnet.c01b_to_bc01(l_conv3)
l_hidden1 = lasagne.layers.DenseLayer(
l_conv3,
num_units=512,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.Uniform(0.01),
b=lasagne.init.Constant(0.1))
l_out = lasagne.layers.DenseLayer(l_hidden1,
num_units=output_dim,
nonlinearity=None,
W=lasagne.init.Uniform(0.01),
b=lasagne.init.Constant(0.1))
return l_out
def build_nips_network(self, input_width, input_height, output_dim,
num_frames, batch_size):
"""
Build a network consistent with the 2013 NIPS paper.
"""
# value of None in shape means that it is not fixed at compile time
from lasagne.layers import cuda_convnet
l_in = lasagne.layers.InputLayer(
# shape=(batch_size, num_frames, input_width, input_height)
shape=(None, num_frames, input_width, input_height))
l_conv1 = cuda_convnet.Conv2DCCLayer(
l_in,
num_filters=16,
filter_size=(8, 8),
stride=(4, 4),
nonlinearity=lasagne.nonlinearities.rectify,
#W=lasagne.init.HeUniform(c01b=True),
W=lasagne.init.Normal(.01),
b=lasagne.init.Constant(.1),
dimshuffle=True)
l_conv2 = cuda_convnet.Conv2DCCLayer(
l_conv1,
num_filters=32,
filter_size=(4, 4),
stride=(2, 2),
nonlinearity=lasagne.nonlinearities.rectify,
#W=lasagne.init.HeUniform(c01b=True),
W=lasagne.init.Normal(.01),
b=lasagne.init.Constant(.1),
dimshuffle=True)
l_hidden1 = lasagne.layers.DenseLayer(
l_conv2,
num_units=256,
nonlinearity=lasagne.nonlinearities.rectify,
#W=lasagne.init.HeUniform(),
W=lasagne.init.Normal(.01),
b=lasagne.init.Constant(.1))
l_out = lasagne.layers.DenseLayer(
l_hidden1,
num_units=output_dim,
nonlinearity=None,
#W=lasagne.init.HeUniform(),
W=lasagne.init.Normal(.01),
b=lasagne.init.Constant(.1))
return l_out
def build_nips_network(self, input_width, input_height, output_dim, num_frames, batch_size): """ Build a network based on google atari deep learning paper """ from lasagne.layers import cuda_convnet l_in = lasagne.layers.InputLayer( shape=(batch_size, num_frames, input_width, input_height) ) l_conv1 = cuda_convnet.Conv2DCCLayer( l_in, #previous layer num_filters=16, #16 8x8 filters filter_size=(8, 8), stride=(4, 4), nonlinearity=lasagne.nonlinearities.rectify, #convolution layer # W is the weights and b is bias term W=lasagne.init.Normal(.01), b=lasagne.init.Constant(.1), dimshuffle=True ) l_conv2 = cuda_convnet.Conv2DCCLayer( l_conv1, #previous layer num_filters=32, #32 4x4 filters filter_size=(4, 4), stride=(2, 2), nonlinearity=lasagne.nonlinearities.rectify, #convolution layer W=lasagne.init.Normal(.01), b=lasagne.init.Constant(.1), dimshuffle=True ) l_hidden1 = lasagne.layers.DenseLayer( l_conv2, #previous layer num_units=256, #256 hidden units nonlinearity=lasagne.nonlinearities.rectify, W=lasagne.init.Normal(.01), b=lasagne.init.Constant(.1) ) l_out = lasagne.layers.DenseLayer( l_hidden1, num_units=output_dim, nonlinearity=None, W=lasagne.init.Normal(.01), b=lasagne.init.Constant(.1) ) return l_out
def build_network():
l_in = lasagne.layers.InputLayer(shape=(None, 8, GRIDSIZE * 2,
ARRAYWIDTH * 2))
l_conv1 = cuda_convnet.Conv2DCCLayer(
l_in,
num_filters=32,
filter_size=(6, 6),
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1),
)
l_conv2 = cuda_convnet.Conv2DCCLayer(
l_conv1,
num_filters=64,
filter_size=(2, 2),
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1),
)
l_hidden1 = lasagne.layers.DenseLayer(
l_conv2,
num_units=512,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
l_hidden2 = lasagne.layers.DenseLayer(
l_hidden1,
num_units=512,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
l_out = lasagne.layers.DenseLayer(l_hidden2,
num_units=NUMBEROFACTIONS,
nonlinearity=None,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
return l_out
def build_model(input_width, input_height, output_dim,
batch_size=BATCH_SIZE, dimshuffle=True):
l_in = lasagne.layers.InputLayer(
shape=(batch_size, 1, input_width, input_height),
)
if not dimshuffle:
l_in = cuda_convnet.bc01_to_c01b(l_in)
l_conv1 = cuda_convnet.Conv2DCCLayer(
l_in,
num_filters=32,
filter_size=(5, 5),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.Uniform(),
dimshuffle=dimshuffle,
)
l_pool1 = cuda_convnet.MaxPool2DCCLayer(
l_conv1,
ds=(2, 2),
dimshuffle=dimshuffle,
)
l_conv2 = cuda_convnet.Conv2DCCLayer(
l_pool1,
num_filters=32,
filter_size=(5, 5),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.Uniform(),
dimshuffle=dimshuffle,
)
l_pool2 = cuda_convnet.MaxPool2DCCLayer(
l_conv2,
ds=(2, 2),
dimshuffle=dimshuffle,
)
if not dimshuffle:
l_pool2 = cuda_convnet.c01b_to_bc01(l_pool2)
l_hidden1 = lasagne.layers.DenseLayer(
l_pool2,
num_units=256,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.Uniform(),
)
l_hidden1_dropout = lasagne.layers.DropoutLayer(l_hidden1, p=0.5)
# l_hidden2 = lasagne.layers.DenseLayer(
# l_hidden1_dropout,
# num_units=256,
# nonlinearity=lasagne.nonlinearities.rectify,
# )
# l_hidden2_dropout = lasagne.layers.DropoutLayer(l_hidden2, p=0.5)
l_out = lasagne.layers.DenseLayer(
l_hidden1_dropout,
num_units=output_dim,
nonlinearity=lasagne.nonlinearities.softmax,
W=lasagne.init.Uniform(),
)
return l_out
def build_model(input_width=96,
input_height=96,
output_dim=1,
batch_size=64,
dimshuffle=True):
l_in = lasagne.layers.InputLayer(shape=(batch_size, 3, input_width,
input_height), )
if not dimshuffle:
l_in = cuda_convnet.bc01_to_c01b(l_in)
l_conv1 = cuda_convnet.Conv2DCCLayer(
l_in,
num_filters=32,
filter_size=(3, 3),
nonlinearity=lasagne.nonlinearities.rectify,
dimshuffle=dimshuffle,
)
l_pool1 = cuda_convnet.MaxPool2DCCLayer(
l_conv1,
pool_size=(2, 2),
dimshuffle=dimshuffle,
)
l_conv2 = cuda_convnet.Conv2DCCLayer(
l_pool1,
num_filters=64,
filter_size=(2, 2),
nonlinearity=lasagne.nonlinearities.rectify,
dimshuffle=dimshuffle,
)
l_pool2 = cuda_convnet.MaxPool2DCCLayer(
l_conv2,
pool_size=(2, 2),
dimshuffle=dimshuffle,
)
l_conv3 = cuda_convnet.Conv2DCCLayer(
l_pool2,
num_filters=128,
filter_size=(2, 2),
nonlinearity=lasagne.nonlinearities.rectify,
dimshuffle=dimshuffle,
)
l_pool3 = cuda_convnet.MaxPool2DCCLayer(
l_conv3,
pool_size=(2, 2),
dimshuffle=dimshuffle,
)
if not dimshuffle:
l_poo3 = cuda_convnet.c01b_to_bc01(l_pool3)
l_hidden1 = lasagne.layers.DenseLayer(
l_pool3,
num_units=512,
nonlinearity=lasagne.nonlinearities.rectify,
)
l_hidden1_dropout = lasagne.layers.DropoutLayer(l_hidden1, p=0.5)
l_hidden2 = lasagne.layers.DenseLayer(
l_hidden1_dropout,
num_units=512,
nonlinearity=lasagne.nonlinearities.rectify,
)
l_hidden2_dropout = lasagne.layers.DropoutLayer(l_hidden2, p=0.5)
l_out = lasagne.layers.DenseLayer(
l_hidden2_dropout,
num_units=output_dim,
nonlinearity=lasagne.nonlinearities.linear,
)
return l_out
def build_large_network(self, input_width, input_height, output_dim,
num_frames, batch_size):
"""
Build a large network consistent with the DeepMind Nature paper.
"""
from lasagne.layers import cuda_convnet
l_in = lasagne.layers.InputLayer(shape=(batch_size, num_frames,
input_width, input_height))
l_conv1 = cuda_convnet.Conv2DCCLayer(
l_in,
num_filters=32,
filter_size=(8, 8),
stride=(4, 4),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(), # Defaults to Glorot
b=lasagne.init.Constant(.1),
dimshuffle=True)
l_conv2 = cuda_convnet.Conv2DCCLayer(
l_conv1,
num_filters=64,
filter_size=(4, 4),
stride=(2, 2),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1),
dimshuffle=True)
l_conv3 = cuda_convnet.Conv2DCCLayer(
l_conv2,
num_filters=64,
filter_size=(3, 3),
stride=(1, 1),
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1),
dimshuffle=True)
l_hidden1 = lasagne.layers.DenseLayer(
l_conv3,
num_units=512,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
l_out = lasagne.layers.DenseLayer(l_hidden1,
num_units=output_dim,
nonlinearity=None,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
# compressor
l_dropout = lasagne.layers.DropoutLayer(
l_conv3,
p=0.3,
)
l_compress_1 = lasagne.layers.DenseLayer(
l_dropout,
num_units=512,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
l_compress_2 = lasagne.layers.DenseLayer(
l_compress_1,
num_units=512,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
l_representation = lasagne.layers.DenseLayer(
l_compress_2,
num_units=256,
nonlinearity=None,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
l_uncompress_1 = lasagne.layers.DenseLayer(
l_representation,
num_units=512,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
l_uncompress_2 = lasagne.layers.DenseLayer(
l_uncompress_1,
num_units=512,
nonlinearity=lasagne.nonlinearities.rectify,
W=lasagne.init.HeUniform(),
b=lasagne.init.Constant(.1))
return l_out, l_conv3, l_uncompress_2
