def get_maxout(dim_input):
config = {
'batch_size':
bsize,
'input_space':
Conv2DSpace(shape=dim_input[:2],
num_channels=dim_input[2],
axes=['c', 0, 1, 'b']),
'layers': [
MaxoutConvC01B(layer_name='h0',
num_channels=96,
num_pieces=2,
irange=.005,
tied_b=1,
max_kernel_norm=.9,
kernel_shape=[8, 8],
pool_shape=[4, 4],
pool_stride=[2, 2],
W_lr_scale=.05,
b_lr_scale=.05),
MaxoutConvC01B(layer_name='h1',
num_channels=128,
num_pieces=2,
irange=.005,
tied_b=1,
max_kernel_norm=0.9,
kernel_shape=[7, 7],
pad=3,
pool_shape=[4, 4],
pool_stride=[2, 2],
W_lr_scale=.05,
b_lr_scale=.05),
MaxoutConvC01B(layer_name='h2',
num_channels=160,
num_pieces=3,
irange=.005,
tied_b=1,
max_kernel_norm=0.9,
kernel_shape=[6, 6],
pad=2,
pool_shape=[2, 2],
pool_stride=[2, 2],
W_lr_scale=.05,
b_lr_scale=.05),
MaxoutConvC01B(layer_name='h3',
num_channels=192,
num_pieces=4,
irange=.005,
tied_b=1,
max_kernel_norm=0.9,
kernel_shape=[5, 5],
pad=1,
pool_shape=[2, 2],
pool_stride=[2, 2],
W_lr_scale=.05,
b_lr_scale=.05),
Maxout(layer_name='h4',
irange=.005,
num_units=500,
num_pieces=5,
max_col_norm=1.9),
Softmax(layer_name='y',
n_classes=nclass,
irange=.005,
max_col_norm=1.9)
]
}
return MLP(**config)
def generateConvRegressor(teacher_hintlayer, student_layer):
layer_name = 'hint_regressor'
out_ch = teacher_hintlayer.get_output_space().num_channels
ks0 = student_layer.get_output_space().shape[0] - teacher_hintlayer.get_output_space().shape[0] + 1
ks1 = student_layer.get_output_space().shape[1] - teacher_hintlayer.get_output_space().shape[1] + 1
ks = [ks0, ks1]
irng = 0.05
mkn = 0.9
tb = 1
if isinstance(teacher_hintlayer, MaxoutConvC01B):
hint_reg_layer = MaxoutConvC01B(num_channels=out_ch,
num_pieces=teacher_hintlayer.num_pieces,
kernel_shape=ks,
pool_shape=[1,1],
pool_stride=[1,1],
layer_name=layer_name,
irange=irng,
max_kernel_norm=mkn,
tied_b=teacher_hintlayer.tied_b)
elif isinstance(teacher_hintlayer, ConvRectifiedLinear):
nonlin = RectifierConvNonlinearity()
hint_reg_layer = ConvElemwise(output_channels = out_ch,
kernel_shape = ks,
layer_name = layer_name,
nonlinearity = nonlin,
irange = irng,
max_kernel_norm = mkn,
tied_b = tb)
elif isinstance(teacher_hintlayer, ConvElemwise):
nonlin = teacher_hintlayer.nonlinearity
#if isinstance(nonlin,TanhConvNonlinearity):
# nonlin = SigmoidConvNonlinearity()
hint_reg_layer = ConvElemwise(output_channels = out_ch,
kernel_shape = ks,
layer_name = layer_name,
nonlinearity = nonlin,
irange = irng,
max_kernel_norm = mkn,
tied_b = tb)
elif isinstance(teacher_hintlayer, ConvElemwisePL2):
nonlin = teacher_hintlayer.nonlinearity
#if isinstance(nonlin,TanhConvNonlinearity):
# nonlin = SigmoidConvNonlinearity()
hint_reg_layer = ConvElemwisePL2(output_channels = out_ch,
kernel_shape = ks,
layer_name = layer_name,
nonlinearity = nonlin,
irange = irng,
max_kernel_norm = mkn,
tied_b = tb)
elif isinstance(teacher_hintlayer, CudNNElemwise):
nonlin = teacher_hintlayer.nonlinearity
#if isinstance(nonlin,TanhConvNonlinearity):
# nonlin = SigmoidConvNonlinearity()
hint_reg_layer = CudNNElemwise(output_channels = out_ch,
kernel_shape = ks,
layer_name = layer_name,
nonlinearity = nonlin,
irange = irng,
max_kernel_norm = mkn,
tied_b = tb)
else:
raise AssertionError("Unknown layer type")
return hint_reg_layer
MonitorBasedSaveBest(channel_name='valid_y_misclass',
save_path=save_best_path),
MomentumAdjustor(start=1, saturate=250, final_momentum=.7)
]
model = MLP(batch_size=batch_size,
input_space=Conv2DSpace(shape=[48, 48],
num_channels=num_chan,
axes=['c', 0, 1, 'b']),
layers=[
MaxoutConvC01B(layer_name='h0',
pad=0,
num_channels=64,
num_pieces=2,
kernel_shape=[8, 8],
pool_shape=[4, 4],
pool_stride=[2, 2],
irange=.005,
max_kernel_norm=.9,
W_lr_scale=0.5,
b_lr_scale=0.5),
MaxoutConvC01B(layer_name='h1',
pad=0,
num_channels=64,
num_pieces=2,
kernel_shape=[8, 8],
pool_shape=[3, 3],
pool_stride=[2, 2],
irange=.005,
max_kernel_norm=.9,
W_lr_scale=0.5,
translation = 9.
center_shape = (img_dim - 2, img_dim - 2)
preprocess = [
preprocessing.GlobalContrastNormalization(sqrt_bias=10., use_std=True),
preprocessing.LeCunLCN([img_dim, img_dim], batch_size=5000)
]
#number of random test augmentations to predict
test_examples = 2
#convolutional layers
l1 = MaxoutConvC01B(layer_name='l1',
tied_b=1,
num_channels=32,
num_pieces=2,
pad=0,
kernel_shape=[4, 4],
pool_shape=[2, 2],
pool_stride=[2, 2],
max_kernel_norm=1.9365,
irange=.025)
l2 = MaxoutConvC01B(layer_name='l2',
tied_b=1,
num_channels=64,
num_pieces=2,
pad=3,
kernel_shape=[4, 4],
pool_shape=[2, 2],
pool_stride=[2, 2],
max_kernel_norm=1.9365,
irange=.025)
l3 = MaxoutConvC01B(layer_name='l3',
import theano
from adversarial.deconv import Deconv
from pylearn2.datasets.mnist import MNIST
from pylearn2.space import Conv2DSpace
from pylearn2.models.mlp import MLP
from pylearn2.models.maxout import MaxoutConvC01B
from pylearn2.gui import patch_viewer
import ipdb
input_space = Conv2DSpace(shape = (28, 28), num_channels=1, axes = ('c', 0, 1, 'b'))
conv = MaxoutConvC01B(layer_name = 'conv',
num_channels = 16,
num_pieces = 1,
kernel_shape = (4, 4),
pool_shape = (1, 1),
pool_stride=(1, 1),
irange = 0.05)
deconv = Deconv(layer_name = 'deconv',
num_channels = 1,
kernel_shape = (4, 4),
irange = 0.05)
mlp = MLP(input_space =input_space,
layers = [conv, deconv])
mlp.layers[1].transformer._filters.set_value(mlp.layers[0].transformer._filters.get_value())
x = input_space.get_theano_batch()
out = mlp.fprop(x)