def test_mkl_lrn_forward():
if theano.config.floatX == 'float32':
x = tensor.ftensor4()
else:
x = tensor.dtensor4()
y = lrn(x)
f = theano.function([x], y, mode=mode_with_mkl)
topo = f.maker.fgraph.toposort()
inputs = f.maker.fgraph.inputs
outputs = f.maker.fgraph.outputs
assert len(inputs) == 1
assert len(outputs) == 1
assert len(topo) == 3
assert isinstance(topo[0].op, U2ILRN)
assert isinstance(topo[1].op, mkl.mkl_lrn.LRN)
assert isinstance(topo[2].op, I2U)
assert outputs[0].owner == topo[2]
imval = numpy.random.rand(4, 2, 4, 4).astype(theano.config.floatX)
f(imval)
print('test_mkl_lrn_forward() pass..')
def test_mkl_lrn_forward():
if theano.config.floatX == 'float32':
x = tensor.ftensor4()
else:
x = tensor.dtensor4()
y = lrn(x)
f = theano.function([x], y, mode=mode_with_mkl)
topo = f.maker.fgraph.toposort()
inputs = f.maker.fgraph.inputs
outputs = f.maker.fgraph.outputs
assert len(inputs) == 1
assert len(outputs) == 1
assert len(topo) == 3
assert isinstance(topo[0].op, U2ILRN)
assert isinstance(topo[1].op, mkl.mkl_lrn.LRN)
assert isinstance(topo[2].op, I2U)
assert outputs[0].owner == topo[2]
imval = numpy.random.rand(4, 2, 4, 4).astype(theano.config.floatX)
f(imval)
print('test_mkl_lrn_forward() pass..')
def __init__(self, input, image_shape, filter_shape, convstride, padsize,
group, poolsize, poolstride, bias_init, lrn=False,
):
'''
conv, pooling, relu and norm layers
'''
self.filter_size = filter_shape
self.convstride = convstride
self.padsize = padsize
self.poolsize = poolsize
self.poolstride = poolstride
self.channel = image_shape[1]
self.lrn = lrn
assert group in [1, 2]
self.filter_shape = list(filter_shape)
self.image_shape = list(image_shape)
if group == 1:
self.W = Weight(self.filter_shape)
self.b = Weight(self.filter_shape[0], bias_init, std=0)
new_filter_shape = self.filter_shape
else:
# making new filter shape 5D tensor, when group is more than 1
new_filter_shape = (group, self.filter_shape[0] / group,
self.filter_shape[1] / group, self.filter_shape[2],
self.filter_shape[3])
self.W = Weight(new_filter_shape)
self.b = Weight(self.filter_shape[0], bias_init, std=0)
conv_out = AbstractConvGroup(imshp=image_shape,
kshp=new_filter_shape,
subsample=(convstride, convstride),
border_mode=(padsize, padsize),
filter_flip=False,
group=group)(input, self.W.val, self.b.val)
# ReLu
self.output = T.nnet.relu(conv_out, 0)
# LRN
if self.lrn:
self.output = LRN.lrn(self.output, alpha=1e-4, beta=0.75, k=1, n=5)
# Pooling
if self.poolsize != 1:
self.output = pool.pool_2d(input=self.output,
ds=(poolsize, poolsize),
ignore_border=False,
st=(poolstride, poolstride),
mode='max')
self.params = [self.W.val, self.b.val]
self.weight_type = ['W', 'b']
print "conv layer with shape_in: {}".format(str(image_shape))
def test_mkl_lrn_backward():
predefineOps = [
U2ILRN, mkl.mkl_lrn.LRN, I2U, Shape_i, Shape_i, Shape_i, Shape_i,
Alloc, I2UGrad, mkl.mkl_lrn.LRNGrad, U2IGrad
]
if theano.config.floatX == 'float32':
x = tensor.ftensor4()
else:
x = tensor.dtensor4()
y = lrn(x)
z = tensor.grad(tensor.sum(y), [x])
f = theano.function([x], z, mode=mode_with_mkl)
topo = f.maker.fgraph.toposort()
inputs = f.maker.fgraph.inputs
outputs = f.maker.fgraph.outputs
assert len(inputs) == 1
assert len(outputs) == 1
assert len(topo) == 11
for i, node in enumerate(topo):
assert isinstance(node.op, predefineOps[i])
# U2I_LRN
assert len(topo[0].inputs) == 1
assert topo[0].inputs[0] == inputs[0]
# LRN
assert len(topo[1].inputs) == 1
assert topo[1].inputs[0].owner == topo[0]
# I2UGrad
assert len(topo[8].inputs) == 2
assert topo[8].inputs[0].owner == topo[1]
assert topo[8].inputs[1].owner == topo[7]
# LRNGrad
assert len(topo[9].inputs) == 2
assert topo[9].inputs[0].owner == topo[0]
assert topo[9].inputs[1].owner == topo[8]
# U2IGrad
assert len(topo[10].inputs) == 2
assert topo[10].inputs[0] == inputs[0]
assert topo[10].inputs[1].owner == topo[9]
# Output
assert outputs[0].owner == topo[10]
imval = numpy.random.rand(4, 2, 4, 4).astype(theano.config.floatX)
f(imval)
print('test_mkl_lrn_backward() pass..')
def test_mkl_lrn_backward():
predefineOps = [U2ILRN, mkl.mkl_lrn.LRN, I2U,
Shape_i, Shape_i, Shape_i, Shape_i, Alloc, I2UGrad, mkl.mkl_lrn.LRNGrad, U2IGrad]
if theano.config.floatX == 'float32':
x = tensor.ftensor4()
else:
x = tensor.dtensor4()
y = lrn(x)
z = tensor.grad(tensor.sum(y), [x])
f = theano.function([x], z, mode=mode_with_mkl)
topo = f.maker.fgraph.toposort()
inputs = f.maker.fgraph.inputs
outputs = f.maker.fgraph.outputs
assert len(inputs) == 1
assert len(outputs) == 1
assert len(topo) == 11
for i, node in enumerate(topo):
assert isinstance(node.op, predefineOps[i])
# U2I_LRN
assert len(topo[0].inputs) == 1
assert topo[0].inputs[0] == inputs[0]
# LRN
assert len(topo[1].inputs) == 1
assert topo[1].inputs[0].owner == topo[0]
# I2UGrad
assert len(topo[8].inputs) == 2
assert topo[8].inputs[0].owner == topo[1]
assert topo[8].inputs[1].owner == topo[7]
# LRNGrad
assert len(topo[9].inputs) == 2
assert topo[9].inputs[0].owner == topo[0]
assert topo[9].inputs[1].owner == topo[8]
# U2IGrad
assert len(topo[10].inputs) == 2
assert topo[10].inputs[0] == inputs[0]
assert topo[10].inputs[1].owner == topo[9]
# Output
assert outputs[0].owner == topo[10]
imval = numpy.random.rand(4, 2, 4, 4).astype(theano.config.floatX)
f(imval)
print('test_mkl_lrn_backward() pass..')
import theano
from theano import tensor as T
from theano.sandbox import mkl
from theano.tensor.nnet.lrn import lrn
from nose.plugins.skip import SkipTest
if not mkl.mkl_available:
raise SkipTest('Optional package MKL disabled')
x = T.ftensor4()
y = lrn(x)
theano.printing.pydotprint(y,
outfile='lrn_fwd_before.png',
var_with_name_simple=True)
f = theano.function([x], y)
theano.printing.pydotprint(f,
outfile='lrn_fwd_after.png',
var_with_name_simple=True)
z = T.grad(T.sum(y), [x])
theano.printing.pydotprint(z,
outfile='lrn_bwd_before.png',
var_with_name_simple=True)
f1 = theano.function([x], z)
theano.printing.pydotprint(f1,
outfile='lrn_bwd_after.png',
var_with_name_simple=True)
def __init__(
self,
input,
image_shape,
filter_shape,
convstride,
padsize,
group,
poolsize,
poolstride,
bias_init,
lrn=False,
):
'''
conv, pooling, relu and norm layers
'''
self.filter_size = filter_shape
self.convstride = convstride
self.padsize = padsize
self.poolsize = poolsize
self.poolstride = poolstride
self.channel = image_shape[1]
self.lrn = lrn
assert group in [1, 2]
self.filter_shape = list(filter_shape)
self.image_shape = list(image_shape)
if group == 1:
self.W = Weight(self.filter_shape)
self.b = Weight(self.filter_shape[0], bias_init, std=0)
new_filter_shape = self.filter_shape
else:
# making new filter shape 5D tensor, when group is more than 1
new_filter_shape = (group, self.filter_shape[0] / group,
self.filter_shape[1] / group,
self.filter_shape[2], self.filter_shape[3])
self.W = Weight(new_filter_shape)
self.b = Weight(self.filter_shape[0], bias_init, std=0)
conv_out = AbstractConvGroup(imshp=image_shape,
kshp=new_filter_shape,
subsample=(convstride, convstride),
border_mode=(padsize, padsize),
filter_flip=False,
group=group)(input, self.W.val,
self.b.val)
# ReLu
self.output = T.nnet.relu(conv_out, 0)
# LRN
if self.lrn:
self.output = LRN.lrn(self.output, alpha=1e-4, beta=0.75, k=1, n=5)
# Pooling
if self.poolsize != 1:
self.output = pool.pool_2d(input=self.output,
ds=(poolsize, poolsize),
ignore_border=False,
st=(poolstride, poolstride),
mode='max')
self.params = [self.W.val, self.b.val]
self.weight_type = ['W', 'b']
print "conv layer with shape_in: {}".format(str(image_shape))
import theano
from theano import tensor as T
from theano.sandbox import mkl
from theano.tensor.nnet.lrn import lrn
from nose.plugins.skip import SkipTest
if not mkl.mkl_available:
raise SkipTest('Optional package MKL disabled')
x = T.ftensor4()
y = lrn(x)
theano.printing.pydotprint(y, outfile='lrn_fwd_before.png', var_with_name_simple=True)
f = theano.function([x], y)
theano.printing.pydotprint(f, outfile='lrn_fwd_after.png', var_with_name_simple=True)
z = T.grad(T.sum(y), [x])
theano.printing.pydotprint(z, outfile='lrn_bwd_before.png', var_with_name_simple=True)
f1 = theano.function([x], z)
theano.printing.pydotprint(f1, outfile='lrn_bwd_after.png', var_with_name_simple=True)