Skip to content

P
pytensor
提交 8d8f527f authored 作者: vdumoulin's avatar vdumoulin
浏览文件
操作

Merge pull request #3621 from shabanian/tensor_signal_pep8

pep8 tensor signal list directory
上级 805991f1 c81f8c36
隐藏空白字符变更
内嵌 并排
正在显示 包含 93 行增加 和 91 行删除
theano/tensor/signal/conv.py
浏览文件 @ 8d8f527f
...@@ -3,8 +3,6 @@ Contains a wrapper function for tensor.nnet.ConvOp, which can be used to perform ...@@ -3,8 +3,6 @@ Contains a wrapper function for tensor.nnet.ConvOp, which can be used to perform
generic 2D convolution. generic 2D convolution.
""" """
__docformat__ = "restructuredtext en"
import warnings import warnings
import theano import theano
...@@ -12,6 +10,10 @@ import theano.tensor as tensor ...@@ -12,6 +10,10 @@ import theano.tensor as tensor
from theano.tensor.nnet import conv from theano.tensor.nnet import conv
import logging import logging
__docformat__ = "restructuredtext en"
_logger = logging.getLogger("theano.tensor.signal.conv") _logger = logging.getLogger("theano.tensor.signal.conv")
...@@ -52,7 +54,7 @@ def conv2d(input, filters, image_shape=None, filter_shape=None, ...@@ -52,7 +54,7 @@ def conv2d(input, filters, image_shape=None, filter_shape=None,
assert input.ndim in (2, 3) assert input.ndim in (2, 3)
assert filters.ndim in (2, 3) assert filters.ndim in (2, 3)
### use shape information if it is given to us ### # use shape information if it is given to us ###
if filter_shape and image_shape: if filter_shape and image_shape:
if input.ndim == 3: if input.ndim == 3:
bsize = image_shape[0] bsize = image_shape[0]
...@@ -69,7 +71,7 @@ def conv2d(input, filters, image_shape=None, filter_shape=None, ...@@ -69,7 +71,7 @@ def conv2d(input, filters, image_shape=None, filter_shape=None,
nkern, kshp = None, None nkern, kshp = None, None
bsize, imshp = None, None bsize, imshp = None, None
### reshape tensors to 4D, for compatibility with ConvOp ### # reshape tensors to 4D, for compatibility with ConvOp ###
if input.ndim == 3: if input.ndim == 3:
sym_bsize = input.shape[0] sym_bsize = input.shape[0]
else: else:
...@@ -86,10 +88,10 @@ def conv2d(input, filters, image_shape=None, filter_shape=None, ...@@ -86,10 +88,10 @@ def conv2d(input, filters, image_shape=None, filter_shape=None,
new_filter_shape = tensor.join(0, tensor.stack([sym_nkern, 1]), filters.shape[-2:]) new_filter_shape = tensor.join(0, tensor.stack([sym_nkern, 1]), filters.shape[-2:])
filters4D = tensor.reshape(filters, new_filter_shape, ndim=4) filters4D = tensor.reshape(filters, new_filter_shape, ndim=4)
### perform actual convolution ### # perform actual convolution ###
op = conv.ConvOp(output_mode=border_mode, op = conv.ConvOp(output_mode=border_mode,
dx=subsample[0], dy=subsample[1], dx=subsample[0], dy=subsample[1],
imshp=imshp, kshp=kshp, nkern=nkern, bsize=bsize, **kargs) imshp=imshp, kshp=kshp, nkern=nkern, bsize=bsize, **kargs)
output = op(input4D, filters4D) output = op(input4D, filters4D)
......
theano/tensor/signal/downsample.py
浏览文件 @ 8d8f527f
""" """
Ops for downsampling images. Ops for downsampling images.
Planned: Planned:
...@@ -18,6 +18,7 @@ from theano import gof, Op, tensor, Variable, Apply ...@@ -18,6 +18,7 @@ from theano import gof, Op, tensor, Variable, Apply
from theano.tensor.opt import register_canonicalize from theano.tensor.opt import register_canonicalize
def max_pool2D(*args, **kwargs): def max_pool2D(*args, **kwargs):
import sys import sys
print("DEPRECATION: max_pool2D renamed to max_pool_2d", file=sys.stderr) print("DEPRECATION: max_pool2D renamed to max_pool_2d", file=sys.stderr)
...@@ -206,8 +207,8 @@ class DownsampleFactorMax(Op): ...@@ -206,8 +207,8 @@ class DownsampleFactorMax(Op):
if isinstance(r, theano.Variable): if isinstance(r, theano.Variable):
nr = tensor.switch(tensor.ge(st[0], ds[0]), nr = tensor.switch(tensor.ge(st[0], ds[0]),
(r - 1) // st[0] + 1, (r - 1) // st[0] + 1,
tensor.maximum(0, (r - 1 - ds[0]) tensor.maximum(0, (r - 1 - ds[0]) //
// st[0] + 1) + 1) st[0] + 1) + 1)
elif st[0] >= ds[0]: elif st[0] >= ds[0]:
nr = (r - 1) // st[0] + 1 nr = (r - 1) // st[0] + 1
else: else:
...@@ -216,8 +217,8 @@ class DownsampleFactorMax(Op): ...@@ -216,8 +217,8 @@ class DownsampleFactorMax(Op):
if isinstance(c, theano.Variable): if isinstance(c, theano.Variable):
nc = tensor.switch(tensor.ge(st[1], ds[1]), nc = tensor.switch(tensor.ge(st[1], ds[1]),
(c - 1) // st[1] + 1, (c - 1) // st[1] + 1,
tensor.maximum(0, (c - 1 - ds[1]) tensor.maximum(0, (c - 1 - ds[1]) //
// st[1] + 1) + 1) st[1] + 1) + 1)
elif st[1] >= ds[1]: elif st[1] >= ds[1]:
nc = (c - 1) // st[1] + 1 nc = (c - 1) // st[1] + 1
else: else:
...@@ -289,7 +290,7 @@ class DownsampleFactorMax(Op): ...@@ -289,7 +290,7 @@ class DownsampleFactorMax(Op):
y = numpy.zeros( y = numpy.zeros(
(x.shape[0], x.shape[1], img_rows, img_cols), (x.shape[0], x.shape[1], img_rows, img_cols),
dtype=x.dtype) dtype=x.dtype)
y[:, :, pad_h:(img_rows-pad_h), pad_w:(img_cols-pad_w)] = x y[:, :, pad_h:(img_rows - pad_h), pad_w:(img_cols - pad_w)] = x
else: else:
y = x y = x
func = numpy.max func = numpy.max
...@@ -312,7 +313,7 @@ class DownsampleFactorMax(Op): ...@@ -312,7 +313,7 @@ class DownsampleFactorMax(Op):
if not inc_pad: if not inc_pad:
col_st = builtins.max(col_st, self.padding[1]) col_st = builtins.max(col_st, self.padding[1])
col_end = builtins.min(col_end, col_end = builtins.min(col_end,
x.shape[-1] + pad_w) x.shape[-1] + pad_w)
zz[n, k, r, c] = func(y[ zz[n, k, r, c] = func(y[
n, k, row_st:row_end, col_st:col_end]) n, k, row_st:row_end, col_st:col_end])
...@@ -336,6 +337,7 @@ class DownsampleFactorMax(Op): ...@@ -336,6 +337,7 @@ class DownsampleFactorMax(Op):
st=self.st, padding=self.padding, st=self.st, padding=self.padding,
mode=self.mode)( mode=self.mode)(
x, gz)] x, gz)]
def c_headers(self): def c_headers(self):
return ['<algorithm>'] return ['<algorithm>']
...@@ -522,6 +524,7 @@ class DownsampleFactorMax(Op): ...@@ -522,6 +524,7 @@ class DownsampleFactorMax(Op):
def c_code_cache_version(self): def c_code_cache_version(self):
return (0, 6, 8, 3) return (0, 6, 8, 3)
class PoolGrad(Op): class PoolGrad(Op):
__props__ = ('ds', 'ignore_border', 'st', 'padding', 'mode') __props__ = ('ds', 'ignore_border', 'st', 'padding', 'mode')
...@@ -582,8 +585,8 @@ class PoolGrad(Op): ...@@ -582,8 +585,8 @@ class PoolGrad(Op):
if isinstance(r, theano.Variable): if isinstance(r, theano.Variable):
nr = tensor.switch(tensor.ge(st[0], ds[0]), nr = tensor.switch(tensor.ge(st[0], ds[0]),
(r - 1) // st[0] + 1, (r - 1) // st[0] + 1,
tensor.maximum(0, (r - 1 - ds[0]) tensor.maximum(0, (r - 1 - ds[0]) //
// st[0] + 1) + 1) st[0] + 1) + 1)
elif st[0] >= ds[0]: elif st[0] >= ds[0]:
nr = (r - 1) // st[0] + 1 nr = (r - 1) // st[0] + 1
else: else:
...@@ -592,8 +595,8 @@ class PoolGrad(Op): ...@@ -592,8 +595,8 @@ class PoolGrad(Op):
if isinstance(c, theano.Variable): if isinstance(c, theano.Variable):
nc = tensor.switch(tensor.ge(st[1], ds[1]), nc = tensor.switch(tensor.ge(st[1], ds[1]),
(c - 1) // st[1] + 1, (c - 1) // st[1] + 1,
tensor.maximum(0, (c - 1 - ds[1]) tensor.maximum(0, (c - 1 - ds[1]) //
// st[1] + 1) + 1) st[1] + 1) + 1)
elif st[1] >= ds[1]: elif st[1] >= ds[1]:
nc = (c - 1) // st[1] + 1 nc = (c - 1) // st[1] + 1
else: else:
...@@ -656,7 +659,7 @@ class MaxPoolGrad(PoolGrad): ...@@ -656,7 +659,7 @@ class MaxPoolGrad(PoolGrad):
y = numpy.zeros( y = numpy.zeros(
(x.shape[0], x.shape[1], img_rows, img_cols), (x.shape[0], x.shape[1], img_rows, img_cols),
dtype=x.dtype) dtype=x.dtype)
y[:, :, pad_h:(img_rows-pad_h), pad_w:(img_cols-pad_w)] = x y[:, :, pad_h:(img_rows - pad_h), pad_w:(img_cols - pad_w)] = x
else: else:
y = x y = x
gx = numpy.zeros_like(y) gx = numpy.zeros_like(y)
...@@ -673,7 +676,7 @@ class MaxPoolGrad(PoolGrad): ...@@ -673,7 +676,7 @@ class MaxPoolGrad(PoolGrad):
if (maxout[n, k, r, c] == y[n, k, row_ind, col_ind]): if (maxout[n, k, r, c] == y[n, k, row_ind, col_ind]):
gx[n, k, row_ind, col_ind] += gz[n, k, r, c] gx[n, k, row_ind, col_ind] += gz[n, k, r, c]
# unpad the image # unpad the image
gx = gx[:, :, pad_h:(img_rows-pad_h), pad_w:(img_cols-pad_w)] gx = gx[:, :, pad_h:(img_rows - pad_h), pad_w:(img_cols - pad_w)]
gx_stg[0] = gx gx_stg[0] = gx
def grad(self, inp, grads): def grad(self, inp, grads):
...@@ -804,6 +807,7 @@ class MaxPoolGrad(PoolGrad): ...@@ -804,6 +807,7 @@ class MaxPoolGrad(PoolGrad):
DownsampleFactorMaxGrad = MaxPoolGrad DownsampleFactorMaxGrad = MaxPoolGrad
class AveragePoolGrad(PoolGrad): class AveragePoolGrad(PoolGrad):
def __init__(self, ds, ignore_border, st=None, padding=(0, 0), mode='average_inc_pad'): def __init__(self, ds, ignore_border, st=None, padding=(0, 0), mode='average_inc_pad'):
...@@ -848,7 +852,7 @@ class AveragePoolGrad(PoolGrad): ...@@ -848,7 +852,7 @@ class AveragePoolGrad(PoolGrad):
y = numpy.zeros( y = numpy.zeros(
(x.shape[0], x.shape[1], img_rows, img_cols), (x.shape[0], x.shape[1], img_rows, img_cols),
dtype=x.dtype) dtype=x.dtype)
y[:, :, pad_h:(img_rows-pad_h), pad_w:(img_cols-pad_w)] = x y[:, :, pad_h:(img_rows - pad_h), pad_w:(img_cols - pad_w)] = x
else: else:
y = x y = x
gx = numpy.zeros_like(y) gx = numpy.zeros_like(y)
...@@ -865,16 +869,16 @@ class AveragePoolGrad(PoolGrad): ...@@ -865,16 +869,16 @@ class AveragePoolGrad(PoolGrad):
col_st = c * st1 col_st = c * st1
else: else:
col_st = builtins.max(c * st1, col_st = builtins.max(c * st1,
self.padding[1]) self.padding[1])
col_end = builtins.min(col_st + ds1, img_cols) col_end = builtins.min(col_st + ds1, img_cols)
if sum_mode: if sum_mode:
val = gz[n, k, r, c] val = gz[n, k, r, c]
else: else:
val = gz[n, k, r, c] / ((row_end - row_st) * val = gz[n, k, r, c] / ((row_end - row_st) *
(col_end - col_st)) (col_end - col_st))
gx[n, k, row_st:row_end, col_st:col_end] += val gx[n, k, row_st:row_end, col_st:col_end] += val
# unpad the image # unpad the image
gx = gx[:, :, pad_h:(img_rows-pad_h), pad_w:(img_cols-pad_w)] gx = gx[:, :, pad_h:(img_rows - pad_h), pad_w:(img_cols - pad_w)]
gx_stg[0] = gx gx_stg[0] = gx
def grad(self, inp, grads): def grad(self, inp, grads):
...@@ -885,6 +889,7 @@ class AveragePoolGrad(PoolGrad): ...@@ -885,6 +889,7 @@ class AveragePoolGrad(PoolGrad):
self.ds, ignore_border=self.ignore_border, self.ds, ignore_border=self.ignore_border,
st=self.st, padding=self.padding, mode=self.mode)(ggx)] st=self.st, padding=self.padding, mode=self.mode)(ggx)]
class DownsampleFactorMaxGradGrad(Op): class DownsampleFactorMaxGradGrad(Op):
__props__ = ('ds', 'ignore_border', 'st', 'padding', 'mode') __props__ = ('ds', 'ignore_border', 'st', 'padding', 'mode')
...@@ -893,7 +898,7 @@ class DownsampleFactorMaxGradGrad(Op): ...@@ -893,7 +898,7 @@ class DownsampleFactorMaxGradGrad(Op):
""" """
Return the shape of the output from this op, for input of given Return the shape of the output from this op, for input of given
shape and flags. shape and flags.
Parameters Parameters
---------- ----------
imgshape : tuple, list, or similar of integer or scalar Theano variable imgshape : tuple, list, or similar of integer or scalar Theano variable
...@@ -946,8 +951,8 @@ class DownsampleFactorMaxGradGrad(Op): ...@@ -946,8 +951,8 @@ class DownsampleFactorMaxGradGrad(Op):
if isinstance(r, theano.Variable): if isinstance(r, theano.Variable):
nr = tensor.switch(tensor.ge(st[0], ds[0]), nr = tensor.switch(tensor.ge(st[0], ds[0]),
(r - 1) // st[0] + 1, (r - 1) // st[0] + 1,
tensor.maximum(0, (r - 1 - ds[0]) tensor.maximum(0, (r - 1 - ds[0]) //
// st[0] + 1) + 1) st[0] + 1) + 1)
elif st[0] >= ds[0]: elif st[0] >= ds[0]:
nr = (r - 1) // st[0] + 1 nr = (r - 1) // st[0] + 1
else: else:
...@@ -956,8 +961,8 @@ class DownsampleFactorMaxGradGrad(Op): ...@@ -956,8 +961,8 @@ class DownsampleFactorMaxGradGrad(Op):
if isinstance(c, theano.Variable): if isinstance(c, theano.Variable):
nc = tensor.switch(tensor.ge(st[1], ds[1]), nc = tensor.switch(tensor.ge(st[1], ds[1]),
(c - 1) // st[1] + 1, (c - 1) // st[1] + 1,
tensor.maximum(0, (c - 1 - ds[1]) tensor.maximum(0, (c - 1 - ds[1]) //
// st[1] + 1) + 1) st[1] + 1) + 1)
elif st[1] >= ds[1]: elif st[1] >= ds[1]:
nc = (c - 1) // st[1] + 1 nc = (c - 1) // st[1] + 1
else: else:
...@@ -966,7 +971,7 @@ class DownsampleFactorMaxGradGrad(Op): ...@@ -966,7 +971,7 @@ class DownsampleFactorMaxGradGrad(Op):
rval = list(imgshape[:-2]) + [nr, nc] rval = list(imgshape[:-2]) + [nr, nc]
return rval return rval
def __init__(self, ds, ignore_border, st=None, padding=(0,0), mode='max'): def __init__(self, ds, ignore_border, st=None, padding=(0, 0), mode='max'):
self.ds = tuple(ds) self.ds = tuple(ds)
if not all([isinstance(d, int) for d in ds]): if not all([isinstance(d, int) for d in ds]):
raise ValueError( raise ValueError(
...@@ -1009,7 +1014,7 @@ class DownsampleFactorMaxGradGrad(Op): ...@@ -1009,7 +1014,7 @@ class DownsampleFactorMaxGradGrad(Op):
self.st, self.padding) self.st, self.padding)
if (z[0] is None) or (z[0].shape != z_shape): if (z[0] is None) or (z[0].shape != z_shape):
z[0] = numpy.zeros(z_shape, dtype=x.dtype) z[0] = numpy.zeros(z_shape, dtype=x.dtype)
ggz = z[0] # grad wrt maxout_grad has the same shape as maxout ggz = z[0] # grad wrt maxout_grad has the same shape as maxout
# number of pooling output rows # number of pooling output rows
pr = ggz.shape[-2] pr = ggz.shape[-2]
# number of pooling output cols # number of pooling output cols
...@@ -1025,11 +1030,11 @@ class DownsampleFactorMaxGradGrad(Op): ...@@ -1025,11 +1030,11 @@ class DownsampleFactorMaxGradGrad(Op):
y_padded = numpy.zeros( y_padded = numpy.zeros(
(x.shape[0], x.shape[1], img_rows, img_cols), (x.shape[0], x.shape[1], img_rows, img_cols),
dtype=x.dtype) + x.min() - 1 dtype=x.dtype) + x.min() - 1
y_padded[:, :, pd0:(img_rows-pd0), pd1:(img_cols-pd1)] = x y_padded[:, :, pd0:(img_rows - pd0), pd1:(img_cols - pd1)] = x
ggx_padded = numpy.zeros( ggx_padded = numpy.zeros(
(x.shape[0], x.shape[1], img_rows, img_cols), (x.shape[0], x.shape[1], img_rows, img_cols),
dtype=x.dtype) dtype=x.dtype)
ggx_padded[:, :, pd0:(img_rows-pd0), pd1:(img_cols-pd1)] = ggx ggx_padded[:, :, pd0:(img_rows - pd0), pd1:(img_cols - pd1)] = ggx
else: else:
y_padded = x y_padded = x
...@@ -1054,7 +1059,7 @@ class DownsampleFactorMaxGradGrad(Op): ...@@ -1054,7 +1059,7 @@ class DownsampleFactorMaxGradGrad(Op):
if self.mode != 'max': if self.mode != 'max':
raise theano.gof.utils.MethodNotDefined() raise theano.gof.utils.MethodNotDefined()
x, maxout, ggx = inp x, maxout, ggx = inp
z, = out # the grad of grad z, = out # the grad of grad
fail = sub['fail'] fail = sub['fail']
ignore_border = int(self.ignore_border) ignore_border = int(self.ignore_border)
ds0, ds1 = self.ds ds0, ds1 = self.ds
...@@ -1130,10 +1135,11 @@ class DownsampleFactorMaxGradGrad(Op): ...@@ -1130,10 +1135,11 @@ class DownsampleFactorMaxGradGrad(Op):
} }
} }
} }
"""%locals() """ % locals()
def c_code_cache_version(self): def c_code_cache_version(self):
return (0,1) return (0, 1)
@register_canonicalize('fast_compile') @register_canonicalize('fast_compile')
@gof.local_optimizer([MaxPoolGrad]) @gof.local_optimizer([MaxPoolGrad])
...@@ -1141,7 +1147,7 @@ def local_average_pool_grad(node): ...@@ -1141,7 +1147,7 @@ def local_average_pool_grad(node):
# To assure backward compatibility with # To assure backward compatibility with
# DownsampleFactorMaxGrad # DownsampleFactorMaxGrad
if (not isinstance(node.op, MaxPoolGrad) or node.op.mode not in if (not isinstance(node.op, MaxPoolGrad) or node.op.mode not in
['sum','average_exc_pad', 'average_inc_pad']): ['sum', 'average_exc_pad', 'average_inc_pad']):
return False return False
return [AveragePoolGrad(ds=node.op.ds, return [AveragePoolGrad(ds=node.op.ds,
ignore_border=node.op.ignore_border, ignore_border=node.op.ignore_border,
......
theano/tensor/signal/tests/test_conv.py
浏览文件 @ 8d8f527f
...@@ -31,7 +31,7 @@ class TestSignalConv2D(unittest.TestCase): ...@@ -31,7 +31,7 @@ class TestSignalConv2D(unittest.TestCase):
if filter_dim != 3: if filter_dim != 3:
nkern = 1 nkern = 1
############# THEANO IMPLEMENTATION ############ # THEANO IMPLEMENTATION ############
# we create a symbolic function so that verify_grad can work # we create a symbolic function so that verify_grad can work
def sym_conv2d(input, filters): def sym_conv2d(input, filters):
return conv.conv2d(input, filters) return conv.conv2d(input, filters)
...@@ -44,9 +44,8 @@ class TestSignalConv2D(unittest.TestCase): ...@@ -44,9 +44,8 @@ class TestSignalConv2D(unittest.TestCase):
filter_data = numpy.random.random(filter_shape) filter_data = numpy.random.random(filter_shape)
theano_output = theano_conv(image_data, filter_data) theano_output = theano_conv(image_data, filter_data)
############# REFERENCE IMPLEMENTATION ############ # REFERENCE IMPLEMENTATION ############
out_shape2d = numpy.array(image_shape[-2:]) -\ out_shape2d = numpy.array(image_shape[-2:]) - numpy.array(filter_shape[-2:]) + 1
numpy.array(filter_shape[-2:]) + 1
ref_output = numpy.zeros(tuple(out_shape2d)) ref_output = numpy.zeros(tuple(out_shape2d))
# reshape as 3D input tensors to make life easier # reshape as 3D input tensors to make life easier
...@@ -76,7 +75,7 @@ class TestSignalConv2D(unittest.TestCase): ...@@ -76,7 +75,7 @@ class TestSignalConv2D(unittest.TestCase):
self.assertTrue(_allclose(theano_output4d[b, k, :, :], self.assertTrue(_allclose(theano_output4d[b, k, :, :],
output2d)) output2d))
############# TEST GRADIENT ############ # TEST GRADIENT ############
if verify_grad: if verify_grad:
utt.verify_grad(sym_conv2d, [image_data, filter_data]) utt.verify_grad(sym_conv2d, [image_data, filter_data])
...@@ -87,8 +86,8 @@ class TestSignalConv2D(unittest.TestCase): ...@@ -87,8 +86,8 @@ class TestSignalConv2D(unittest.TestCase):
signal.conv.conv2d can support inputs and filters of type signal.conv.conv2d can support inputs and filters of type
matrix or tensor3. matrix or tensor3.
""" """
if (not theano.tensor.nnet.conv.imported_scipy_signal and if(not theano.tensor.nnet.conv.imported_scipy_signal and
theano.config.cxx == ""): theano.config.cxx == ""):
raise SkipTest("conv2d tests need SciPy or a c++ compiler") raise SkipTest("conv2d tests need SciPy or a c++ compiler")
self.validate((1, 4, 5), (2, 2, 3), out_dim=4, verify_grad=True) self.validate((1, 4, 5), (2, 2, 3), out_dim=4, verify_grad=True)
......
theano/tensor/signal/tests/test_downsample.py
浏览文件 @ 8d8f527f
...@@ -63,9 +63,9 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -63,9 +63,9 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
def pad_img(x): def pad_img(x):
y = numpy.zeros( y = numpy.zeros(
(x.shape[0], x.shape[1], (x.shape[0], x.shape[1],
x.shape[2]+pad_h*2, x.shape[3]+pad_w*2), x.shape[2] + pad_h * 2, x.shape[3] + pad_w * 2),
dtype=x.dtype) dtype=x.dtype)
y[:, :, pad_h:(x.shape[2]+pad_h), pad_w:(x.shape[3]+pad_w)] = x y[:, :, pad_h:(x.shape[2] + pad_h), pad_w:(x.shape[3] + pad_w)] = x
return y return y
img_rows = h + 2 * pad_h img_rows = h + 2 * pad_h
...@@ -78,7 +78,6 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -78,7 +78,6 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
ds0, ds1 = ds ds0, ds1 = ds
st0, st1 = st st0, st1 = st
output_val = numpy.zeros(out_shp) output_val = numpy.zeros(out_shp)
tt = []
y = pad_img(x) y = pad_img(x)
func = numpy.max func = numpy.max
if mode == 'sum': if mode == 'sum':
...@@ -117,8 +116,6 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -117,8 +116,6 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
if st is None: if st is None:
st = ds st = ds
xi = 0
yi = 0
img_rows = input.shape[-2] img_rows = input.shape[-2]
img_cols = input.shape[-1] img_cols = input.shape[-1]
...@@ -364,12 +361,12 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -364,12 +361,12 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
imval = rng.rand(1, 2, 16, 16) imval = rng.rand(1, 2, 16, 16)
for maxpoolshp, ignore_border, mode, stride in product(maxpoolshps, for maxpoolshp, ignore_border, mode, stride in product(maxpoolshps,
[True, False], [True, False],
['max', ['max',
'sum', 'sum',
'average_inc_pad', 'average_inc_pad',
'average_exc_pad'], 'average_exc_pad'],
stridesizes): stridesizes):
def mp(input): def mp(input):
return DownsampleFactorMax(maxpoolshp, return DownsampleFactorMax(maxpoolshp,
ignore_border=ignore_border, ignore_border=ignore_border,
...@@ -581,10 +578,12 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -581,10 +578,12 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
stridesize = stridesizes[i] stridesize = stridesizes[i]
paddingsize = paddingsizes[i] paddingsize = paddingsizes[i]
grad_shape = DownsampleFactorMaxGradGrad.out_shape( grad_shape = DownsampleFactorMaxGradGrad.out_shape(imval.shape,
imval.shape, maxpoolsize, st=stridesize, maxpoolsize, st=stridesize,
ignore_border=True, padding=paddingsize) ignore_border=True,
padding=paddingsize)
grad_val = rng.rand(*grad_shape) * 10.0 grad_val = rng.rand(*grad_shape) * 10.0
def mp(input, grad): def mp(input, grad):
out = DownsampleFactorMax( out = DownsampleFactorMax(
maxpoolsize, ignore_border=True, maxpoolsize, ignore_border=True,
...@@ -592,7 +591,7 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -592,7 +591,7 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
padding=paddingsize, padding=paddingsize,
)(input) )(input)
grad_op = MaxPoolGrad(maxpoolsize, ignore_border=True, grad_op = MaxPoolGrad(maxpoolsize, ignore_border=True,
st=stridesize, padding=paddingsize) st=stridesize, padding=paddingsize)
return grad_op(input, out, grad) return grad_op(input, out, grad)
utt.verify_grad(mp, [imval, grad_val], rng=rng) utt.verify_grad(mp, [imval, grad_val], rng=rng)
...@@ -610,16 +609,17 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -610,16 +609,17 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
stridesize = stridesizes[i] stridesize = stridesizes[i]
paddingsize = paddingsizes[i] paddingsize = paddingsizes[i]
#'average_exc_pad' with non-zero padding is not implemented # 'average_exc_pad' with non-zero padding is not implemented
for mode in ['sum', 'average_inc_pad']: for mode in ['sum', 'average_inc_pad']:
grad_shape = DownsampleFactorMax.out_shape( grad_shape = DownsampleFactorMax.out_shape(imval.shape,
imval.shape, avgpoolsize, st=stridesize, avgpoolsize, st=stridesize,
ignore_border=True, padding=paddingsize) ignore_border=True, padding=paddingsize)
grad_val = rng.rand(*grad_shape) * 10.0 grad_val = rng.rand(*grad_shape) * 10.0
def mp(input, grad): def mp(input, grad):
grad_op = AveragePoolGrad(avgpoolsize, ignore_border=True, grad_op = AveragePoolGrad(avgpoolsize, ignore_border=True,
st=stridesize, padding=paddingsize, st=stridesize, padding=paddingsize,
mode=mode) mode=mode)
return grad_op(input, grad) return grad_op(input, grad)
utt.verify_grad(mp, [imval, grad_val], rng=rng) utt.verify_grad(mp, [imval, grad_val], rng=rng)
...@@ -637,7 +637,8 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -637,7 +637,8 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
# The value has been manually computed from the theoretical gradient, # The value has been manually computed from the theoretical gradient,
# and confirmed by the implementation. # and confirmed by the implementation.
assert numpy.allclose(fn_hess([1, 2]), [[0., 0.], [0., 982.7667]])
assert numpy.allclose(fn_hess( [1, 2]), [[0., 0.], [0., 982.7667]])
def test_max_pool_2d_2D(self): def test_max_pool_2d_2D(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = numpy.random.RandomState(utt.fetch_seed())
...@@ -683,9 +684,11 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -683,9 +684,11 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
op_output = function([input], op)(test_input_array) op_output = function([input], op)(test_input_array)
assert numpy.all(op_output == test_answer_array), ( assert numpy.all(op_output == test_answer_array), (
"op_output is %s, test_answer_array is %s" % ( "op_output is %s, test_answer_array is %s" % (
op_output, numpy_output_val
op_output, test_answer_array
) )
) )
def mp(input): def mp(input):
return max_pool_2d_same_size(input, patch_size) return max_pool_2d_same_size(input, patch_size)
utt.verify_grad(mp, [test_input_array], rng=rng) utt.verify_grad(mp, [test_input_array], rng=rng)
...@@ -712,14 +715,6 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -712,14 +715,6 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
assert numpy.all(output_val == numpy_output_val), ( assert numpy.all(output_val == numpy_output_val), (
"output_val is %s, numpy_output_val is %s" "output_val is %s, numpy_output_val is %s"
% (output_val, numpy_output_val)) % (output_val, numpy_output_val))
c = tensor.sum(output)
c_val = function([images], c)(imval)
g = tensor.grad(c, images)
g_val = function([images],
[g.shape,
tensor.min(g, axis=(0, 1, 2)),
tensor.max(g, axis=(0, 1, 2))]
)(imval)
# removed as already tested in test_max_pool_2d_2D # removed as already tested in test_max_pool_2d_2D
# This make test in debug mode too slow. # This make test in debug mode too slow.
...@@ -780,14 +775,14 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -780,14 +775,14 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
for i, maxpoolshp in enumerate(maxpoolshps): for i, maxpoolshp in enumerate(maxpoolshps):
for j, ignore_border in enumerate([True, False]): for j, ignore_border in enumerate([True, False]):
for k, padding in enumerate([(0,0), (1,1), (1,2)]): for k, padding in enumerate([(0, 0), (1, 1), (1, 2)]):
if out_shapes[k][i][j] is None: if out_shapes[k][i][j] is None:
continue continue
# checking shapes generated by DownsampleFactorMax # checking shapes generated by DownsampleFactorMax
self._compile_and_check([image], self._compile_and_check([image],
[DownsampleFactorMax(maxpoolshp, [DownsampleFactorMax(maxpoolshp,
ignore_border=ignore_border, ignore_border=ignore_border,
padding=padding)(image)], padding=padding)(image)],
[image_val], DownsampleFactorMax) [image_val], DownsampleFactorMax)
# checking shapes generated by MaxPoolGrad # checking shapes generated by MaxPoolGrad
...@@ -795,8 +790,8 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -795,8 +790,8 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
gz_val = rng.rand(*out_shapes[k][i][j]) gz_val = rng.rand(*out_shapes[k][i][j])
self._compile_and_check([image, maxout, gz], self._compile_and_check([image, maxout, gz],
[MaxPoolGrad(maxpoolshp, [MaxPoolGrad(maxpoolshp,
ignore_border=ignore_border, ignore_border=ignore_border,
padding=padding) padding=padding)
(image, maxout, gz)], (image, maxout, gz)],
[image_val, maxout_val, gz_val], [image_val, maxout_val, gz_val],
MaxPoolGrad, MaxPoolGrad,
...@@ -817,26 +812,26 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -817,26 +812,26 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
maxout = theano.tensor.tensor4() maxout = theano.tensor.tensor4()
grad = theano.tensor.tensor4() grad = theano.tensor.tensor4()
compilation_mode=theano.compile.get_default_mode().including( compilation_mode = theano.compile.get_default_mode().including(
'local_average_pool_grad') 'local_average_pool_grad')
for mode in ['max', 'sum', 'average_inc_pad', 'average_exc_pad']: for mode in ['max', 'sum', 'average_inc_pad', 'average_exc_pad']:
f = theano.function([im, maxout, grad], f = theano.function([im, maxout, grad],
DownsampleFactorMaxGrad(ds=(3,3), DownsampleFactorMaxGrad(ds=(3, 3),
ignore_border=False, ignore_border=False,
mode=mode)(im, maxout, grad), mode=mode)(im, maxout, grad),
mode=compilation_mode) mode=compilation_mode)
if mode == 'max': if mode == 'max':
assert any(isinstance(n.op, MaxPoolGrad) assert any(isinstance(n.op, MaxPoolGrad)
for n in f.maker.fgraph.toposort()) for n in f.maker.fgraph.toposort())
assert not any(isinstance(n.op, AveragePoolGrad) assert not any(isinstance(n.op, AveragePoolGrad)
for n in f.maker.fgraph.toposort()) for n in f.maker.fgraph.toposort())
else: else:
assert not any(isinstance(n.op, MaxPoolGrad) assert not any(isinstance(n.op, MaxPoolGrad)
for n in f.maker.fgraph.toposort()) for n in f.maker.fgraph.toposort())
assert any(isinstance(n.op, AveragePoolGrad) assert any(isinstance(n.op, AveragePoolGrad)
for n in f.maker.fgraph.toposort()) for n in f.maker.fgraph.toposort())
if __name__ == '__main__': if __name__ == '__main__':
unittest.main() unittest.main()
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论