Skip to content

P
pytensor
提交 54d5d47d authored 作者: Sina Honari's avatar Sina Honari
浏览文件
操作

issue #2196: adding functionality for the case when stride size is greater than the pooling size

上级 0dbb6196
显示空白字符变更
内嵌 并排
正在显示 包含 100 行增加 和 16 行删除
theano/tensor/signal/downsample.py
浏览文件 @ 54d5d47d
...@@ -68,7 +68,7 @@ class DownsampleFactorMax(Op): ...@@ -68,7 +68,7 @@ class DownsampleFactorMax(Op):
""" """
@staticmethod @staticmethod
def out_shape(imgshape, ds, st, ignore_border=False): def out_shape(imgshape, ds, ignore_border=False, st=None):
"""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.
...@@ -96,14 +96,39 @@ class DownsampleFactorMax(Op): ...@@ -96,14 +96,39 @@ class DownsampleFactorMax(Op):
if len(imgshape) < 2: if len(imgshape) < 2:
raise TypeError('imgshape must have at least two elements ' raise TypeError('imgshape must have at least two elements '
'(rows, cols)') '(rows, cols)')
if st == None:
st = ds
r, c = imgshape[-2:] r, c = imgshape[-2:]
rval = list(imgshape[:-2]) + [(r - ds[0]) // st[0] + 1, (c - ds[1]) // st[1] + 1] if st[0] >= ds[0]:
nr = r // st[0]
else:
nr = (r - ds[0]) // st[0] + 1
if st[1] >= ds[1]:
nc = c // st[1]
else:
nc = (c - ds[1]) // st[1] + 1
rval = list(imgshape[:-2]) + [nr, nc]
if not ignore_border: if not ignore_border:
if st[0] >= ds[0]:
if isinstance(r, theano.Variable):
rval[-2] = tensor.switch(r % st[0], rval[-2] + 1, rval[-2])
elif r % ds[0]:
rval[-2] += 1
else:
if isinstance(r, theano.Variable): if isinstance(r, theano.Variable):
rval[-2] = tensor.switch((r - ds[0]) % st[0], rval[-2] + 1, rval[-2]) rval[-2] = tensor.switch((r - ds[0]) % st[0], rval[-2] + 1, rval[-2])
elif (r - ds[0]) % st[0]: elif (r - ds[0]) % st[0]:
rval[-2] += 1 rval[-2] += 1
if st[1] >= ds[1]:
if isinstance(c, theano.Variable):
rval[-1] = tensor.switch(c % st[1], rval[-1] + 1, rval[-1])
elif c % ds[1]:
rval[-1] += 1
else:
if isinstance(c, theano.Variable): if isinstance(c, theano.Variable):
rval[-1] = tensor.switch((c - ds[1]) % st[1], rval[-1] + 1, rval[-1]) rval[-1] = tensor.switch((c - ds[1]) % st[1], rval[-1] + 1, rval[-1])
elif (c - ds[1]) % st[1]: elif (c - ds[1]) % st[1]:
...@@ -148,7 +173,7 @@ class DownsampleFactorMax(Op): ...@@ -148,7 +173,7 @@ class DownsampleFactorMax(Op):
return hash(type(self)) ^ hash(self.ds) ^ hash(self.st) ^ hash(self.ignore_border) return hash(type(self)) ^ hash(self.ds) ^ hash(self.st) ^ hash(self.ignore_border)
def __str__(self): def __str__(self):
return '%s{%s,%s}' % (self.__class__.__name__, return '%s{%s,%s,%s}' % (self.__class__.__name__,
self.ds, self.st, self.ignore_border) self.ds, self.st, self.ignore_border)
def make_node(self, x): def make_node(self, x):
...@@ -165,10 +190,10 @@ class DownsampleFactorMax(Op): ...@@ -165,10 +190,10 @@ class DownsampleFactorMax(Op):
if len(x.shape) != 4: if len(x.shape) != 4:
raise NotImplementedError( raise NotImplementedError(
'DownsampleFactorMax requires 4D input for now') 'DownsampleFactorMax requires 4D input for now')
z_shape = self.out_shape(x.shape, self.ds, self.st, self.ignore_border) z_shape = self.out_shape(x.shape, self.ds, self.ignore_border, self.st)
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(self.out_shape(x.shape, self.ds, self.st, z[0] = numpy.zeros(self.out_shape(x.shape, self.ds,
self.ignore_border)) self.ignore_border, self.st))
z[0] = theano._asarray(z[0], dtype=x.dtype) z[0] = theano._asarray(z[0], dtype=x.dtype)
zz = z[0] zz = z[0]
...@@ -182,32 +207,36 @@ class DownsampleFactorMax(Op): ...@@ -182,32 +207,36 @@ class DownsampleFactorMax(Op):
img_cols = x.shape[-1] img_cols = x.shape[-1]
if self.ignore_border: if self.ignore_border:
if st0 >= ds0:
x_usable2 = (x.shape[2] // ds0 * ds0)
else:
x_usable2 = (x.shape[2] - ds0) // st0 * st0 + ds0 x_usable2 = (x.shape[2] - ds0) // st0 * st0 + ds0
else: else:
x_usable2 = x.shape[2] x_usable2 = x.shape[2]
if self.ignore_border: if self.ignore_border:
if st1 >= ds1:
x_usable3 = (x.shape[3] // ds1 * ds1)
else:
x_usable3 = (x.shape[3] - ds1) // st1 * st1 + ds1 x_usable3 = (x.shape[3] - ds1) // st1 * st1 + ds1
else: else:
x_usable3 = x.shape[3] x_usable3 = x.shape[3]
for n in xrange(x.shape[0]): for n in xrange(x.shape[0]):
for k in xrange(x.shape[1]): for k in xrange(x.shape[1]):
for r in xrange(pr): for r in xrange(pr):
row_st = r * st0 row_st = r * st0
row_end = __builtin__.min(row_st + ds0, img_rows)
for c in xrange(pc): for c in xrange(pc):
col_st = c * st1 col_st = c * st1
for i in xrange(ds0): col_end = __builtin__.min(col_st + ds1, img_cols)
row_ind = row_st + i for row_ind in xrange(row_st, row_end):
if row_ind >= img_rows: for col_ind in xrange(col_st, col_end):
continue
for j in xrange(ds1):
col_ind = col_st + j
if col_ind >= img_cols:
continue
zz[n, k, r, c] = __builtin__.max(zz[n, k, r, c], zz[n, k, r, c] = __builtin__.max(zz[n, k, r, c],
x[n, k, row_ind, col_ind]) x[n, k, row_ind, col_ind])
def infer_shape(self, node, in_shapes): def infer_shape(self, node, in_shapes):
shp = self.out_shape(in_shapes[0], self.ds, self.st, self.ignore_border) shp = self.out_shape(in_shapes[0], self.ds, self.ignore_border, self.st)
return [shp] return [shp]
def grad(self, inp, grads): def grad(self, inp, grads):
...@@ -290,7 +319,7 @@ class DownsampleFactorMax(Op): ...@@ -290,7 +319,7 @@ class DownsampleFactorMax(Op):
} }
""" % locals() """ % locals()
def c_code_cache_version_tmp(self): def c_code_cache_version(self):
return (0, 1) return (0, 1)
......
theano/tensor/signal/tests/test_downsample.py
浏览文件 @ 54d5d47d
import unittest import unittest
import __builtin__
import numpy import numpy
import theano.tensor as tensor import theano.tensor as tensor
from theano.tests import unittest_tools as utt from theano.tests import unittest_tools as utt
...@@ -37,6 +38,60 @@ class TestDownsampleFactorMax(utt.InferShapeTester): ...@@ -37,6 +38,60 @@ class TestDownsampleFactorMax(utt.InferShapeTester):
output_val[k][i, j] = numpy.max(patch) output_val[k][i, j] = numpy.max(patch)
return output_val return output_val
@staticmethod
def numpy_max_pool_2d_stride(input, ds, ignore_border=False, st=None):
'''Helper function, implementing max_pool_2d in pure numpy
this function provides st input to indicate the stide size
for the pooling regions. if not indicated, st == sd.'''
if len(input.shape) < 2:
raise NotImplementedError('input should have at least 2 dim,'
' shape is %s'\
% str(input.shape))
if st == None:
st = ds
xi = 0
yi = 0
if not ignore_border:
if st[0] >= ds[0]:
if input.shape[-2] % st[0]:
xi += 1
else:
if (input.shape[-2] - ds[0]) % st[0]:
xi += 1
if st[1] >= ds[1]:
if input.shape[-1] % st[1]:
yi += 1
else:
if (input.shape[-1] % - ds[1]) % st[1]:
yi += 1
out_shp = list(input.shape[:-2])
if st[0] >= ds[0]:
out_shp.append(input.shape[-2] / ds[0] + xi)
else:
out_shp.append((input.shape[-2] - ds[0]) / st[0] + 1 + xi)
if st[1] >= ds[1]:
out_shp.append(input.shape[-1] / ds[1] + yi)
else:
out_shp.append((input.shape[-1] - ds[1]) / st[1] + 1 + yi)
output_val = numpy.zeros(out_shp)
img_rows = input.shape[-2]
img_cols = input.shape[-1]
for k in numpy.ndindex(*input.shape[:-2]):
for i in range(output_val.shape[-2]):
ii_st = i * ds[0]
ii_end = __builtin__.min(ii_st + ds[0], img_rows)
for j in range(output_val.shape[-1]):
jj_st = j * ds[1]
jj_end = __builtin__.min(jj_st + ds[1], img_cols)
patch = input[k][ii_st:ii_end, jj_st:jj_end]
output_val[k][i, j] = numpy.max(patch)
return output_val
def test_DownsampleFactorMax(self): def test_DownsampleFactorMax(self):
rng = numpy.random.RandomState(utt.fetch_seed()) rng = numpy.random.RandomState(utt.fetch_seed())
# generate random images # generate random images
......
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