def test_AveragePoolGrad_grad_stride(self, example, ignore_border, mode):
# checks the gradient of the gradient for
# the case that stride is used
rng = np.random.RandomState(utt.fetch_seed())
(avgpoolshp, stride, inputsize) = example
imval = rng.rand(*inputsize)
grad_shape = Pool.out_shape(
imval.shape,
avgpoolshp,
ndim=len(avgpoolshp),
ignore_border=ignore_border,
stride=stride,
)
# skip the grad verification when the output is empty
if np.prod(grad_shape) != 0:
grad_val = rng.rand(*grad_shape)
def mp(input, grad):
grad_op = AveragePoolGrad(ndim=len(avgpoolshp),
ignore_border=ignore_border,
mode=mode)
return grad_op(input, grad, avgpoolshp, stride)
utt.verify_grad(mp, [imval, grad_val], rng=rng)
def test_DownsampleFactorMaxPaddingStride_grad_grad(self):
rng = np.random.RandomState(utt.fetch_seed())
# maxpool, stride, pad, input sizes
examples = (
((3, ), (2, ), (2, ), (10, )),
(
(3, ),
(2, ),
(2, ),
(
2,
10,
),
),
(
(3, ),
(2, ),
(2, ),
(
2,
1,
10,
),
),
((5, 3), (3, 2), (2, 2), (1, 1, 10, 10)),
((3, 5), (2, 3), (2, 1), (1, 1, 10, 5)),
((5, 3, 3), (3, 2, 2), (2, 2, 2), (1, 1, 10, 5, 5)),
((3, 3, 5), (2, 2, 3), (2, 2, 1), (1, 1, 5, 5, 10)),
)
for (maxpoolshp, stridesize, padsize, inputsize) in examples:
imval = rng.rand(*inputsize) * 10.0
grad_shape = Pool.out_shape(
imval.shape,
maxpoolshp,
ndim=len(maxpoolshp),
stride=stridesize,
ignore_border=True,
pad=padsize,
)
grad_val = rng.rand(*grad_shape) * 10.0
def mp(input, grad):
out = Pool(
ndim=len(maxpoolshp),
ignore_border=True,
)(input, maxpoolshp, stridesize, padsize)
grad_op = MaxPoolGrad(ndim=len(maxpoolshp), ignore_border=True)
return grad_op(input, out, grad, maxpoolshp, stridesize,
padsize)
utt.verify_grad(mp, [imval, grad_val], rng=rng)
def test_AveragePoolPaddingStride_grad_grad(self):
rng = np.random.RandomState(utt.fetch_seed())
# avgpool, stride, pad, input sizes
examples = (
((3, ), (2, ), (2, ), (10, )),
(
(3, ),
(2, ),
(2, ),
(
2,
10,
),
),
(
(3, ),
(2, ),
(2, ),
(
2,
1,
10,
),
),
((5, 3), (3, 2), (2, 2), (1, 1, 10, 10)),
((3, 5), (2, 3), (2, 1), (1, 1, 10, 5)),
((5, 3, 2), (3, 2, 1), (2, 2, 2), (1, 1, 10, 5, 5)),
)
for (avgpoolshp, stridesize, padsize, inputsize) in examples:
imval = rng.rand(*inputsize) * 10.0
# 'average_exc_pad' with non-zero padding is not implemented
for mode in ["sum", "average_inc_pad"]:
grad_shape = Pool.out_shape(
imval.shape,
avgpoolshp,
ndim=len(avgpoolshp),
stride=stridesize,
ignore_border=True,
pad=padsize,
)
grad_val = rng.rand(*grad_shape) * 10.0
def mp(input, grad):
grad_op = AveragePoolGrad(ndim=len(avgpoolshp),
ignore_border=True,
mode=mode)
return grad_op(input, grad, avgpoolshp, stridesize,
padsize)
utt.verify_grad(mp, [imval, grad_val], rng=rng)
def test_DownsampleFactorMaxGrad_grad(self):
rng = np.random.RandomState(utt.fetch_seed())
# maxpool, input sizes
examples = (
((2, ), (2, )),
((2, ), (2, 3)),
((1, 1), (2, 3, 3, 4)),
((3, 2), (2, 3, 3, 4)),
((2, 3), (2, 3, 3, 4)),
((1, 1, 1), (2, 3, 3, 4)),
((3, 2, 2), (2, 3, 3, 4)),
((2, 3, 2), (2, 3, 3, 4)),
((2, 2, 3), (2, 3, 3, 4)),
((2, 2, 3), (2, 1, 3, 3, 4)),
)
for (maxpoolshp, inputsize) in examples:
imval = rng.rand(*inputsize) * 10.0
# more variance means numeric gradient will be more accurate
for ignore_border in [True, False]:
# print 'maxpoolshp =', maxpoolshp
# print 'ignore_border =', ignore_border
# The shape of the gradient will be the shape of the output
grad_shape = Pool.out_shape(
imval.shape,
maxpoolshp,
ndim=len(maxpoolshp),
ignore_border=ignore_border,
)
grad_val = rng.rand(*grad_shape) * 10.0
def mp(input, grad):
out = Pool(ndim=len(maxpoolshp),
ignore_border=ignore_border)(input, maxpoolshp)
grad_op = MaxPoolGrad(ndim=len(maxpoolshp),
ignore_border=ignore_border)
return grad_op(input, out, grad, maxpoolshp)
utt.verify_grad(mp, [imval, grad_val], rng=rng)
def test_AveragePoolGrad_grad(self):
rng = np.random.RandomState(utt.fetch_seed())
# avgpool, input sizes
examples = (
((2, ), (2, )),
((2, ), (2, 3)),
((1, 1), (2, 3, 3, 4)),
((3, 2), (2, 3, 3, 4)),
((2, 3), (2, 3, 3, 4)),
((3, 2, 2), (2, 3, 3, 4)),
((2, 2, 3), (2, 3, 3, 4)),
)
for (avgpoolshp, inputsize) in examples:
imval = rng.rand(*inputsize) * 10.0
# more variance means numeric gradient will be more accurate
for ignore_border in [True, False]:
for mode in ["sum", "average_inc_pad", "average_exc_pad"]:
# print 'maxpoolshp =', maxpoolshp
# print 'ignore_border =', ignore_border
# The shape of the gradient will be the shape of the output
grad_shape = Pool.out_shape(
imval.shape,
avgpoolshp,
ndim=len(avgpoolshp),
ignore_border=ignore_border,
)
grad_val = rng.rand(*grad_shape) * 10.0
def mp(input, grad):
grad_op = AveragePoolGrad(ndim=len(avgpoolshp),
ignore_border=ignore_border,
mode=mode)
return grad_op(input, grad, avgpoolshp)
utt.verify_grad(mp, [imval, grad_val], rng=rng)
def infer_shape(self, fgraph, node, in_shapes):
ws, stride, pad = [node.inputs[1], node.inputs[2], node.inputs[3]]
shp = Pool.out_shape(in_shapes[0], ws, self.ignore_border, stride, pad,
self.ndim)
return [shp]
def test_out_shape(self):
assert Pool.out_shape((9, 8, 6), (2, 2)) == [9, 4, 3]
assert Pool.out_shape((8, 6), (2, 2)) == [4, 3]
def test_DownsampleFactorMax(self):
rng = np.random.RandomState(utt.fetch_seed())
# maxpool, input size
examples = (
((2, ), (16, )),
(
(2, ),
(
4,
16,
),
),
(
(2, ),
(
4,
2,
16,
),
),
((1, 1), (4, 2, 16, 16)),
((2, 2), (4, 2, 16, 16)),
((3, 3), (4, 2, 16, 16)),
((3, 2), (4, 2, 16, 16)),
((3, 2, 2), (3, 2, 16, 16, 16)),
((2, 2, 3, 2), (3, 2, 6, 6, 6, 5)),
)
for example, ignore_border, mode in product(
examples,
[True, False],
["max", "sum", "average_inc_pad", "average_exc_pad"],
):
(maxpoolshp, inputsize) = example
imval = rng.rand(*inputsize)
images = aesara.shared(imval)
# Pure Numpy computation
numpy_output_val = self.numpy_max_pool_nd(imval,
maxpoolshp,
ignore_border,
mode=mode)
# The pool_2d or pool_3d helper methods
if len(maxpoolshp) == 2:
output = pool_2d(images, maxpoolshp, ignore_border, mode=mode)
f = function(
[],
[
output,
],
)
output_val = f()
utt.assert_allclose(output_val, numpy_output_val)
elif len(maxpoolshp) == 3:
output = pool_3d(images, maxpoolshp, ignore_border, mode=mode)
f = function(
[],
[
output,
],
)
output_val = f()
utt.assert_allclose(output_val, numpy_output_val)
# Pool op
maxpool_op = Pool(ndim=len(maxpoolshp),
ignore_border=ignore_border,
mode=mode)(images, maxpoolshp)
output_shape = Pool.out_shape(
imval.shape,
maxpoolshp,
ndim=len(maxpoolshp),
ignore_border=ignore_border,
)
utt.assert_allclose(np.asarray(output_shape),
numpy_output_val.shape)
f = function([], maxpool_op)
output_val = f()
utt.assert_allclose(output_val, numpy_output_val)