Skip to content

P
pytensor
提交 19c1624f authored 作者: Gijs van Tulder's avatar Gijs van Tulder
浏览文件
操作

Update gpuarray/cuda pooling tests.

Now use the CPU implementation as the reference (and other changes).
上级 8150bbd3
隐藏空白字符变更
内嵌 并排
正在显示 包含 234 行增加 和 350 行删除
theano/gpuarray/tests/test_dnn.py
浏览文件 @ 19c1624f
...@@ -10,7 +10,6 @@ import theano ...@@ -10,7 +10,6 @@ import theano
from six import StringIO from six import StringIO
import theano.tensor as T import theano.tensor as T
import theano.tests.unittest_tools as utt import theano.tests.unittest_tools as utt
from theano.sandbox.neighbours import images2neibs
from theano.tensor.signal.pool import pool_2d, pool_3d from theano.tensor.signal.pool import pool_2d, pool_3d
from theano.tensor.signal.pool import Pool, MaxPoolGrad, AveragePoolGrad from theano.tensor.signal.pool import Pool, MaxPoolGrad, AveragePoolGrad
...@@ -132,63 +131,6 @@ def test_dnn_conv_inplace(): ...@@ -132,63 +131,6 @@ def test_dnn_conv_inplace():
assert len([n for n in topo if isinstance(n.op, GpuAllocEmpty)]) == 2 assert len([n for n in topo if isinstance(n.op, GpuAllocEmpty)]) == 2
def pool_2d_i2n(input, ds=(2, 2), strides=None,
pad=(0, 0),
pool_function=T.max, mode='ignore_borders'):
if strides is None:
strides = ds
if strides[0] > ds[0] or strides[1] > ds[1]:
raise RuntimeError(
"strides should be smaller than or equal to ds,"
" strides=(%d, %d) and ds=(%d, %d)" %
(strides + ds))
shape = input.shape
if pad != (0, 0):
assert pool_function is T.max
pad_x = pad[0]
pad_y = pad[1]
a = T.alloc(-numpy.inf, shape[0], shape[1], shape[2] + pad_x * 2,
shape[3] + pad_y * 2)
input = T.set_subtensor(a[:, :,
pad_x:pad_x + shape[2],
pad_y:pad_y + shape[3]],
input)
shape = input.shape
neibs = images2neibs(input, ds, strides, mode=mode)
pooled_neibs = pool_function(neibs, axis=1)
output_width = (shape[2] - ds[0]) // strides[0] + 1
output_height = (shape[3] - ds[1]) // strides[1] + 1
pooled_output = pooled_neibs.reshape((shape[0], shape[1],
output_width, output_height))
return pooled_output
def pool3d2d(input, ds=(2, 2, 2), strides=None, pad=(0, 0, 0),
pool_function=T.max, mode='ignore_borders'):
if strides is None:
strides = ds
assert input.ndim == 5
shape = input.shape
# reshape to B, C*0, 1, 2 and do the pooling on 1, 2
first = input.reshape((shape[0], shape[1] * shape[2], shape[3], shape[4]))
pooled1 = pool_2d_i2n(first, ds=ds[1:], strides=strides[1:], pad=pad[1:],
pool_function=pool_function, mode=mode)
shp1 = pooled1.shape
# reshape to B, C, 0, 1*2 and do the pooling on 0
second = pooled1.reshape((shape[0], shape[1], shape[2], shp1[2] * shp1[3]))
pooled2 = pool_2d_i2n(second, ds=(ds[0], 1), strides=(strides[0], 1),
pad=(pad[0], 0), pool_function=pool_function, mode=mode)
shp2 = pooled2.shape
return pooled2.reshape((shape[0], shape[1], shp2[2], shp1[2], shp1[3]))
def test_pooling(): def test_pooling():
if not dnn.dnn_available(test_ctx_name): if not dnn.dnn_available(test_ctx_name):
raise SkipTest(dnn.dnn_available.msg) raise SkipTest(dnn.dnn_available.msg)
...@@ -202,12 +144,8 @@ def test_pooling(): ...@@ -202,12 +144,8 @@ def test_pooling():
x = T.ftensor4() x = T.ftensor4()
for mode, pad in product(modes, for mode, pad in product(modes,
((0, 0), (1, 0), (0, 1), (2, 3), (3, 2))): ((0, 0), (1, 0), (0, 1), (2, 3), (3, 2))):
if mode == 'max': if pad != (0, 0) and mode == 'average_exc_pad':
func = T.max # Not implemented
else:
func = T.mean
if pad != (0, 0) and func is T.mean:
continue continue
for ws in (4, 2, 5): for ws in (4, 2, 5):
...@@ -218,29 +156,32 @@ def test_pooling(): ...@@ -218,29 +156,32 @@ def test_pooling():
# Not implemented # Not implemented
continue continue
# We will check that the opt introduced it. # We will check that the opt introduced it.
out1 = pool_2d(x, (ws, ws), out = pool_2d(x, (ws, ws),
st=(stride, stride), st=(stride, stride),
ignore_border=True, ignore_border=True,
padding=pad, mode=mode) padding=pad, mode=mode)
out2 = pool_2d_i2n(x, ds=(ws, ws), strides=(stride, stride),
pad=pad,
pool_function=func)
mode_without_gpu2 = mode_without_gpu.including() mode_without_gpu2 = mode_without_gpu.including()
mode_without_gpu2.check_isfinite = False mode_without_gpu2.check_isfinite = False
f1 = theano.function([x], out1, mode=mode_with_gpu)
# GPU implementation
f_gpu = theano.function([x], out, mode=mode_with_gpu)
assert any([isinstance(node.op, dnn.GpuDnnPool) assert any([isinstance(node.op, dnn.GpuDnnPool)
for node in f1.maker.fgraph.apply_nodes]) for node in f_gpu.maker.fgraph.apply_nodes])
f2 = theano.function([x], out2, mode=mode_without_gpu2)
# CPU implementation
f_cpu = theano.function([x], out, mode=mode_without_gpu2)
assert not any([isinstance(node.op, dnn.GpuDnnPool) assert not any([isinstance(node.op, dnn.GpuDnnPool)
for node in f2.maker.fgraph.apply_nodes]) for node in f_cpu.maker.fgraph.apply_nodes])
assert any([isinstance(node.op, Pool)
for node in f_cpu.maker.fgraph.apply_nodes])
for shp in [(1, 10, 100, 100), for shp in [(1, 10, 100, 100),
(1, 3, 99, 99), (1, 3, 99, 99),
(32, 1, 147, 197), (32, 1, 147, 197),
]: ]:
data = numpy.random.normal(0, 1, shp).astype("float32") data = numpy.random.normal(0, 1, shp).astype("float32")
a = f1(data) a = f_cpu(data).__array__()
b = f2(data) b = f_gpu(data).__array__()
utt.assert_allclose(a, b) utt.assert_allclose(a, b)
# Test the grad # Test the grad
...@@ -254,7 +195,7 @@ def test_pooling(): ...@@ -254,7 +195,7 @@ def test_pooling():
# Not implemented # Not implemented
continue continue
# This test the CPU grad + opt + GPU implemtentation # This tests the CPU grad + opt + GPU implementation
def fn(x): def fn(x):
return pool_2d(x, (ws, ws), ignore_border=True, return pool_2d(x, (ws, ws), ignore_border=True,
padding=pad, mode=mode) padding=pad, mode=mode)
...@@ -279,27 +220,69 @@ def test_pooling(): ...@@ -279,27 +220,69 @@ def test_pooling():
mode=mode_with_gpu) mode=mode_with_gpu)
assert any([isinstance(node.op, dnn.GpuDnnPoolGrad) assert any([isinstance(node.op, dnn.GpuDnnPoolGrad)
for node in fg.maker.fgraph.toposort()]) for node in fg.maker.fgraph.toposort()])
g_out = fg(data)
# Compare against the CPU result def test_pooling_with_tensor_vars():
out = pool_2d(x, (ws, ws),
padding=pad,
ignore_border=True, mode=mode)
fc = theano.function([x], theano.grad(out.sum(), x),
mode=mode_without_gpu)
if mode == 'max':
assert any([isinstance(node.op, MaxPoolGrad)
for node in fc.maker.fgraph.toposort()])
else:
assert any([isinstance(node.op, AveragePoolGrad)
for node in fc.maker.fgraph.toposort()])
c_out = fc(data)
utt.assert_allclose(c_out, g_out)
def test_pooling_3d():
if not dnn.dnn_available(test_ctx_name): if not dnn.dnn_available(test_ctx_name):
raise SkipTest(dnn.dnn_available.msg) raise SkipTest(dnn.dnn_available.msg)
x = T.ftensor4()
ws = theano.shared(numpy.array([2, 2], dtype='int32'))
st = theano.shared(numpy.array([1, 1], dtype='int32'))
pad = theano.shared(numpy.array([0, 0], dtype='int32'))
mode = 'max'
def fn(x):
dnn_op = dnn.dnn_pool(
x, ws=ws,
stride=st,
pad=pad,
mode=mode)
return dnn_op
for shp in [(1, 1, 2, 2),
(1, 1, 3, 3)]:
data = numpy.random.normal(0, 1, shp).astype("float32") * 10
theano.tests.unittest_tools.verify_grad(
fn, [data], mode=mode_with_gpu)
mode_without_gpu2 = mode_without_gpu.including()
mode_without_gpu2.check_isfinite = False
# GPU implementation
f_gpu = theano.function([x], fn(x), mode=mode_with_gpu)
assert any([isinstance(node.op, dnn.GpuDnnPool)
for node in f_gpu.maker.fgraph.apply_nodes])
# CPU implementation
out_cpu = pool_2d(x, ws, ignore_border=True, st=st, padding=pad, mode=mode)
f_cpu = theano.function([x], out_cpu, mode=mode_without_gpu2)
assert not any([isinstance(node.op, dnn.GpuDnnPool)
for node in f_cpu.maker.fgraph.apply_nodes])
assert any([isinstance(node.op, Pool)
for node in f_cpu.maker.fgraph.apply_nodes])
i = 1
for shp in [(1, 10, 100, 100),
(1, 3, 99, 99),
(32, 1, 147, 197)]:
data = numpy.random.normal(0, 1, shp).astype("float32")
# Change the window size dynamically
ws.set_value(numpy.array([i, i]).astype('int32'))
a = f_gpu(data).__array__()
b = f_cpu(data).__array__()
utt.assert_allclose(a, b)
i += 1
def test_pooling3d():
# 3d pooling requires version 3 or newer.
if not dnn.dnn_available(test_ctx_name) or dnn.version(raises=False) < 3000:
raise SkipTest(dnn.dnn_available.msg)
# We force the FAST_RUN as we don't want the reference to run in DebugMode.
mode_without_gpu_ref = theano.compile.mode.get_mode(
'FAST_RUN').excluding('gpuarray')
# 'average_exc_pad' is disabled for versions < 4004 # 'average_exc_pad' is disabled for versions < 4004
if dnn.version(raises=False) < 4004: if dnn.version(raises=False) < 4004:
...@@ -310,13 +293,9 @@ def test_pooling_3d(): ...@@ -310,13 +293,9 @@ def test_pooling_3d():
x = T.ftensor5() x = T.ftensor5()
for mode, pad in product(modes, for mode, pad in product(modes,
((0, 0, 0), (1, 0, 0), (0, 1, 0), (0, 0, 1), ((0, 0, 0), (1, 0, 0), (0, 1, 0), (0, 0, 1),
(3, 2, 2), (2, 3, 2), (2, 2, 3))): (2, 3, 2), (3, 2, 2), (2, 2, 3))):
if mode == 'max': if pad != (0, 0, 0) and mode == 'average_exc_pad':
func = T.max # Not implemented
else:
func = T.mean
if pad != (0, 0, 0) and func is T.mean:
continue continue
for ws in (4, 2, 5): for ws in (4, 2, 5):
...@@ -326,35 +305,41 @@ def test_pooling_3d(): ...@@ -326,35 +305,41 @@ def test_pooling_3d():
if pad[0] > stride or pad[1] > stride or pad[2] > stride: if pad[0] > stride or pad[1] > stride or pad[2] > stride:
# Not implemented # Not implemented
continue continue
# We will check that the opt introduced it. out = pool_3d(x, (ws, ws, ws),
out1 = pool_3d(x, (ws, ws, ws), st=(stride, stride, stride),
st=(stride, stride, stride), ignore_border=True,
ignore_border=True, padding=pad, mode=mode)
padding=pad, mode=mode)
out2 = pool3d2d(x, ds=(ws, ws, ws), strides=(stride, stride, stride), # GPU implementation
pad=pad, f_gpu = theano.function([x], out, mode=mode_with_gpu)
pool_function=func)
mode_without_gpu2 = mode_without_gpu.including()
mode_without_gpu2.check_isfinite = False
f1 = theano.function([x], out1, mode=mode_with_gpu)
assert any([isinstance(node.op, dnn.GpuDnnPool) assert any([isinstance(node.op, dnn.GpuDnnPool)
for node in f1.maker.fgraph.apply_nodes]) for node in f_gpu.maker.fgraph.apply_nodes])
f2 = theano.function([x], out2, mode=mode_without_gpu2)
# CPU implementation
f_cpu = theano.function([x], out, mode=mode_without_gpu_ref)
assert not any([isinstance(node.op, dnn.GpuDnnPool) assert not any([isinstance(node.op, dnn.GpuDnnPool)
for node in f2.maker.fgraph.apply_nodes]) for node in f_cpu.maker.fgraph.apply_nodes])
for shp in [(1, 10, 30, 30, 30), assert any([isinstance(node.op, Pool)
(1, 3, 29, 29, 29), for node in f_cpu.maker.fgraph.apply_nodes])
(3, 1, 47, 97, 47),
for shp in [(1, 5, 50, 20, 50),
(1, 3, 99, 99, 29),
(2, 1, 147, 97, 37),
]: ]:
data = numpy.random.normal(0, 1, shp).astype("float32") data = numpy.random.normal(0, 1, shp).astype("float32")
a = f1(data) a = f_cpu(data).__array__()
b = f2(data) b = f_gpu(data).__array__()
utt.assert_allclose(a, b,
utt.assert_allclose(a, b) atol=numpy.finfo(numpy.float32).eps)
# Test the grad # Test the grad
for shp in [(1, 1, 2, 2, 2), for shp in [(1, 1, 2, 2, 2),
(1, 1, 3, 3, 3)]: (1, 1, 3, 3, 3),
(1, 1, 3, 3, 4),
(1, 1, 3, 4, 3),
(1, 1, 4, 3, 3),
(1, 1, 4, 4, 4),
(1, 1, 5, 5, 5)]:
data = numpy.random.normal(0, 1, shp).astype("float32") * 10 data = numpy.random.normal(0, 1, shp).astype("float32") * 10
ws = 2 ws = 2
...@@ -363,17 +348,6 @@ def test_pooling_3d(): ...@@ -363,17 +348,6 @@ def test_pooling_3d():
# Not implemented # Not implemented
continue continue
# This test the CPU grad + opt + GPU implemtentation
def fn(x):
return pool_3d(x, (ws, ws, ws), ignore_border=True,
padding=pad, mode=mode)
utt.verify_grad(fn, [data], mode=mode_with_gpu)
# Confirm that the opt would have inserted it.
fg = theano.function([x], theano.grad(fn(x).sum(), x),
mode=mode_with_gpu)
assert any([isinstance(node.op, dnn.GpuDnnPoolGrad)
for node in fg.maker.fgraph.toposort()])
# Test the GPU grad + GPU implementation # Test the GPU grad + GPU implementation
def fn(x): def fn(x):
dnn_op = dnn.dnn_pool( dnn_op = dnn.dnn_pool(
...@@ -388,69 +362,6 @@ def test_pooling_3d(): ...@@ -388,69 +362,6 @@ def test_pooling_3d():
mode=mode_with_gpu) mode=mode_with_gpu)
assert any([isinstance(node.op, dnn.GpuDnnPoolGrad) assert any([isinstance(node.op, dnn.GpuDnnPoolGrad)
for node in fg.maker.fgraph.toposort()]) for node in fg.maker.fgraph.toposort()])
g_out = fg(data)
# Compare against the CPU result
out = pool_3d(x, (ws, ws, ws),
padding=pad,
ignore_border=True, mode=mode)
fc = theano.function([x], theano.grad(out.sum(), x),
mode=mode_without_gpu)
if mode == 'max':
assert any([isinstance(node.op, MaxPoolGrad)
for node in fc.maker.fgraph.toposort()])
else:
assert any([isinstance(node.op, AveragePoolGrad)
for node in fc.maker.fgraph.toposort()])
c_out = fc(data)
utt.assert_allclose(c_out, g_out)
def test_pooling_with_tensor_vars():
if not dnn.dnn_available(test_ctx_name):
raise SkipTest(dnn.dnn_available.msg)
x = T.ftensor4()
ws = theano.shared(numpy.array([2, 2], dtype='int32'))
st = theano.shared(numpy.array([1, 1], dtype='int32'))
pad = theano.shared(numpy.array([0, 0], dtype='int32'))
mode = 'max'
def fn(x):
dnn_op = dnn.dnn_pool(x,
ws=ws,
stride=st,
pad=pad,
mode=mode)
return dnn_op
for shp in [(1, 1, 2, 2),
(1, 1, 3, 3)]:
data = numpy.random.normal(0, 1, shp).astype("float32") * 10
theano.tests.unittest_tools.verify_grad(
fn, [data], mode=mode_with_gpu)
out2 = pool_2d_i2n(x, ds=(2, 2), strides=(1, 1),
pad=(0, 0),
pool_function=T.max)
mode_without_gpu2 = mode_without_gpu.including()
mode_without_gpu2.check_isfinite = False
f1 = theano.function([x], fn(x), mode=mode_with_gpu)
assert any([isinstance(node.op, dnn.GpuDnnPool)
for node in f1.maker.fgraph.apply_nodes])
f2 = theano.function([x], out2, mode=mode_without_gpu2)
assert not any([isinstance(node.op, dnn.GpuDnnPool)
for node in f2.maker.fgraph.apply_nodes])
for shp in [(1, 10, 100, 100),
(1, 3, 99, 99),
(32, 1, 147, 197),
]:
data = numpy.random.normal(0, 1, shp).astype("float32")
a = f1(data).__array__()
b = f2(data).__array__()
utt.assert_allclose(a, b)
def test_pooling_opt(): def test_pooling_opt():
......
theano/sandbox/cuda/tests/test_dnn.py
浏览文件 @ 19c1624f
...@@ -14,7 +14,6 @@ import numpy ...@@ -14,7 +14,6 @@ import numpy
import theano import theano
import theano.tensor as T import theano.tensor as T
import theano.tests.unittest_tools as utt import theano.tests.unittest_tools as utt
from theano.sandbox.neighbours import images2neibs
from theano.tensor.signal.pool import pool_2d, pool_3d from theano.tensor.signal.pool import pool_2d, pool_3d
from theano.tensor.signal.pool import Pool, MaxPoolGrad, AveragePoolGrad from theano.tensor.signal.pool import Pool, MaxPoolGrad, AveragePoolGrad
import theano.sandbox.cuda.dnn as dnn import theano.sandbox.cuda.dnn as dnn
...@@ -169,62 +168,6 @@ def test_dnn_conv_inplace(): ...@@ -169,62 +168,6 @@ def test_dnn_conv_inplace():
assert len([n for n in topo if isinstance(n.op, GpuAllocEmpty)]) == 2 assert len([n for n in topo if isinstance(n.op, GpuAllocEmpty)]) == 2
def pool3d2d(input, ds=(2, 2, 2), strides=None, pad=(0, 0, 0),
pool_function=T.max, mode='ignore_borders'):
if strides is None:
strides = ds
assert input.ndim == 5
shape = input.shape
# resahpe to B, C*0, 1, 2 and do the pooling on 1, 2
first = input.reshape((shape[0], shape[1] * shape[2], shape[3], shape[4]))
pooled1 = pool_2d_i2n(first, ds=ds[1:], strides=strides[1:], pad=pad[1:],
pool_function=pool_function, mode=mode)
shp1 = pooled1.shape
# reshape to B, C, 0, 1*2 and do the pooling on 0
second = pooled1.reshape((shape[0], shape[1], shape[2], shp1[2] * shp1[3]))
pooled2 = pool_2d_i2n(second, ds=(ds[0], 1), strides=(strides[0], 1),
pad=(pad[0], 0), pool_function=pool_function, mode=mode)
shp2 = pooled2.shape
return pooled2.reshape((shape[0], shape[1], shp2[2], shp1[2], shp1[3]))
def pool_2d_i2n(input, ds=(2, 2), strides=None, pad=(0, 0),
pool_function=T.max, mode='ignore_borders'):
if strides is None:
strides = ds
if strides[0] > ds[0] or strides[1] > ds[1]:
raise RuntimeError(
"strides should be smaller than or equal to ds,"
" strides=(%d, %d) and ds=(%d, %d)" %
(strides + ds))
shape = input.shape
if pad != (0, 0):
assert pool_function is T.max
pad_x = pad[0]
pad_y = pad[1]
a = T.alloc(-numpy.inf, shape[0], shape[1], shape[2] + pad_x * 2,
shape[3] + pad_y * 2)
input = T.set_subtensor(a[:, :,
pad_x:pad_x + shape[2],
pad_y:pad_y + shape[3]],
input)
shape = input.shape
neibs = images2neibs(input, ds, strides, mode=mode)
pooled_neibs = pool_function(neibs, axis=1)
output_width = (shape[2] - ds[0]) // strides[0] + 1
output_height = (shape[3] - ds[1]) // strides[1] + 1
pooled_output = pooled_neibs.reshape((shape[0], shape[1],
output_width, output_height))
return pooled_output
def test_pooling(): def test_pooling():
if not cuda.dnn.dnn_available(): if not cuda.dnn.dnn_available():
raise SkipTest(cuda.dnn.dnn_available.msg) raise SkipTest(cuda.dnn.dnn_available.msg)
...@@ -237,13 +180,9 @@ def test_pooling(): ...@@ -237,13 +180,9 @@ def test_pooling():
x = T.ftensor4() x = T.ftensor4()
for mode, pad in product(modes, for mode, pad in product(modes,
((0, 0), (1, 0), (1, 0), (2, 3), (3, 2))): ((0, 0), (1, 0), (0, 1), (2, 3), (3, 2))):
if mode == 'max': if pad != (0, 0) and mode == 'average_exc_pad':
func = T.max # Not implemented
else:
func = T.mean
if pad != (0, 0) and func is T.mean:
continue continue
for ws in (4, 2, 5): for ws in (4, 2, 5):
...@@ -254,29 +193,32 @@ def test_pooling(): ...@@ -254,29 +193,32 @@ def test_pooling():
# Not implemented # Not implemented
continue continue
# We will check that the opt introduced it. # We will check that the opt introduced it.
out1 = pool_2d(x, (ws, ws), out = pool_2d(x, (ws, ws),
st=(stride, stride), st=(stride, stride),
ignore_border=True, ignore_border=True,
padding=pad, mode=mode) padding=pad, mode=mode)
out2 = pool_2d_i2n(x, ds=(ws, ws), strides=(stride, stride),
pad=pad,
pool_function=func)
mode_without_gpu2 = mode_without_gpu.including() mode_without_gpu2 = mode_without_gpu.including()
mode_without_gpu2.check_isfinite = False mode_without_gpu2.check_isfinite = False
f1 = theano.function([x], out1, mode=mode_with_gpu)
# GPU implementation
f_gpu = theano.function([x], out, mode=mode_with_gpu)
assert any([isinstance(node.op, cuda.dnn.GpuDnnPool) assert any([isinstance(node.op, cuda.dnn.GpuDnnPool)
for node in f1.maker.fgraph.apply_nodes]) for node in f_gpu.maker.fgraph.apply_nodes])
f2 = theano.function([x], out2, mode=mode_without_gpu2)
# CPU implementation
f_cpu = theano.function([x], out, mode=mode_without_gpu2)
assert not any([isinstance(node.op, cuda.dnn.GpuDnnPool) assert not any([isinstance(node.op, cuda.dnn.GpuDnnPool)
for node in f2.maker.fgraph.apply_nodes]) for node in f_cpu.maker.fgraph.apply_nodes])
assert any([isinstance(node.op, Pool)
for node in f_cpu.maker.fgraph.apply_nodes])
for shp in [(1, 10, 100, 100), for shp in [(1, 10, 100, 100),
(1, 3, 99, 99), (1, 3, 99, 99),
(32, 1, 147, 197), (32, 1, 147, 197),
]: ]:
data = numpy.random.normal(0, 1, shp).astype("float32") data = numpy.random.normal(0, 1, shp).astype("float32")
a = f1(data).__array__() a = f_cpu(data).__array__()
b = f_gpu(data).__array__()
b = f2(data).__array__()
utt.assert_allclose(a, b) utt.assert_allclose(a, b)
# Test the grad # Test the grad
...@@ -290,12 +232,11 @@ def test_pooling(): ...@@ -290,12 +232,11 @@ def test_pooling():
# Not implemented # Not implemented
continue continue
# This test the CPU grad + opt + GPU implemtentation # This tests the CPU grad + opt + GPU implementation
def fn(x): def fn(x):
return pool_2d(x, (ws, ws), ignore_border=True, return pool_2d(x, (ws, ws), ignore_border=True,
padding=pad, mode=mode) padding=pad, mode=mode)
theano.tests.unittest_tools.verify_grad(fn, [data], utt.verify_grad(fn, [data], mode=mode_with_gpu)
mode=mode_with_gpu)
# Confirm that the opt would have inserted it. # Confirm that the opt would have inserted it.
fg = theano.function([x], theano.grad(fn(x).sum(), x), fg = theano.function([x], theano.grad(fn(x).sum(), x),
mode=mode_with_gpu) mode=mode_with_gpu)
...@@ -310,29 +251,12 @@ def test_pooling(): ...@@ -310,29 +251,12 @@ def test_pooling():
pad=pad, pad=pad,
mode=mode) mode=mode)
return dnn_op return dnn_op
theano.tests.unittest_tools.verify_grad( utt.verify_grad(fn, [data], mode=mode_with_gpu)
fn, [data], mode=mode_with_gpu)
# Confirm that we get the good op. # Confirm that we get the good op.
fg = theano.function([x], theano.grad(fn(x).sum(), x), fg = theano.function([x], theano.grad(fn(x).sum(), x),
mode=mode_with_gpu) mode=mode_with_gpu)
assert any([isinstance(node.op, cuda.dnn.GpuDnnPoolGrad) assert any([isinstance(node.op, cuda.dnn.GpuDnnPoolGrad)
for node in fg.maker.fgraph.toposort()]) for node in fg.maker.fgraph.toposort()])
g_out = fg(data)
# Compare again the CPU result
out = pool_2d(x, (ws, ws),
padding=pad,
ignore_border=True, mode=mode)
fc = theano.function([x], theano.grad(out.sum(), x),
mode=mode_without_gpu)
if mode == 'max':
assert any([isinstance(node.op, MaxPoolGrad)
for node in fc.maker.fgraph.toposort()])
else:
assert any([isinstance(node.op, AveragePoolGrad)
for node in fc.maker.fgraph.toposort()])
c_out = fc(data)
utt.assert_allclose(c_out, g_out)
def test_pooling_with_tensor_vars(): def test_pooling_with_tensor_vars():
...@@ -361,24 +285,26 @@ def test_pooling_with_tensor_vars(): ...@@ -361,24 +285,26 @@ def test_pooling_with_tensor_vars():
mode_without_gpu2 = mode_without_gpu.including() mode_without_gpu2 = mode_without_gpu.including()
mode_without_gpu2.check_isfinite = False mode_without_gpu2.check_isfinite = False
# GPU implementation
f_gpu = theano.function([x], fn(x), mode=mode_with_gpu) f_gpu = theano.function([x], fn(x), mode=mode_with_gpu)
assert any([isinstance(node.op, cuda.dnn.GpuDnnPool) assert any([isinstance(node.op, cuda.dnn.GpuDnnPool)
for node in f_gpu.maker.fgraph.apply_nodes]) for node in f_gpu.maker.fgraph.apply_nodes])
# CPU implementation
out_cpu = pool_2d(x, ws, ignore_border=True, st=st, padding=pad, mode=mode)
f_cpu = theano.function([x], out_cpu, mode=mode_without_gpu2)
assert not any([isinstance(node.op, cuda.dnn.GpuDnnPool)
for node in f_cpu.maker.fgraph.apply_nodes])
assert any([isinstance(node.op, Pool)
for node in f_cpu.maker.fgraph.apply_nodes])
i = 1 i = 1
for shp in [(1, 10, 100, 100), for shp in [(1, 10, 100, 100),
(1, 3, 99, 99), (1, 3, 99, 99),
(32, 1, 147, 197)]: (32, 1, 147, 197)]:
data = numpy.random.normal(0, 1, shp).astype("float32") data = numpy.random.normal(0, 1, shp).astype("float32")
out = pool_2d_i2n(x, ds=(i, i), strides=(1, 1),
pad=(0, 0),
pool_function=T.max)
f_cpu = theano.function([x], out, mode=mode_without_gpu2)
assert not any([isinstance(node.op, cuda.dnn.GpuDnnPool)
for node in f_cpu.maker.fgraph.apply_nodes])
# Change the window size dynamically for gpu op # Change the window size dynamically
ws.set_value(numpy.array([i, i]).astype('int32')) ws.set_value(numpy.array([i, i]).astype('int32'))
a = f_gpu(data).__array__() a = f_gpu(data).__array__()
b = f_cpu(data).__array__() b = f_cpu(data).__array__()
...@@ -426,12 +352,8 @@ def test_pooling3d(): ...@@ -426,12 +352,8 @@ def test_pooling3d():
for mode, pad in product(modes, for mode, pad in product(modes,
((0, 0, 0), (1, 0, 0), (0, 1, 0), (0, 0, 1), ((0, 0, 0), (1, 0, 0), (0, 1, 0), (0, 0, 1),
(2, 3, 2), (3, 2, 2), (2, 2, 3))): (2, 3, 2), (3, 2, 2), (2, 2, 3))):
if mode == 'max': if pad != (0, 0, 0) and mode == 'average_exc_pad':
func = T.max # Not implemented
else:
func = T.mean
if pad != (0, 0, 0) and func is T.mean:
continue continue
for ws in (4, 2, 5): for ws in (4, 2, 5):
...@@ -441,28 +363,30 @@ def test_pooling3d(): ...@@ -441,28 +363,30 @@ def test_pooling3d():
if pad[0] > stride or pad[1] > stride or pad[2] > stride: if pad[0] > stride or pad[1] > stride or pad[2] > stride:
# Not implemented # Not implemented
continue continue
out1 = pool_3d(x, (ws, ws, ws), out = pool_3d(x, (ws, ws, ws),
st=(stride, stride, stride), st=(stride, stride, stride),
ignore_border=True, ignore_border=True,
padding=pad, mode=mode) padding=pad, mode=mode)
out2 = pool3d2d(x, ds=(ws, ws, ws), strides=(stride, stride, stride),
pad=pad, # GPU implementation
pool_function=func) f_gpu = theano.function([x], out, mode=mode_with_gpu)
f1 = theano.function([x], out1, mode=mode_with_gpu)
assert any([isinstance(node.op, cuda.dnn.GpuDnnPool) assert any([isinstance(node.op, cuda.dnn.GpuDnnPool)
for node in f1.maker.fgraph.apply_nodes]) for node in f_gpu.maker.fgraph.apply_nodes])
f2 = theano.function([x], out2, mode=mode_without_gpu_ref)
# CPU implementation
f_cpu = theano.function([x], out, mode=mode_without_gpu_ref)
assert not any([isinstance(node.op, cuda.dnn.GpuDnnPool) assert not any([isinstance(node.op, cuda.dnn.GpuDnnPool)
for node in f2.maker.fgraph.apply_nodes]) for node in f_cpu.maker.fgraph.apply_nodes])
for shp in [(1, 10, 100, 100, 100), assert any([isinstance(node.op, Pool)
(1, 3, 99, 99, 99), for node in f_cpu.maker.fgraph.apply_nodes])
(32, 1, 147, 197, 37),
for shp in [(1, 5, 50, 20, 50),
(1, 3, 99, 99, 29),
(2, 1, 147, 97, 37),
]: ]:
data = numpy.random.normal(0, 1, shp).astype("float32") data = numpy.random.normal(0, 1, shp).astype("float32")
a = f1(data).__array__() a = f_cpu(data).__array__()
b = f_gpu(data).__array__()
b = f2(data).__array__()
utt.assert_allclose(a, b, utt.assert_allclose(a, b,
atol=numpy.finfo(numpy.float32).eps) atol=numpy.finfo(numpy.float32).eps)
...@@ -490,29 +414,12 @@ def test_pooling3d(): ...@@ -490,29 +414,12 @@ def test_pooling3d():
pad=pad, pad=pad,
mode=mode) mode=mode)
return dnn_op return dnn_op
theano.tests.unittest_tools.verify_grad( utt.verify_grad(fn, [data], mode=mode_with_gpu)
fn, [data], mode=mode_with_gpu)
# Confirm that we get the good op. # Confirm that we get the good op.
fg = theano.function([x], theano.grad(fn(x).sum(), x), fg = theano.function([x], theano.grad(fn(x).sum(), x),
mode=mode_with_gpu) mode=mode_with_gpu)
assert any([isinstance(node.op, cuda.dnn.GpuDnnPoolGrad) assert any([isinstance(node.op, cuda.dnn.GpuDnnPoolGrad)
for node in fg.maker.fgraph.toposort()]) for node in fg.maker.fgraph.toposort()])
g_out = fg(data)
# Compare again the CPU result
out = pool_3d(x, (ws, ws, ws),
padding=pad,
ignore_border=True, mode=mode)
fc = theano.function([x], theano.grad(out.sum(), x),
mode=mode_without_gpu_ref)
if mode == 'max':
assert any([isinstance(node.op, MaxPoolGrad)
for node in fc.maker.fgraph.toposort()])
else:
assert any([isinstance(node.op, AveragePoolGrad)
for node in fc.maker.fgraph.toposort()])
c_out = fc(data)
utt.assert_allclose(c_out, g_out)
def test_pooling_opt(): def test_pooling_opt():
...@@ -1269,6 +1176,33 @@ class TestDnnInferShapes(utt.InferShapeTester): ...@@ -1269,6 +1176,33 @@ class TestDnnInferShapes(utt.InferShapeTester):
dnn.GpuDnnPool dnn.GpuDnnPool
) )
def test_pool_3d(self):
if not dnn.dnn_available():
raise SkipTest(dnn.dnn_available.msg)
img = T.ftensor5('img')
img_val = numpy.asarray(
numpy.random.rand(2, 3, 4, 5, 6),
dtype='float32'
)
# 'average_exc_pad' is disabled for versions < 4004
if cuda.dnn.version() < (4004, 4004):
modes = ['max', 'average_inc_pad']
else:
modes = ['max', 'average_inc_pad', 'average_exc_pad']
for params in product(
[(1, 1, 1), (2, 2, 2), (3, 3, 3)],
[(1, 1, 1), (2, 2, 2), (3, 3, 3)],
modes
):
self._compile_and_check(
[img],
[dnn.GpuDnnPool(mode=params[2])(img, params[0], params[1], (0, 0, 0))],
[img_val],
dnn.GpuDnnPool
)
def test_pool_grad(self): def test_pool_grad(self):
if not dnn.dnn_available(): if not dnn.dnn_available():
raise SkipTest(dnn.dnn_available.msg) raise SkipTest(dnn.dnn_available.msg)
...@@ -1308,6 +1242,45 @@ class TestDnnInferShapes(utt.InferShapeTester): ...@@ -1308,6 +1242,45 @@ class TestDnnInferShapes(utt.InferShapeTester):
dnn.GpuDnnPoolGrad dnn.GpuDnnPoolGrad
) )
def test_pool_3d_grad(self):
if not dnn.dnn_available():
raise SkipTest(dnn.dnn_available.msg)
img = T.ftensor5('img')
img_grad = T.ftensor5('img_grad')
out = T.ftensor5('out')
img_val = numpy.asarray(
numpy.random.rand(2, 3, 4, 5, 6),
dtype='float32'
)
img_grad_val = numpy.asarray(
numpy.random.rand(2, 3, 4, 5, 6),
dtype='float32'
)
out_val = numpy.asarray(
numpy.random.rand(2, 3, 4, 5, 6),
dtype='float32'
)
for params in product(
[(1, 1, 1), (2, 2, 2), (3, 3, 3)],
[(1, 1, 1), (2, 2, 2), (3, 3, 3)],
['max', 'average_inc_pad']
):
pool_grad = dnn.GpuDnnPoolGrad(mode=params[2])(
img,
out,
img_grad,
params[0],
params[1],
(0, 0, 0)
)
self._compile_and_check(
[img, img_grad, out],
[pool_grad],
[img_val, img_grad_val, out_val],
dnn.GpuDnnPoolGrad
)
# this has been a problem in the past # this has been a problem in the past
def test_dnn_conv_border_mode(): def test_dnn_conv_border_mode():
......
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