Skip to content

P
pytensor
提交 330f40aa authored 作者: Frederic's avatar Frederic
浏览文件
操作

merged the test for gpumax and gpusum.

上级 e0516eef
显示空白字符变更
内嵌 并排
正在显示 包含 61 行增加 和 218 行删除
theano/sandbox/cuda/tests/test_basic_ops.py
浏览文件 @ 330f40aa
...@@ -42,7 +42,15 @@ def tes_use(): ...@@ -42,7 +42,15 @@ def tes_use():
tcn.use() tcn.use()
def test_sum(): def tensor_pattern_to_gpu_pattern(shape, pattern):
gpu_pattern = [0 for elem in shape]
for idx in pattern:
gpu_pattern[idx] = 1
gpu_pattern = tuple(gpu_pattern)
return gpu_pattern
def test_careduce():
""" """
test sum pattern 1, 11, 10, 01, 001, 010, 100, 110, 011, 111, test sum pattern 1, 11, 10, 01, 001, 010, 100, 110, 011, 111,
0011, 0101, 0111, 1011, 1111 0011, 0101, 0111, 1011, 1111
...@@ -56,6 +64,7 @@ def test_sum(): ...@@ -56,6 +64,7 @@ def test_sum():
TODO: test with broadcast TODO: test with broadcast
""" """
for scalar_op in [theano.scalar.add, theano.scalar.maximum]:
for shape, pattern in [((1,1),(1,)), for shape, pattern in [((1,1),(1,)),
((1,0),(1,)), ((1,0),(1,)),
((0,1),(1,)), ((0,1),(1,)),
...@@ -109,174 +118,26 @@ def test_sum(): ...@@ -109,174 +118,26 @@ def test_sum():
((1100,2,3,4,5),[0,1,2,3,4]),((2,1100,3,4,5),[0,1,2,3,4]),((2,3,1100,4,5),[0,1,2,3,4]),((2,3,4,1100,5),[0,1,2,3,4]),((2,3,4,5,1100),[0,1,2,3,4]),#11111 ((1100,2,3,4,5),[0,1,2,3,4]),((2,1100,3,4,5),[0,1,2,3,4]),((2,3,1100,4,5),[0,1,2,3,4]),((2,3,4,1100,5),[0,1,2,3,4]),((2,3,4,5,1100),[0,1,2,3,4]),#11111
]: ]:
a = tensor.TensorType('float32', (False,) * len(shape))()
b = T.Sum(pattern)(a)
val = numpy.random.rand(numpy.prod(shape)).reshape(shape)
# val = numpy.ones(shape)
# val = numpy.arange(numpy.prod(shape)).reshape(shape)
val = theano._asarray(val, dtype='float32')
f = theano.function([a], b, mode=mode_with_gpu)
f2 = theano.function([a], b, mode=mode_without_gpu)
assert tcn.GpuCAReduce in [x.op.__class__ for x in f.maker.fgraph.toposort()]
assert T.Sum in [x.op.__class__ for x in f2.maker.fgraph.toposort()]
if val.size == 0:
assert _allclose(f2(val), f(val)), ('shape', shape, 'pattern', pattern)
else:
try:
#We raise the error threashold as we sum big matrix
#and this cause small rounding difference with some seed
#example in debug mode with unittests.rseed=9275
orig_rtol = theano.tensor.basic.float32_rtol
theano.tensor.basic.float32_rtol = 2e-5
assert _allclose(f2(val), f(val)), ('shape', shape,
'pattern', pattern,
sum([shape[i] for i in pattern]),
f2(val), f(val), val)
finally:
theano.tensor.basic.float32_rtol = orig_rtol
#test with dimshuffle
#we shuffle the 2 outer dims.
for shape, pattern in [#((5,),[0]),
((5,4),[0,1]),((5,4),[0]),
((5,4,3),[0]),((5,4,3),[0,1]),((5,4,3),[2]),((5,4,3),[0,1,2]),
((5,4,3,2),[0,1,2,3]), ((5,4,3,2),[0,2,3])]:
a = tensor.TensorType('float32', (False,) * len(shape))()
dim_pattern = range(len(shape))
dim_pattern[0] = 1
dim_pattern[1] = 0
a = a.dimshuffle(dim_pattern)
b = T.Sum(pattern)(a)
val = numpy.random.rand(numpy.prod(shape)).reshape(shape)
# val = numpy.ones(shape)
# val = numpy.arange(numpy.prod(shape)).reshape(shape)
val = theano._asarray(val, dtype='float32')
f = theano.function([a], b, mode=mode_with_gpu)
f2 = theano.function([a], b, mode=mode_without_gpu)
assert tcn.GpuCAReduce in [x.op.__class__ for x in f.maker.fgraph.toposort()]
assert T.Sum in [x.op.__class__ for x in f2.maker.fgraph.toposort()]
assert _allclose(f2(val), f(val)), ('shape', shape,
'pattern', pattern,
sum([shape[i] for i in pattern]))
#test with broadcast
for shape, pattern in [((5,),[0]),
((5,4),[0,1]),((5,4),[0]),
((5,4,3),[0]),((5,4,3),[0,1]),((5,4,3),[2]),((5,4,3),[0,1,2]),
((5,4,3,2),[0,1,2,3]), ((5,4,3,2),[0,2,3])]:
shape = numpy.asarray(shape) * 2
a = tensor.TensorType('float32', (False,) * len(shape))()
a2 = tcn.CudaNdarrayType((False,) * len(shape))()
b = T.Sum(pattern)(a)
b2 = T.Sum(pattern)(a2)
val = numpy.random.rand(numpy.prod(shape)).reshape(shape)
# val = numpy.ones(shape)
# val = numpy.arange(numpy.prod(shape)).reshape(shape)
val = theano._asarray(val, dtype='float32')
val2 = cuda.CudaNdarray(val)
if len(shape) == 1:
val = val[::2]
val2 = val2[::2]
elif len(shape) == 2:
val = val[::2, ::2]
val2 = val2[::2, ::2]
elif len(shape) == 3:
val = val[::2, ::2, ::2]
val2 = val2[::2, ::2, ::2]
elif len(shape) == 4:
val = val[::2, ::2, ::2, ::2]
val2 = val2[::2, ::2, ::2, ::2]
f = theano.function([a], b, mode=mode_without_gpu)
f2 = theano.function([a2], b2, mode=mode_with_gpu)
assert tcn.GpuCAReduce in [x.op.__class__ for x in f2.maker.fgraph.toposort()]
assert T.Sum in [x.op.__class__ for x in f.maker.fgraph.toposort()]
assert _allclose(f2(val2), f(val)), ('shape', shape,
'pattern', pattern,
sum([shape[i] for i in pattern]))
def test_max():
"""
test GpuMax pattern 01, 011, 0111 (tensor.max pattern (1,), (1,2), (1,2,3) )
TODO: are others currently implemented by reshape?
"""
def tensor_pattern_to_gpu_pattern(shape, pattern):
gpu_pattern = [ 0 for elem in shape ]
for idx in pattern:
gpu_pattern[idx] = 1
gpu_pattern = tuple(gpu_pattern)
return gpu_pattern
known_fail = False
for shape, pattern in [((1,1),(1,)),
((1,0),(1,)),
((0,1),(1,)),
((0,0),(1,)),
((0,0,0),(1,2)),
((0,0,0,0),(1,2,3)),
((2,1),(1,)),
((1,2),(1,)),
((100,3,1300),[1]),
((0,),[0]),((5,),[0]),
((0,0),[0,1]),((1,0),[0,1]),((5,4),[0,1]),((33,31),[0,1]),((5,4),[1]),((5,4),[0]),#need something bigger then 32 for some opt test.
((5,4,3),[0]),((5,4,3),[1]),((5,4,3),[0,1]),((5,4,3),[2]),((5,4,3),[1,2]),((5,4,3),[0,1,2]),
((0,0,0,0),[0,1,2,3]),
((5,4,3,20),[2,3]), ((5,4,3,2),[0,1,2,3]), ((5,4,3,2),[0,2,3]),((5,4,3,2),[1,2,3]),
((5,4,3,10,11),[1,2]),
((5,4,3,20),[2,3]), ((5,4,3,2),[0,1,2,3]), ((5,4,3,2),[0,2,3]),((5,4,3,2),[1,2,3]),
#test shape bigger then 4096 on each dimension to make sure that we work correctly when we don't have enough thread/block in each dimensions
((4100,3),[0]),((3,4101),[0]),#10
((1024,33),[0]),((33,1024),[0]),#10
((1025,33),[0]),((33,1025),[0]),#10
((4100,3),[1]),((3,4101),[1]),#01
((1024,33),[1]),((33,1024),[1]),#01
((1025,33),[1]),((33,1025),[1]),#01
((4100,3),[0,1]),((3,4101),[0,1]),#11
((1024,33),[0,1]),((33,1024),[0,1]),#01
((1025,33),[0,1]),((33,1025),[0,1]),#01
((4100,4,3),[0]),((5,4100,3),[0]),((5,4,4100),[0]),#100 op = tensor.CAReduce(scalar_op, axis=pattern)
((4100,4,3),[1]),((5,4100,3),[1]),((5,4,4100),[1]),#010 pat = tensor_pattern_to_gpu_pattern(shape, pattern)
((4100,4,3),[2]),((5,4100,3),[2]),((5,4,4100),[2]),#001 #GpuCAReduce{maximum} support only those patterns
((4100,4,3),[0,1]),((5,4100,3),[0,1]),((5,4,4100),[0,1]),#110 if scalar_op is theano.scalar.maximum and pat not in [
((4100,4,3),[1,2]),((5,4100,3),[1,2]),((5,4,4100),[1,2]),#011 (0, 1), (0, 1, 1), (0, 1, 1)]:
#((4100,4,3),[0,2]),((5,4100,3),[0,2]),((5,4,4100),[0,2]),#101 ##not implemented
((4100,4,3),[0,1,2]),((5,4100,3),[0,1,2]),((5,4,4100),[0,1,2]),#111
((4100,4,3,2),[2,3]),((4,4100,3,2),[2,3]),((4,3,4100,2),[2,3]),((4,3,2,4100),[2,3]),#0011
((4100,4,3,2),[1,3]),((4,4100,3,2),[1,3]),((4,3,4100,2),[1,3]),((4,3,2,4100),[1,3]),#0101
((4100,4,3,2),[0,2,3]),((4,4100,3,2),[0,2,3]),((4,3,4100,2),[0,2,3]),#((4,3,2,4100),[0,2,3]),#1011
((4100,4,3,2),[1,2,3]),((4,4100,3,2),[1,2,3]),((4,3,4100,2),[1,2,3]),((4,3,2,4100),[1,2,3]),#0111
((4100,2,3,4),[0,1,2,3]),((2,4100,3,4),[0,1,2,3]),((2,3,4100,4),[0,1,2,3]),((2,3,4,4100),[0,1,2,3]),#1111
#test pattern implemented by reshape
((4100,4,3,2),[0]),((4,4100,3,2),[0]),((4,3,4100,2),[0]),((4,3,2,4100),[0]),#1000
((4100,4,3,2),[1]),((4,4100,3,2),[1]),((4,3,4100,2),[1]),((4,3,2,4100),[1]),#0100
((4100,4,3,2),[2]),((4,4100,3,2),[2]),((4,3,4100,2),[2]),((4,3,2,4100),[2]),#0010
((4100,4,3,2),[3]),((4,4100,3,2),[3]),((4,3,4100,2),[3]),((4,3,2,4100),[3]),#0001
((1100,2,3,4,5),[0,1,2,3,4]),((2,1100,3,4,5),[0,1,2,3,4]),((2,3,1100,4,5),[0,1,2,3,4]),((2,3,4,1100,5),[0,1,2,3,4]),((2,3,4,5,1100),[0,1,2,3,4]),#11111
]:
# Don't test patterns that aren't implemented for max yet
if tensor_pattern_to_gpu_pattern(shape, pattern) not in \
[ (0,1), (0,1,1), (0,1,1) ]:
continue continue
a = tensor.TensorType('float32', (False,) * len(shape))() a = tensor.TensorType('float32', (False,) * len(shape))()
b = T.max(a, pattern) b = op(a)
val = numpy.random.rand(numpy.prod(shape)).reshape(shape) val = numpy.random.rand(numpy.prod(shape)).reshape(shape)
# val = numpy.ones(shape) # val = numpy.ones(shape)
# val = numpy.arange(numpy.prod(shape)).reshape(shape) # val = numpy.arange(numpy.prod(shape)).reshape(shape)
val = theano._asarray(val, dtype='float32') val = theano._asarray(val, dtype='float32')
f = theano.function([a], b, mode=mode_with_gpu) f = theano.function([a], b, mode=mode_with_gpu)
f2 = theano.function([a], b, mode=mode_without_gpu)
assert tcn.GpuCAReduce in [x.op.__class__
for x in f.maker.fgraph.toposort()]
assert op.__class__ in [x.op.__class__
for x in f2.maker.fgraph.toposort()]
f_caused_value_error = False f_caused_value_error = False
try: try:
f_out = f(val) f_out = f(val)
...@@ -284,19 +145,13 @@ def test_max(): ...@@ -284,19 +145,13 @@ def test_max():
exc = e exc = e
f_caused_value_error = True f_caused_value_error = True
f2 = theano.function([a], b, mode=mode_without_gpu) f2_caused_value_error = False
try: try:
f2_out = f2(val) f2_out = f2(val)
f2_caused_value_error = False
except ValueError, e: except ValueError, e:
exc2 = e exc2 = e
f2_caused_value_error = True f2_caused_value_error = True
assert tcn.GpuCAReduce in [x.op.__class__ for x in f.maker.fgraph.toposort()]
assert T.CAReduce in [x.op.__class__ for x in f2.maker.fgraph.toposort()]
# Check that 0 shape matrices are invalid in the same cases
if f_caused_value_error != f2_caused_value_error: if f_caused_value_error != f2_caused_value_error:
if f_caused_value_error: if f_caused_value_error:
print 'f caused this value error:' print 'f caused this value error:'
...@@ -308,36 +163,20 @@ def test_max(): ...@@ -308,36 +163,20 @@ def test_max():
print exc2 print exc2
else: else:
print 'f should not have raised a value error' print 'f should not have raised a value error'
print 'shape was: ',shape print 'shape was: ', shape
print 'pattern was: ',pattern print 'pattern was: ', pattern
assert False assert False
if f_caused_value_error:
continue
if val.size == 0:
assert f2(val).size == f(val).size
assert f2(val).shape == f(val).shape
else:
try: try:
#We raise the error threashold as we sum big matrix #We raise the error threashold as we sum big matrix
#and this cause small rounding difference with some seed #and this cause small rounding difference with some seed
#example in debug mode with unittests.rseed=9275 #example in debug mode with unittests.rseed=9275
orig_rtol = theano.tensor.basic.float32_rtol orig_rtol = theano.tensor.basic.float32_rtol
theano.tensor.basic.float32_rtol = 2e-5 theano.tensor.basic.float32_rtol = 2e-5
f2_val = f2(val) assert _allclose(f_out, f2_out), ('shape', shape,
f_val = f(val) 'pattern', pattern,
if not _allclose(f2_val, f_val): sum([shape[i] for i in pattern]),
print 'failed for the following arguments: ' f2(val), f(val), val)
print 'shape:',shape
print 'pattern: ',pattern
print 'input:'
print val
print 'correct output: '
print f2_val
print 'actual output: '
print f_val
assert False
finally: finally:
theano.tensor.basic.float32_rtol = orig_rtol theano.tensor.basic.float32_rtol = orig_rtol
...@@ -345,53 +184,55 @@ def test_max(): ...@@ -345,53 +184,55 @@ def test_max():
#test with dimshuffle #test with dimshuffle
#we shuffle the 2 outer dims. #we shuffle the 2 outer dims.
for shape, pattern in [#((5,),[0]), for shape, pattern in [#((5,),[0]),
((5,4),(0,1)),((5,4),[0]), ((5,4),[0,1]),((5,4),[0]),
((5,4,3),[0]),((5,4,3),[0,1]),((5,4,3),[2]),((5,4,3),[0,1,2]), ((5,4,3),[0]),((5,4,3),[0,1]),((5,4,3),[2]),((5,4,3),[0,1,2]),
((5,4,3,2),[0,1,2,3]), ((5,4,3,2),[0,2,3])]: ((5,4,3,2),[0,1,2,3]), ((5,4,3,2),[0,2,3])]:
# Don't test patterns that aren't implemented for max yet op = tensor.CAReduce(scalar_op, axis=pattern)
if tensor_pattern_to_gpu_pattern(shape, pattern) not in \ pat = tensor_pattern_to_gpu_pattern(shape, pattern)
[ (0,1), (0,1,1), (0,1,1) ]: #GpuCAReduce{maximum} support only those patterns
if scalar_op is theano.scalar.maximum and pat not in [
(0, 1), (0, 1, 1), (0, 1, 1)]:
continue continue
a = tensor.TensorType('float32', (False,) * len(shape))() a = tensor.TensorType('float32', (False,) * len(shape))()
dim_pattern = range(len(shape)) dim_pattern = range(len(shape))
dim_pattern[0] = 1 dim_pattern[0] = 1
dim_pattern[1] = 0 dim_pattern[1] = 0
a = a.dimshuffle(dim_pattern) a = a.dimshuffle(dim_pattern)
b = T.max(a, pattern) b = op(a)
val = numpy.random.rand(numpy.prod(shape)).reshape(shape) val = numpy.random.rand(numpy.prod(shape)).reshape(shape)
# val = numpy.ones(shape) # val = numpy.ones(shape)
# val = numpy.arange(numpy.prod(shape)).reshape(shape) # val = numpy.arange(numpy.prod(shape)).reshape(shape)
val = theano._asarray(val, dtype='float32') val = theano._asarray(val, dtype='float32')
f = theano.function([a], b, mode=mode_with_gpu) f = theano.function([a], b, mode=mode_with_gpu)
f2 = theano.function([a], b, mode=mode_without_gpu) f2 = theano.function([a], b, mode=mode_without_gpu)
assert tcn.GpuCAReduce in [x.op.__class__ for x in f.maker.fgraph.toposort()] assert tcn.GpuCAReduce in [x.op.__class__
assert T.CAReduce in [x.op.__class__ for x in f2.maker.fgraph.toposort()] for x in f.maker.fgraph.toposort()]
assert op.__class__ in [x.op.__class__
for x in f2.maker.fgraph.toposort()]
assert _allclose(f2(val), f(val)), ('shape', shape, assert _allclose(f2(val), f(val)), ('shape', shape,
'pattern', pattern, 'pattern', pattern,
sum([shape[i] for i in pattern])) sum([shape[i] for i in pattern]))
#test with broadcast #test with broadcast
for shape, pattern in [((5,),(0,)), for shape, pattern in [((5,),[0]),
((5,4),(0,1)), ((5,4),[0,1]),((5,4),[0]),
((5,4),(0,)), ((5,4,3),[0]),((5,4,3),[0,1]),
((5,4,3),(0,)), ((5,4,3),[2]),((5,4,3),[0,1,2]),
((5,4,3),(0,1)), ((5,4,3,2),[0,1,2,3]), ((5,4,3,2),[0,2,3])]:
((5,4,3),(2,)), op = tensor.CAReduce(scalar_op, axis=pattern)
((5,4,3),(0,1,2)), pat = tensor_pattern_to_gpu_pattern(shape, pattern)
((5,4,3,2),(0,1,2,3)), #GpuCAReduce{maximum} support only those patterns
((5,4,3,2),(0,2,3))]: if scalar_op is theano.scalar.maximum and pat not in [
# Don't test patterns that aren't implemented for max yet (0, 1), (0, 1, 1), (0, 1, 1)]:
if tensor_pattern_to_gpu_pattern(shape, pattern) not in \
[ (0,1), (0,1,1), (0,1,1) ]:
continue continue
shape = numpy.asarray(shape) * 2 shape = numpy.asarray(shape) * 2
a = tensor.TensorType('float32', (False,) * len(shape))() a = tensor.TensorType('float32', (False,) * len(shape))()
a2 = tcn.CudaNdarrayType((False,) * len(shape))() a2 = tcn.CudaNdarrayType((False,) * len(shape))()
b = T.max(a, pattern) b = op(a)
b2 = T.max(a2, pattern) b2 = op(a2)
val = numpy.random.rand(numpy.prod(shape)).reshape(shape) val = numpy.random.rand(numpy.prod(shape)).reshape(shape)
# val = numpy.ones(shape) # val = numpy.ones(shape)
# val = numpy.arange(numpy.prod(shape)).reshape(shape) # val = numpy.arange(numpy.prod(shape)).reshape(shape)
val = theano._asarray(val, dtype='float32') val = theano._asarray(val, dtype='float32')
val2 = cuda.CudaNdarray(val) val2 = cuda.CudaNdarray(val)
if len(shape) == 1: if len(shape) == 1:
...@@ -408,8 +249,10 @@ def test_max(): ...@@ -408,8 +249,10 @@ def test_max():
val2 = val2[::2, ::2, ::2, ::2] val2 = val2[::2, ::2, ::2, ::2]
f = theano.function([a], b, mode=mode_without_gpu) f = theano.function([a], b, mode=mode_without_gpu)
f2 = theano.function([a2], b2, mode=mode_with_gpu) f2 = theano.function([a2], b2, mode=mode_with_gpu)
assert tcn.GpuCAReduce in [x.op.__class__ for x in f2.maker.fgraph.toposort()] assert tcn.GpuCAReduce in [x.op.__class__
assert T.CAReduce in [x.op.__class__ for x in f.maker.fgraph.toposort()] for x in f2.maker.fgraph.toposort()]
assert op.__class__ in [x.op.__class__
for x in f.maker.fgraph.toposort()]
assert _allclose(f2(val2), f(val)), ('shape', shape, assert _allclose(f2(val2), f(val)), ('shape', shape,
'pattern', pattern, 'pattern', pattern,
sum([shape[i] for i in pattern])) sum([shape[i] for i in pattern]))
......
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