Skip to content

P
pytensor
提交 8f0b0888 authored 作者: Frédéric Bastien's avatar Frédéric Bastien 提交者： GitHub
浏览文件
操作

Merge pull request #4039 from Sentient07/gpu_optimization

Making DimShuffles view by default
上级 8682627a 3da3054e
隐藏空白字符变更
内嵌 并排
正在显示 包含 106 行增加 和 222 行删除
doc/extending/graphstructures.txt
浏览文件 @ 8f0b0888
......@@ -99,7 +99,7 @@ multiplication is done between the inputs:
>>> y.owner.inputs[0]
x
>>> y.owner.inputs[1]
DimShuffle{x,x}.0
InplaceDimShuffle{x,x}.0
Note that the second input is not 2 as we would have expected. This is
because 2 was first :term:`broadcasted <broadcasting>` to a matrix of
......
doc/library/printing.txt
浏览文件 @ 8f0b0888
......@@ -118,7 +118,7 @@ Elemwise{mul} [id A] ''
|x [id E]
|Elemwise{sub} [id I] ''
|TensorConstant{2} [id F]
|DimShuffle{} [id J] ''
|InplaceDimShuffle{} [id J] ''
|TensorConstant{1} [id K]
>>> theano.printing.debugprint(gy, depth=2) # doctest: +NORMALIZE_WHITESPACE
......
doc/library/tensor/basic.txt
浏览文件 @ 8f0b0888
......@@ -615,13 +615,13 @@ dimensions, see :meth:`_tensor_py_operators.dimshuffle`.
>>> tensor = theano.tensor.tensor3()
>>> theano.tensor.shape_padaxis(tensor, axis=0)
DimShuffle{x,0,1,2}.0
InplaceDimShuffle{x,0,1,2}.0
>>> theano.tensor.shape_padaxis(tensor, axis=1)
DimShuffle{0,x,1,2}.0
InplaceDimShuffle{0,x,1,2}.0
>>> theano.tensor.shape_padaxis(tensor, axis=3)
DimShuffle{0,1,2,x}.0
InplaceDimShuffle{0,1,2,x}.0
>>> theano.tensor.shape_padaxis(tensor, axis=-1)
DimShuffle{0,1,2,x}.0
InplaceDimShuffle{0,1,2,x}.0
.. autofunction:: unbroadcast(x, *axes)
......
doc/tutorial/printing_drawing.txt
浏览文件 @ 8f0b0888
......@@ -69,10 +69,10 @@ The pre-compilation graph:
>>> theano.printing.debugprint(prediction) # doctest: +NORMALIZE_WHITESPACE, +ELLIPSIS
Elemwise{gt,no_inplace} [id A] ''
|Elemwise{true_div,no_inplace} [id B] ''
| |DimShuffle{x} [id C] ''
| |InplaceDimShuffle{x} [id C] ''
| | |TensorConstant{1} [id D]
| |Elemwise{add,no_inplace} [id E] ''
| |DimShuffle{x} [id F] ''
| |InplaceDimShuffle{x} [id F] ''
| | |TensorConstant{1} [id D]
| |Elemwise{exp,no_inplace} [id G] ''
| |Elemwise{sub,no_inplace} [id H] ''
......@@ -80,9 +80,9 @@ Elemwise{gt,no_inplace} [id A] ''
| | |dot [id J] ''
| | |x [id K]
| | |w [id L]
| |DimShuffle{x} [id M] ''
| |InplaceDimShuffle{x} [id M] ''
| |b [id N]
|DimShuffle{x} [id O] ''
|InplaceDimShuffle{x} [id O] ''
|TensorConstant{0.5} [id P]
The post-compilation graph:
......
theano/sandbox/cuda/basic_ops.py
浏览文件 @ 8f0b0888
......@@ -121,14 +121,7 @@ class GpuFromHost(GpuOp):
check_input = False
def __eq__(self, other):
return type(self) == type(other)
def __hash__(self):
return hash(type(self))
def __str__(self):
return 'GpuFromHost'
__props__ = ()
def make_node(self, x):
if not isinstance(x.type, tensor.TensorType):
......@@ -220,12 +213,6 @@ class GpuElemwise(GpuOp):
self.sync = d.get('sync', True)
self._rehash()
def __eq__(self, other):
return (type(self) == type(other) and
self.scalar_op == other.scalar_op and
self.inplace_pattern == other.inplace_pattern and
self.sync == other.sync)
def _rehash(self):
items = list(self.inplace_pattern.items())
items.sort()
......@@ -242,6 +229,12 @@ class GpuElemwise(GpuOp):
assert h == getattr(self, '_hashval', h)
self._hashval = h
def __eq__(self, other):
return (type(self) == type(other) and
self.scalar_op == other.scalar_op and
self.inplace_pattern == other.inplace_pattern and
self.sync == other.sync)
def __hash__(self):
return self._hashval
......@@ -320,6 +313,8 @@ class GpuDimShuffle(GpuOp):
check_broadcast = False
__props__ = ("input_broadcastable", "new_order")
def __init__(self, input_broadcastable, new_order):
input_broadcastable = tuple(input_broadcastable)
self.input_broadcastable = input_broadcastable
......@@ -342,17 +337,6 @@ class GpuDimShuffle(GpuOp):
self.view_map = {0: [0]}
self._rehash()
def __getstate__(self):
d = dict(self.__dict__)
del d['_hashval']
return d
def __setstate__(self, d):
self.__dict__.update(d)
self._rehash()
def make_node(self, input):
ib = tuple(input.type.broadcastable)
if not ib == self.input_broadcastable:
......@@ -379,21 +363,6 @@ class GpuDimShuffle(GpuOp):
ob.append(ib[value])
return Apply(self, [input], [CudaNdarrayType(broadcastable=ob)()])
def __eq__(self, other):
# it's probably not necessary to compare input_broadcastable
return type(self) == type(other) \
and self.new_order == other.new_order \
and self.input_broadcastable == other.input_broadcastable
def _rehash(self):
self._hashval = (hash(type(self).__name__) ^
hash(type(self).__module__) ^
hash(self.new_order) ^
hash(self.input_broadcastable))
def __hash__(self):
return self._hashval
def __str__(self):
return "GpuDimShuffle{%s}" % ",".join(str(x) for x in self.new_order)
......@@ -568,6 +537,8 @@ class GpuCAReduce(GpuOp):
"""
__props__ = ("reduce_mask", "scalar_op", "pre_scalar_op", )
def __init__(self, reduce_mask, scalar_op, pre_scalar_op=None):
self.reduce_mask = tuple(reduce_mask)
self.scalar_op = scalar_op
......@@ -578,17 +549,11 @@ class GpuCAReduce(GpuOp):
if pre_scalar_op:
assert pre_scalar_op.nin == 1
def __eq__(self, other):
return (type(self) == type(other) and
self.reduce_mask == other.reduce_mask and
self.scalar_op == other.scalar_op and
self.pre_scalar_op == other.pre_scalar_op)
def __hash__(self):
return (hash(type(self)) ^
hash(self.reduce_mask) ^
hash(type(self.scalar_op)) ^
hash(type(self.pre_scalar_op)))
def __setstate__(self, d):
self.__dict__.update(d)
# For unpickling of old ops.
if not hasattr(self, "pre_scalar_op"):
self.pre_scalar_op = None
def __str__(self):
pre = ""
......@@ -598,12 +563,6 @@ class GpuCAReduce(GpuOp):
pre, str(self.scalar_op),
','.join(str(i) for i in self.reduce_mask))
def __setstate__(self, d):
self.__dict__.update(d)
# For unpickling of old ops.
if not hasattr(self, "pre_scalar_op"):
self.pre_scalar_op = None
def make_node(self, x):
x = as_cuda_ndarray_variable(x)
if (x.type.ndim != len(self.reduce_mask)):
......@@ -3655,7 +3614,6 @@ class GpuAllocEmpty(GpuOp):
Implement Alloc on the gpu, but without initializing memory.
"""
__props__ = ()
def make_node(self, *shape):
......@@ -3852,10 +3810,10 @@ class CopyOnNegativeStrides(GpuOp):
If it does, returns a c contiguous copy.
"""
__props__ = ()
view_map = {0: [0]}
check_input = False
__props__ = ()
def grad(self, inputs, dout):
......
theano/sandbox/cuda/blas.py
浏览文件 @ 8f0b0888
......@@ -2207,18 +2207,12 @@ class GpuDownsampleFactorMax(GpuOp):
Implement downsample with max on the gpu.
"""
__props__ = ('ds', 'ignore_border')
def __init__(self, ds, ignore_border=False):
self.ds = tuple(ds)
self.ignore_border = ignore_border
def __eq__(self, other):
return (type(self) == type(other) and
self.ds == other.ds and
self.ignore_border == other.ignore_border)
def __hash__(self):
return hash(type(self)) ^ hash(self.ds) ^ hash(self.ignore_border)
def __str__(self):
return '%s{%s,%s}' % (self.__class__.__name__,
self.ds,
......
theano/sandbox/cuda/opt.py
浏览文件 @ 8f0b0888
......@@ -379,7 +379,7 @@ def local_gpu_split(node):
if (input.owner and isinstance(input.owner.op, HostFromGpu) or
any(c != 'output' and isinstance(c.op, GpuFromHost) for c, idx
in outs_clients)):
new_op = GpuSplit(node.op.len_splits)
new_op = GpuSplit(**node.op._props_dict())
split_res = new_op(as_cuda_ndarray_variable(input),
*node.inputs[1:], return_list=True)
return [host_from_gpu(o) for o in split_res]
......@@ -398,16 +398,18 @@ def local_gpu_dimshuffle_0(node):
input, = node.inputs
if input.owner and isinstance(input.owner.op, HostFromGpu):
# move the add to a GpuAdd
new_op = GpuDimShuffle(node.op.input_broadcastable,
node.op.new_order)
p_dict = node.op._props_dict()
p_dict.pop('inplace', None)
new_op = GpuDimShuffle(**p_dict)
return [host_from_gpu(new_op(as_cuda_ndarray_variable(input)))]
if isinstance(node.op, GpuFromHost):
host_input = node.inputs[0]
if host_input.owner and isinstance(host_input.owner.op,
tensor.DimShuffle):
dimshuffle_node = host_input.owner
new_op = GpuDimShuffle(dimshuffle_node.op.input_broadcastable,
dimshuffle_node.op.new_order)
p_dict = dimshuffle_node.op._props_dict()
p_dict.pop('inplace', None)
new_op = GpuDimShuffle(**p_dict)
return [new_op(
as_cuda_ndarray_variable(dimshuffle_node.inputs[0]))]
return False
......@@ -995,10 +997,8 @@ def local_gpu_reshape(node):
host_input = node.inputs[0]
if host_input.owner and \
isinstance(host_input.owner.op, tensor.Reshape):
rshp = host_input.owner.op
x, shp = host_input.owner.inputs
gpu_reshape = GpuReshape(rshp.ndim)(as_cuda_ndarray_variable(x),
shp)
gpu_reshape = GpuReshape(**host_input.owner.op._props_dict())(as_cuda_ndarray_variable(x), shp)
if gpu_reshape.broadcastable != node.outputs[0].broadcastable:
# this can happen as we always return False for all broadcast
# dim in GpuReshape but not for Reshape
......@@ -1011,7 +1011,7 @@ def local_gpu_reshape(node):
x, shp = node.inputs
if x.owner and isinstance(x.owner.op, HostFromGpu):
gpu_x, = x.owner.inputs
gpu_reshape = GpuReshape(node.op.ndim)(gpu_x, shp)
gpu_reshape = GpuReshape(**node.op._props_dict())(gpu_x, shp)
if gpu_reshape.broadcastable != node.outputs[0].broadcastable:
# this can happen as we always return False for all broadcast
# dim in GpuReshape but not for Reshape
......@@ -1082,7 +1082,7 @@ def local_gpu_subtensor(node):
gpu_x, = x.owner.inputs
coords = node.inputs[1:]
return [host_from_gpu(GpuSubtensor(
node.op.idx_list)(gpu_x, *coords))]
**node.op._props_dict())(gpu_x, *coords))]
return False
......@@ -1131,11 +1131,9 @@ def local_gpu_advanced_incsubtensor1(node):
compute_capability = device_properties(active_device_no)['major']
if (compute_capability < 2 or y.ndim != 2 or x.ndim != 2):
gpu_op = GpuAdvancedIncSubtensor1(
set_instead_of_inc=set_instead_of_inc)
gpu_op = GpuAdvancedIncSubtensor1(**node.op._props_dict())
else:
gpu_op = GpuAdvancedIncSubtensor1_dev20(
set_instead_of_inc=set_instead_of_inc)
gpu_op = GpuAdvancedIncSubtensor1_dev20(**node.op._props_dict())
return [gpu_op(as_cuda_ndarray_variable(x),
as_cuda_ndarray_variable(y), *coords)]
......@@ -1171,11 +1169,9 @@ def local_gpu_advanced_incsubtensor1(node):
active_device_no = theano.sandbox.cuda.active_device_number()
compute_capability = device_properties(active_device_no)['major']
if (compute_capability < 2 or y.ndim != 2 or x.ndim != 2):
gpu_op = GpuAdvancedIncSubtensor1(
set_instead_of_inc=set_instead_of_inc)
gpu_op = GpuAdvancedIncSubtensor1(**node.op._props_dict())
else:
gpu_op = GpuAdvancedIncSubtensor1_dev20(
set_instead_of_inc=set_instead_of_inc)
gpu_op = GpuAdvancedIncSubtensor1_dev20(**node.op._props_dict())
return [host_from_gpu(gpu_op(gpu_x, gpu_y, *coords))]
return False
......@@ -1196,13 +1192,9 @@ def local_gpu_incsubtensor(node):
# The IncSubtensor upcast to float32 y, so we do it
# explicitly to move it to the GPU.
y = y.astype('float32')
ret = GpuIncSubtensor(
incsubt.idx_list,
inplace=incsubt.inplace,
set_instead_of_inc=incsubt.set_instead_of_inc)(
as_cuda_ndarray_variable(x),
as_cuda_ndarray_variable(y),
*coords)
ret = GpuIncSubtensor(**incsubt._props_dict())(as_cuda_ndarray_variable(x),
as_cuda_ndarray_variable(y),
*coords)
ret.tag.nan_guard_mode_check = getattr(
host_output.tag, 'nan_guard_mode_check', True)
return [ret]
......@@ -1229,10 +1221,7 @@ def local_gpu_incsubtensor(node):
y = tensor.cast(y, 'float32')
gpu_y = as_cuda_ndarray_variable(y)
if go_gpu:
ret = GpuIncSubtensor(
node.op.idx_list, inplace=node.op.inplace,
set_instead_of_inc=node.op.set_instead_of_inc)(
gpu_x, gpu_y, *coords)
ret = GpuIncSubtensor(**node.op._props_dict())(gpu_x, gpu_y, *coords)
val = getattr(node.outputs[0].tag, 'nan_guard_mode_check', True)
ret.tag.nan_guard_mode_check = val
......@@ -2690,7 +2679,7 @@ def gpu_sparse_block_outer_opt(node):
inputs = _clear_host_from_gpu(node.inputs)
return [host_from_gpu(GpuSparseBlockOuter(node.op.inplace)(*inputs))]
return [host_from_gpu(GpuSparseBlockOuter()(*inputs))]
elif isinstance(node.op, GpuFromHost) and \
_owner_isinstance(node.inputs[0], SparseBlockOuter):
......@@ -2698,7 +2687,7 @@ def gpu_sparse_block_outer_opt(node):
meta_node = node.inputs[0].owner
inputs = _clear_host_from_gpu(meta_node.inputs)
return [GpuSparseBlockOuter(meta_node.op.inplace)(*inputs)]
return [GpuSparseBlockOuter()(*inputs)]
@local_optimizer([GpuSparseBlockGemv], inplace=True)
......
theano/tensor/basic.py
浏览文件 @ 8f0b0888
......@@ -3517,7 +3517,7 @@ def transpose(x, axes=None):
"""
if axes is None:
axes = list(range((x.ndim - 1), -1, -1))
ret = DimShuffle(x.broadcastable, axes, inplace=False)(x)
ret = DimShuffle(x.broadcastable, axes)(x)
if x.name and axes == list(range((x.ndim - 1), -1, -1)):
ret.name = x.name + '.T'
return ret
......
theano/tensor/elemwise.py
浏览文件 @ 8f0b0888
......@@ -73,8 +73,7 @@ class DimShuffle(Op):
list can either be an index or 'x'. Indices must be encoded
as python integers, not theano symbolic integers.
inplace : bool, optional
If True, the output will be a view of the input.
If False (default), the output will be a copy of the input.
If True (default), the output will be a view of the input.
Note
----
......@@ -134,13 +133,17 @@ class DimShuffle(Op):
_f16_ok = True
check_input = False
__props__ = ("input_broadcastable", "new_order", "inplace")
def __init__(self, input_broadcastable, new_order, inplace=False):
def __init__(self, input_broadcastable, new_order, inplace=True):
input_broadcastable = tuple(input_broadcastable)
self.input_broadcastable = input_broadcastable
new_order = tuple(new_order)
self.new_order = new_order
self.inplace = inplace
if inplace is True:
self.inplace = inplace
else:
raise ValueError("DimShuffle is inplace by default and hence the inplace for DimShuffle must be true")
for i, j in enumerate(new_order):
if j != 'x':
......@@ -186,17 +189,6 @@ class DimShuffle(Op):
if self.inplace:
self.view_map = {0: [0]}
self._rehash()
def __getstate__(self):
d = dict(self.__dict__)
del d['_hashval']
return d
def __setstate__(self, d):
self.__dict__.update(d)
self._rehash()
def make_node(self, _input):
input = as_tensor_variable(_input)
ib = tuple(input.type.broadcastable)
......@@ -227,23 +219,6 @@ class DimShuffle(Op):
return Apply(self, [input], [output])
def __eq__(self, other):
# it's probably not necessary to compare input_broadcastable
return type(self) == type(other) \
and self.inplace == other.inplace \
and self.new_order == other.new_order \
and self.input_broadcastable == other.input_broadcastable
def _rehash(self):
self._hashval = (hash(type(self).__name__) ^
hash(type(self).__module__) ^
hash(self.inplace) ^
hash(self.new_order) ^
hash(self.input_broadcastable))
def __hash__(self):
return self._hashval
def __str__(self):
if self.inplace:
return "InplaceDimShuffle{%s}" % ",".join(str(x)
......@@ -564,8 +539,7 @@ second dimension
# TODO: use LComplete instead
args.append(dim_shuffle(
input.type.broadcastable,
['x'] * difference + list(range(length)),
inplace=False)(input))
['x'] * difference + list(range(length)))(input))
inputs = args
# HERE: all the broadcast dims have the same length now
......@@ -798,7 +772,8 @@ second dimension
# dimensions
res = theano.tensor.constant(numpy.asarray(r.data),
dtype=r.type.dtype)
return DimShuffle((), ['x'] * nd, inplace=False)(res)
return DimShuffle((), ['x'] * nd)(res)
new_r = Elemwise(node.op, {})(
*[transform(ipt) for ipt in node.inputs])
return new_r
......
theano/tensor/opt.py
浏览文件 @ 8f0b0888
......@@ -600,8 +600,7 @@ def local_dimshuffle_lift(node):
new_inputs = []
for inp in inode.inputs:
new_inp = op.__class__(inp.type.broadcastable,
op.new_order,
op.inplace)(inp)
op.new_order)(inp)
new_inputs.append(apply_local_dimshuffle_lift(new_inp))
copy_stack_trace(node.outputs[0], new_inputs)
ret = inode.op(*new_inputs, **dict(return_list=True))
......@@ -609,14 +608,12 @@ def local_dimshuffle_lift(node):
if inode and isinstance(inode.op, DimShuffle):
new_order = [x == 'x' and 'x' or inode.op.new_order[x] for x in
new_order]
inplace = op.inplace and inode.op.inplace
input = inode.inputs[0]
if is_dimshuffle_useless(new_order, input):
return [input]
elif inode and isinstance(inode.op, DimShuffle):
ret = op.__class__(input.type.broadcastable, new_order,
inplace)(input)
ret = op.__class__(input.type.broadcastable, new_order)(input)
ret = apply_local_dimshuffle_lift(ret)
copy_stack_trace(node.outputs[0], ret)
return [ret]
......@@ -659,7 +656,7 @@ def local_useless_dimshuffle_in_reshape(node):
@register_canonicalize
@gof.local_optimizer([T.DimShuffle])
@gof.local_optimizer([DimShuffle])
def local_lift_transpose_through_dot(node):
"""
dot(x,y).T -> dot(y.T, x.T)
......@@ -688,40 +685,14 @@ def local_lift_transpose_through_dot(node):
copy_stack_trace(node.inputs[0], ret)
return ret
@gof.local_optimizer([DimShuffle])
def dimshuffle_as_view(node):
op = node.op
if not isinstance(op, DimShuffle) or op.inplace:
return False
new_op = op.__class__(op.input_broadcastable, op.new_order, inplace=True)
v = new_op(*node.inputs)
copy_stack_trace(node.outputs[0], v)
return [v]
# Step 60 is the inplace optimization stage.
compile.optdb.register('dimshuffle_as_view',
TopoOptimizer(
dimshuffle_as_view,
failure_callback=TopoOptimizer.warn_inplace),
60,
'fast_run', 'inplace')
register_canonicalize(local_dimshuffle_lift)
register_specialize(local_dimshuffle_lift)
@register_canonicalize
@gof.local_optimizer([T.DimShuffle])
def local_dimshuffle_no_inplace_at_canonicalize(node):
if isinstance(node.op, T.DimShuffle) and node.op.inplace:
return [T.DimShuffle(node.op.input_broadcastable,
node.op.new_order, inplace=False)(node.inputs[0])]
######################
# Casting operations #
######################
@register_canonicalize
@register_specialize
@gof.local_optimizer([T.TensorFromScalar])
......
theano/tensor/tests/test_opt.py
浏览文件 @ 8f0b0888
......@@ -117,7 +117,7 @@ class test_dimshuffle_lift(unittest.TestCase):
x, y, z = inputs()
e = ds(ds(x, (1, 0)), (1, 0))
g = FunctionGraph([x], [e])
self.assertTrue(str(g) == "[DimShuffle{1,0}(DimShuffle{1,0}(x))]")
self.assertTrue(str(g) == "[InplaceDimShuffle{1,0}(InplaceDimShuffle{1,0}(x))]")
dimshuffle_lift.optimize(g)
self.assertTrue(str(g) == "[x]")
# no need to check_stack_trace as graph is supposed to be empty
......@@ -126,11 +126,10 @@ class test_dimshuffle_lift(unittest.TestCase):
x, y, z = inputs()
e = ds(ds(x, (1, 'x', 0)), (2, 0, 'x', 1))
g = FunctionGraph([x], [e])
self.assertTrue(
str(g) == "[DimShuffle{2,0,x,1}(DimShuffle{1,x,0}(x))]",
str(g))
self.assertTrue(str(g) == "[InplaceDimShuffle{2,0,x,1}(InplaceDimShuffle{1,x,0}(x))]",
str(g))
dimshuffle_lift.optimize(g)
self.assertTrue(str(g) == "[DimShuffle{0,1,x,x}(x)]", str(g))
self.assertTrue(str(g) == "[InplaceDimShuffle{0,1,x,x}(x)]", str(g))
# Check stacktrace was copied over correctly after opt was applied
self.assertTrue(check_stack_trace(g, ops_to_check='all'))
......@@ -138,10 +137,9 @@ class test_dimshuffle_lift(unittest.TestCase):
x, y, z = inputs()
e = ds(ds(ds(x, (0, 'x', 1)), (2, 0, 'x', 1)), (1, 0))
g = FunctionGraph([x], [e])
self.assertTrue(
str(g) == "[DimShuffle{1,0}(DimShuffle{2,0,x,1}"
"(DimShuffle{0,x,1}(x)))]",
str(g))
self.assertTrue(str(g) == "[InplaceDimShuffle{1,0}(InplaceDimShuffle{2,0,x,1}"
"(InplaceDimShuffle{0,x,1}(x)))]",
str(g))
dimshuffle_lift.optimize(g)
self.assertTrue(str(g) == "[x]", str(g))
# no need to check_stack_trace as graph is supposed to be empty
......@@ -179,24 +177,22 @@ class test_dimshuffle_lift(unittest.TestCase):
m = T.matrix(dtype="float64")
out = ((v + 42) * (m + 84)).T
g = FunctionGraph([v, m], [out])
init_str_g = ("[DimShuffle{1,0}(Elemwise{mul,no_inplace}"
"(DimShuffle{x,0}(Elemwise{add,no_inplace}"
init_str_g = ("[InplaceDimShuffle{1,0}(Elemwise{mul,no_inplace}"
"(InplaceDimShuffle{x,0}(Elemwise{add,no_inplace}"
"(<TensorType(float64, vector)>, "
"DimShuffle{x}(TensorConstant{42}))), "
"InplaceDimShuffle{x}(TensorConstant{42}))), "
"Elemwise{add,no_inplace}"
"(<TensorType(float64, matrix)>, "
"DimShuffle{x,x}(TensorConstant{84}))))]")
"InplaceDimShuffle{x,x}(TensorConstant{84}))))]")
self.assertTrue(str(g) == init_str_g)
new_out = local_dimshuffle_lift.transform(g.outputs[0].owner)[0]
new_g = FunctionGraph(g.inputs, [new_out])
opt_str_g = ("[Elemwise{mul,no_inplace}(Elemwise{add,no_inplace}"
"(DimShuffle{0,x}(<TensorType(float64, vector)>), "
"DimShuffle{x,x}(TensorConstant{42})), "
"Elemwise{add,no_inplace}(DimShuffle{1,0}"
"(InplaceDimShuffle{0,x}(<TensorType(float64, vector)>), "
"InplaceDimShuffle{x,x}(TensorConstant{42})), "
"Elemwise{add,no_inplace}(InplaceDimShuffle{1,0}"
"(<TensorType(float64, matrix)>), "
"DimShuffle{x,x}(TensorConstant{84})))]")
"InplaceDimShuffle{x,x}(TensorConstant{84})))]")
self.assertTrue(str(new_g) == opt_str_g)
# Check stacktrace was copied over correctly after opt was applied
self.assertTrue(check_stack_trace(new_g, ops_to_check='all'))
......@@ -205,7 +201,7 @@ class test_dimshuffle_lift(unittest.TestCase):
x, _, _ = inputs()
e = ds(x, (0, 1))
g = FunctionGraph([x], [e])
self.assertTrue(str(g) == "[DimShuffle{0,1}(x)]")
self.assertTrue(str(g) == "[InplaceDimShuffle{0,1}(x)]")
dimshuffle_lift.optimize(g)
self.assertTrue(str(g) == "[x]")
# Check stacktrace was copied over correctly after opt was applied
......@@ -219,9 +215,9 @@ class test_dimshuffle_lift(unittest.TestCase):
ds_z = ds(z, (2, 1, 0)) # usefull
ds_u = ds(u, ('x')) # usefull
g = FunctionGraph([x, y, z, u], [ds_x, ds_y, ds_z, ds_u])
self.assertTrue(str(g) == "[DimShuffle{0,x}(x), DimShuffle{2,1,0}(y), DimShuffle{2,1,0}(z), DimShuffle{x}(TensorConstant{1})]")
self.assertTrue(str(g) == "[InplaceDimShuffle{0,x}(x), InplaceDimShuffle{2,1,0}(y), InplaceDimShuffle{2,1,0}(z), InplaceDimShuffle{x}(TensorConstant{1})]")
dimshuffle_lift.optimize(g)
self.assertTrue(str(g) == "[x, y, DimShuffle{2,1,0}(z), DimShuffle{x}(TensorConstant{1})]")
self.assertTrue(str(g) == "[x, y, InplaceDimShuffle{2,1,0}(z), InplaceDimShuffle{x}(TensorConstant{1})]")
# Check stacktrace was copied over correctly after opt was applied
self.assertTrue(hasattr(g.outputs[0].tag, 'trace'))
......@@ -241,10 +237,11 @@ def test_local_useless_dimshuffle_in_reshape():
[reshape_dimshuffle_vector, reshape_dimshuffle_mat,
reshape_dimshuffle_row, reshape_dimshuffle_col])
assert_true(str(g) == "[Reshape{1}(DimShuffle{x,0}(vector), Shape(vector)), "
"Reshape{2}(DimShuffle{x,0,x,1}(mat), Shape(mat)), "
"Reshape{2}(DimShuffle{1,x}(row), Shape(row)), "
"Reshape{2}(DimShuffle{0}(col), Shape(col))]")
print(str(g))
assert_true(str(g) == "[Reshape{1}(InplaceDimShuffle{x,0}(vector), Shape(vector)), "
"Reshape{2}(InplaceDimShuffle{x,0,x,1}(mat), Shape(mat)), "
"Reshape{2}(InplaceDimShuffle{1,x}(row), Shape(row)), "
"Reshape{2}(InplaceDimShuffle{0}(col), Shape(col))]")
useless_dimshuffle_in_reshape = out2in(local_useless_dimshuffle_in_reshape)
useless_dimshuffle_in_reshape.optimize(g)
assert_true(str(g) == "[Reshape{1}(vector, Shape(vector)), "
......@@ -3766,15 +3763,15 @@ class Test_local_canonicalize_alloc(unittest.TestCase):
"TensorConstant{2})]"))
alloc_lift.optimize(g)
self.assertTrue(str(g) == "[DimShuffle{x,0,1}"
self.assertTrue(str(g) == "[InplaceDimShuffle{x,0,1}"
"(Alloc(<TensorType(float64, vector)>, "
"TensorConstant{3}, "
"TensorConstant{2})), "
"DimShuffle{x,x}"
"InplaceDimShuffle{x,x}"
"(<TensorType(float64, scalar)>), "
"DimShuffle{x,0,1}"
"InplaceDimShuffle{x,0,1}"
"(Alloc(<TensorType(float64, matrix)>, "
"TensorConstant{1}, "
"TensorConstant{2})), "
......@@ -6268,9 +6265,9 @@ class Test_local_reshape_to_dimshuffle(unittest.TestCase):
reshape_lift.optimize(g)
useless_reshape.optimize(g)
self.assertTrue(str(g) == "[DimShuffle{x,0}"
self.assertTrue(str(g) == "[InplaceDimShuffle{x,0}"
"(<TensorType(float64, vector)>), "
"DimShuffle{x,0,x,1,x,x}"
"InplaceDimShuffle{x,0,x,1,x,x}"
"(Reshape{2}(<TensorType(float64, matrix)>, "
"TensorConstant{[5 6]}))]")
......@@ -6301,7 +6298,7 @@ class Test_lift_transpose_through_dot(unittest.TestCase):
def test_matrix_matrix(self):
a, b = matrices('ab')
g = self.simple_optimize(FunctionGraph([a, b], [tensor.dot(a, b).T]))
sg = '[dot(DimShuffle{1,0}(b), DimShuffle{1,0}(a))]'
sg = '[dot(InplaceDimShuffle{1,0}(b), InplaceDimShuffle{1,0}(a))]'
assert str(g) == sg, (str(g), sg)
# Check stacktrace was copied over correctly after opt was applied
self.assertTrue(check_stack_trace(g, ops_to_check='all'))
......@@ -6313,7 +6310,7 @@ class Test_lift_transpose_through_dot(unittest.TestCase):
[a, b],
[tensor.dot(a.dimshuffle('x', 0), b).T]),
level='stabilize')
sg = '[dot(DimShuffle{1,0}(b), DimShuffle{0,x}(a))]'
sg = '[dot(InplaceDimShuffle{1,0}(b), InplaceDimShuffle{0,x}(a))]'
assert str(g) == sg, (str(g), sg)
# Check stacktrace was copied over correctly after opt was applied
self.assertTrue(check_stack_trace(g, ops_to_check='all'))
......@@ -6325,7 +6322,7 @@ class Test_lift_transpose_through_dot(unittest.TestCase):
[a, b],
[tensor.dot(b, a.dimshuffle(0, 'x')).T]),
level='stabilize')
sg = '[dot(DimShuffle{x,0}(a), DimShuffle{1,0}(b))]'
sg = '[dot(InplaceDimShuffle{x,0}(a), InplaceDimShuffle{1,0}(b))]'
assert str(g) == sg, (str(g), sg)
# Check stacktrace was copied over correctly after opt was applied
self.assertTrue(check_stack_trace(g, ops_to_check='all'))
......
theano/tests/test_printing.py
浏览文件 @ 8f0b0888
......@@ -321,10 +321,10 @@ def test_scan_debugprint1():
| | | | | |Subtensor{int64} [id H] ''
| | | | | |Shape [id I] ''
| | | | | | |Rebroadcast{0} [id J] ''
| | | | | | |DimShuffle{x,0} [id K] ''
| | | | | | |InplaceDimShuffle{x,0} [id K] ''
| | | | | | |Elemwise{second,no_inplace} [id L] ''
| | | | | | |A [id M]
| | | | | | |DimShuffle{x} [id N] ''
| | | | | | |InplaceDimShuffle{x} [id N] ''
| | | | | | |TensorConstant{1.0} [id O]
| | | | | |Constant{0} [id P]
| | | | |Subtensor{int64} [id Q] ''
......@@ -490,7 +490,7 @@ def test_scan_debugprint3():
for{cpu,scan_fn} [id B] ''
>Elemwise{mul,no_inplace} [id Y] ''
> |DimShuffle{x} [id Z] ''
> |InplaceDimShuffle{x} [id Z] ''
> | |coefficients[t] [id BA] -> [id S]
> |Elemwise{pow,no_inplace} [id BB] ''
> |Subtensor{int64} [id BC] ''
......@@ -504,10 +504,10 @@ def test_scan_debugprint3():
> | | | | | | |Subtensor{int64} [id BJ] ''
> | | | | | | |Shape [id BK] ''
> | | | | | | | |Rebroadcast{0} [id BL] ''
> | | | | | | | |DimShuffle{x,0} [id BM] ''
> | | | | | | | |InplaceDimShuffle{x,0} [id BM] ''
> | | | | | | | |Elemwise{second,no_inplace} [id BN] ''
> | | | | | | | |A_copy [id BO] -> [id W]
> | | | | | | | |DimShuffle{x} [id BP] ''
> | | | | | | | |InplaceDimShuffle{x} [id BP] ''
> | | | | | | | |TensorConstant{1.0} [id BQ]
> | | | | | | |Constant{0} [id BR]
> | | | | | |Subtensor{int64} [id BS] ''
......@@ -520,7 +520,7 @@ def test_scan_debugprint3():
> | | | |A_copy [id BO] -> [id W]
> | | |Constant{1} [id BW]
> | |Constant{-1} [id BX]
> |DimShuffle{x} [id BY] ''
> |InplaceDimShuffle{x} [id BY] ''
> |<TensorType(int64, scalar)> [id BZ] -> [id U]
for{cpu,scan_fn} [id BE] ''
......@@ -636,10 +636,10 @@ def test_scan_debugprint5():
| | | | | | | |Subtensor{int64} [id K] ''
| | | | | | | |Shape [id L] ''
| | | | | | | | |Rebroadcast{0} [id M] ''
| | | | | | | | |DimShuffle{x,0} [id N] ''
| | | | | | | | |InplaceDimShuffle{x,0} [id N] ''
| | | | | | | | |Elemwise{second,no_inplace} [id O] ''
| | | | | | | | |A [id P]
| | | | | | | | |DimShuffle{x} [id Q] ''
| | | | | | | | |InplaceDimShuffle{x} [id Q] ''
| | | | | | | | |TensorConstant{1.0} [id R]
| | | | | | | |Constant{0} [id S]
| | | | | | |Subtensor{int64} [id T] ''
......@@ -675,20 +675,20 @@ def test_scan_debugprint5():
| | | | | |k [id G]
| | | | | |IncSubtensor{Set;:int64:} [id H] ''
| | | | | |A [id P]
| | | | |DimShuffle{x,x} [id BP] ''
| | | | |InplaceDimShuffle{x,x} [id BP] ''
| | | | |TensorConstant{0.0} [id BQ]
| | | |IncSubtensor{Inc;int64} [id BR] ''
| | | | |Elemwise{second,no_inplace} [id BS] ''
| | | | | |Subtensor{int64::} [id BT] ''
| | | | | | |for{cpu,scan_fn} [id BO] ''
| | | | | | |Constant{1} [id BU]
| | | | | |DimShuffle{x,x} [id BV] ''
| | | | | |InplaceDimShuffle{x,x} [id BV] ''
| | | | | |TensorConstant{0.0} [id BQ]
| | | | |Elemwise{second} [id BW] ''
| | | | | |Subtensor{int64} [id BX] ''
| | | | | | |Subtensor{int64::} [id BT] ''
| | | | | | |Constant{-1} [id BY]
| | | | | |DimShuffle{x} [id BZ] ''
| | | | | |InplaceDimShuffle{x} [id BZ] ''
| | | | | |Elemwise{second,no_inplace} [id CA] ''
| | | | | |Sum{acc_dtype=float64} [id CB] ''
| | | | | | |Subtensor{int64} [id BX] ''
......
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