Skip to content

P
pytensor
提交 9ad99d0f authored 作者: Laurent Dinh's avatar Laurent Dinh
浏览文件
操作

Merge branch 'master' of git://github.com/Theano/Theano into scan

上级 ff371509 99a994cb
全部展开 隐藏空白字符变更
内嵌 并排
正在显示 包含 449 行增加 和 246 行删除
doc/library/sandbox/index.txt
浏览文件 @ 9ad99d0f
......@@ -16,6 +16,4 @@
cuda/index
linalg
neighbours
rng_mrg
doc/library/sandbox/rng_mrg.txt 0 → 100644
浏览文件 @ 9ad99d0f
.. _libdoc_rng_mrg:
===================================================================
:mod:`sandbox.rng_mrg` -- MRG random number generator
===================================================================
.. module:: sandbox.rng_mrg
:platform: Unix, Windows
:synopsis: MRG random number generator
.. moduleauthor:: LISA
API
===
.. automodule:: theano.sandbox.rng_mrg
:members:
doc/library/sparse/index.txt
浏览文件 @ 9ad99d0f
......@@ -11,20 +11,20 @@ In the tutorial section, you can find a :ref:`sparse tutorial
The sparse submodule is not loaded when we import Theano. You must
import ``theano.sparse`` to enable it.
The sparse module provide the same functionalities as the tensor
module. The difference lies under the cover because sparse matrices
does not store data in a contiguous array. Note that there are no GPU
implementations for sparse matrices implemented in Theano. The sparse
module has been used in:
The sparse module provides the same functionality as the tensor
module. The difference lies under the covers because sparse matrices
do not store data in a contiguous array. Note that there are no GPU
implementations for sparse matrices in Theano. The sparse module has
been used in:
- NLP: Dense linear transformations of sparse vectors.
- Audio: Filterbank in Fourier domain.
- Audio: Filterbank in the Fourier domain.
Compressed Sparse Format
========================
This section tries to explain how information is store for the two
sparse formats of SciPy supported by Theano. There is more formats
This section tries to explain how information is stored for the two
sparse formats of SciPy supported by Theano. There are more formats
that can be used with SciPy and some documentation about them may be
found `here
<http://deeplearning.net/software/theano/sandbox/sparse.html>`_.
......@@ -50,14 +50,14 @@ attributes: ``data``, ``indices``, ``indptr`` and ``shape``.
CSC Matrix
----------
In the *Compressed Sparse Column* format, ``indices`` stands for index
inside the column vectors of the matrix and ``indptr`` tells where the
column starts in the ``data`` and in the ``indices``
attributes. ``indptr`` can be tought as giving the slice which must be
applied to the other attribute in order to get each column of the
matrix. In other words, ``slice(indptr[i], indptr[i+1])`` correspond
to the slice needed to find the i-th column of the matrix in the
``data`` and in the ``indices`` fields.
In the *Compressed Sparse Column* format, ``indices`` stands for
indexes inside the column vectors of the matrix and ``indptr`` tells
where the column starts in the ``data`` and in the ``indices``
attributes. ``indptr`` can be thought of as giving the slice which
must be applied to the other attribute in order to get each column of
the matrix. In other words, ``slice(indptr[i], indptr[i+1])``
corresponds to the slice needed to find the i-th column of the matrix
in the ``data`` and ``indices`` fields.
The following example builds a matrix and returns its columns. It
prints the i-th column, i.e. a list of indices in the column and their
......@@ -84,18 +84,18 @@ corresponding value in the second list.
CSR Matrix
----------
In the *Compressed Sparse Row* format, ``indices`` stands for index
In the *Compressed Sparse Row* format, ``indices`` stands for indexes
inside the row vectors of the matrix and ``indptr`` tells where the
row starts in the ``data`` and in the ``indices``
attributes. ``indptr`` can be tought as giving the slice which must be
applied to the other attribute in order to get each row of the
matrix. In other words, ``slice(indptr[i], indptr[i+1])`` correspond
attributes. ``indptr`` can be thought of as giving the slice which
must be applied to the other attribute in order to get each row of the
matrix. In other words, ``slice(indptr[i], indptr[i+1])`` corresponds
to the slice needed to find the i-th row of the matrix in the ``data``
and in the ``indices`` fields.
and ``indices`` fields.
The following example builds a matrix and returns its rows. It prints
the i-th row, i.e. a list of indices in the row and their corresponding value
in the second list.
the i-th row, i.e. a list of indices in the row and their
corresponding value in the second list.
>>> data = np.asarray([7, 8, 9])
>>> indices = np.asarray([0, 1, 2])
......@@ -120,7 +120,7 @@ List of Implemented Operations
- Moving from and to sparse
- :class:`DenseFromSparse <theano.sparse.basic.DenseFromSparse>` and ``dense_from_sparse``.
Both grad are implemented. Structured by default.
Both grads are implemented. Structured by default.
- :class:`SparseFromDense <theano.sparse.basic.SparseFromDense>` and ``csr_from_dense``, ``csc_from_dense``.
The grad implemented is structured.
- Theano SparseVariable object have a method ``toarray()`` that is the same as ``dense_from_sparse``.
......@@ -201,51 +201,55 @@ List of Implemented Operations
- One of the inputs must be sparse, the other sparse or dense.
- The grad implemented is regular.
- No C code for perform and no C code for grad.
- Return a dense for perform and a dense for grad.
- Returns a dense for perform and a dense for grad.
- :class:`StructuredDot <theano.sparse.basic.StructuredDot>`
and :func:`structured_dot <theano.sparse.basic.structured_dot>`.
- The first input is sparse, the second can be sparse or dense.
- The grad implemented is structured.
- C code for perform and grad.
- Return a dense for perforn and a sparse for grad.
- It returns a sparse output if both inputs are sparse and
dense one if one of the inputs is dense.
- Returns a sparse grad for sparse inputs and dense grad for
dense inputs.
- :class:`TrueDot <theano.sparse.basic.TrueDot>` and
:func:`true_dot <theano.sparse.basic.true_dot>`.
- The first input is sparse, the second can be sparse or dense.
- The grad implemented is regular.
- No C code for perform and no C code for grad.
- Return a Sparse for perform and a Sparse for grad.
- Flags trough constructor can change the output of
grad to be dense if the second input of the op is dense.
- Returns a Sparse.
- The gradient returns a Sparse for sparse inputs and by
default a dense for dense inputs. The parameter
``grad_preserves_dense`` can be set to False to return a
sparse grad for dense inputs.
- :class:`SamplingDot <theano.sparse.basic.SamplingDot>` and
``sampling_dot``.
- Both input must be dense.
- Both inputs must be dense.
- The grad implemented is structured for `p`.
- Sample of the dot and sample of the gradient.
- C code for perform but not for grad.
- Return sparse for perform and grad.
- Returns sparse for perform and grad.
- :class:`Usmm <theano.sparse.basic.Usmm>` and ``usmm``.
- You *shouldn't* insert this op yourself!
- There is optimization that transform a
- There is an optimization that transform a
:class:`Dot <theano.sparse.basic.Dot>` to ``Usmm`` when possible.
- This op is the equivalent of gemm for sparse dot.
- There is no grad implemented for this op and this is not needed as
you don't insert it yourself.
- There is no grad implemented for this op.
- One of the inputs must be sparse, the other sparse or dense.
- Return a dense for perform
- Returns a dense from perform.
- Slice Operations
- sparse_variable[N, N], return a tensor scalar.
- sparse_variable[N, N], returns a tensor scalar.
There is no grad implemented for this operation.
- sparse_variable[M:N, O:P], return a sparse matrix
- sparse_variable[M:N, O:P], returns a sparse matrix
There is no grad implemented for this operation.
- Sparse variable don't support [M, N:O] and [M:N, O] as we don't support sparse vector
and returning a sparse matrix would break the numpy interface.
Use [M:M+1, N:O] and [M:N, O:O+1] instead.
- Sparse variables don't support [M, N:O] and [M:N, O] as we don't
support sparse vectors and returning a sparse matrix would break
the numpy interface. Use [M:M+1, N:O] and [M:N, O:O+1] instead.
- :class:`Diag <theano.sparse.basic.Diag>` and ``diag``.
The grad implemented is regular.
......
doc/library/tensor/basic.txt
浏览文件 @ 9ad99d0f
......@@ -1452,7 +1452,7 @@ Linear Algebra
print(b.shape) #(5,6,4,3)
print(c.shape) #(2,3,4,5,6,4,3)
See the documentation of numpy.tensordot for more examples.
:note: See the documentation of `numpy.tensordot <http://docs.scipy.org/doc/numpy/reference/generated/numpy.tensordot.html>`_ for more examples.
.. function:: batched_dot(X, Y)
......@@ -1478,6 +1478,40 @@ Linear Algebra
:return: tensor of products
.. function:: batched_tensordot(X, Y, axes=2)
:param x: A Tensor with sizes e.g.: for 3D (dim1, dim3, dim2)
:param y: A Tensor with sizes e.g.: for 3D (dim1, dim2, dim4)
:param axes: an integer or array. If an integer, the number of axes
to sum over. If an array, it must have two array
elements containing the axes to sum over in each tensor.
If an integer i, it is converted to an array containing
the last i dimensions of the first tensor and the first
i dimensions of the second tensor (excluding the first
(batch) dimension):
axes = [range(a.ndim - i, b.ndim), range(1,i+1)]
If an array, its two elements must contain compatible axes
of the two tensors. For example, [[1, 2], [2, 4]] means sum
over the 2nd and 3rd axes of a and the 3rd and 5th axes of b.
(Remember axes are zero-indexed!) The 2nd axis of a and the
3rd axis of b must have the same shape; the same is true for
the 3rd axis of a and the 5th axis of b.
:type axes: int or array-like of length 2
:returns: a tensor with shape equal to the concatenation of a's shape
(less any dimensions that were summed over) and b's shape
(less first dimension and any dimensions that were summed over).
:rtype: tensor of tensordots
A hybrid of batch_dot and tensordot, this function computes the
tensordot product between the two tensors, by iterating over the
first dimension using scan to perform a sequence of tensordots.
:note: See :func:`tensordot` and :func:`batched_dot` for
supplementary documentation.
Gradient / Differentiation
......
doc/tutorial/examples.txt
浏览文件 @ 9ad99d0f
......@@ -5,13 +5,13 @@
More Examples
=============
At this point it would be wise to begin familiarizing yourself
more systematically with Theano's fundamental objects and operations by browsing
this section of the library: :ref:`libdoc_basic_tensor`.
At this point it would be wise to begin familiarizing yourself more
systematically with Theano's fundamental objects and operations by
browsing this section of the library: :ref:`libdoc_basic_tensor`.
As the tutorial unfolds, you should also gradually acquaint yourself with the other
relevant areas of the library and with the relevant subjects of the documentation
entrance page.
As the tutorial unfolds, you should also gradually acquaint yourself
with the other relevant areas of the library and with the relevant
subjects of the documentation entrance page.
Logistic Function
......@@ -30,13 +30,13 @@ the logistic curve, which is given by:
A plot of the logistic function, with x on the x-axis and s(x) on the
y-axis.
You want to compute the function :ref:`elementwise <libdoc_tensor_elementwise>` on matrices of
doubles, which means that you want to apply this function to each
individual element of the matrix.
You want to compute the function :ref:`elementwise
<libdoc_tensor_elementwise>` on matrices of doubles, which means that
you want to apply this function to each individual element of the
matrix.
Well, what you do is this:
.. If you modify this code, also change :
.. theano/tests/test_tutorial.py:T_examples.test_examples_1
......@@ -450,6 +450,10 @@ Other Random Distributions
There are :ref:`other distributions implemented <libdoc_tensor_raw_random>`.
Other Implementations
---------------------
There is 2 other implementations based on :class:`CURAND <theano.sandbox.cuda.rng_curand>` and :ref:`MRG31k3p <libdoc_rng_mrg>`
.. _logistic_regression:
......@@ -457,7 +461,8 @@ There are :ref:`other distributions implemented <libdoc_tensor_raw_random>`.
A Real Example: Logistic Regression
===================================
The preceding elements are featured in this more realistic example. It will be used repeatedly.
The preceding elements are featured in this more realistic example.
It will be used repeatedly.
.. code-block:: python
......
doc/tutorial/multi_cores.txt
浏览文件 @ 9ad99d0f
......@@ -2,30 +2,49 @@
Multi cores support in Theano
=============================
Parallel element wise op with openmp
====================================
BLAS operation
==============
Beacuse element wise ops work on every tensor entry indipedently they can be
easly parallelized using openmp.
BLAS is an interface for some mathematic operations between two
vectors, a vector and a matrix or two matrices (e.g. the dot product
between vector/matrix and matrix/matrix). Many different
implementations of that interface exist and some of them are
parallelized.
To use openmp you must set the openmp flag in Theano configuration.
Theano tries to use that interface as frequently as possible for
performance reasons. So if Theano links to a parallel implementation,
those operations will run in parallel in Theano.
Yuo can use the flag openmp_elemwise_minsize to set the minimum tensor size
for which the operation is parallelized because for short tensor using opemp
can slow down the operation.
The most frequent way to control the number of threads used is via the
``OMP_NUM_THREADS`` environment variable. Set it to the number of
threads you want to use before starting the python process. Some BLAS
implementations support other enviroment variables.
If it is no specified the default value (200000) is used.
For simple(fast) operation you can obtain a speed up for very long tensor
while for more complex operation you ca obtain a good speed up also for not
too long tensor.
There is a script (elemwise_openmp_speedup.py in theano/misc/) which you can
use to choose that value for your machine.
The script run two elemwise operation (a fast and a slow one) for a vector of
size openmp_elemwise_minsize with and without openmp and show the time
difference between the two cases.
Parallel element wise ops with OpenMP
=====================================
Because element wise ops work on every tensor entry independently they
can be easily parallelized using OpenMP.
To use OpenMP you must set the ``openmp`` :ref:`flag <libdoc_config>`
to ``True``.
You can use the flag ``openmp_elemwise_minsize`` to set the minimum
tensor size for which the operation is parallelized because for short
tensors using OpenMP can slow down the operation. The default value is
``200000``.
For simple(fast) operation you can obtain a speed up with very large
tensors while for more complex operation you can obtain a good speed
up also for smaller tensor.
There is a script ``elemwise_openmp_speedup.py`` in ``theano/misc/``
which you can use to tune the value of ``openmp_elemwise_minsize`` for
your machine. The script runs two elemwise operations (a fast one and
a slow one) for a vector of size ``openmp_elemwise_minsize`` with and
without OpenMP and shows the time difference between the cases.
The only way to control the number of threads used is via the
``OMP_NUM_THREADS`` environment variable. Set it to the number of threads
you want to use before starting the python process.
theano/gof/__init__.py
浏览文件 @ 9ad99d0f
......@@ -62,7 +62,6 @@ from theano.gof.opt import (Optimizer, optimizer, SeqOptimizer,
LocalOptimizer, local_optimizer, LocalOptGroup,
OpSub, OpRemove, PatternSub,
NavigatorOptimizer, TopoOptimizer, EquilibriumOptimizer,
InplaceOptimizer, PureThenInplaceOptimizer,
OpKeyOptimizer)
from theano.gof.optdb import \
......
theano/gof/graph.py
浏览文件 @ 9ad99d0f
......@@ -415,18 +415,21 @@ class Variable(Node):
if inputs_to_values is None:
inputs_to_values = {}
if not hasattr(self, '_fn'):
self._fn_inputs = inputs_to_values.keys()
self._fn = theano.function(self._fn_inputs, self)
args = [inputs_to_values[param] for param in self._fn_inputs]
if not hasattr(self, '_fn_cache'):
self._fn_cache = dict()
rval = self._fn(*args)
inputs = tuple(sorted(inputs_to_values.keys(), key=id))
if not inputs in self._fn_cache:
self._fn_cache[inputs] = theano.function(inputs, self)
args = [inputs_to_values[param] for param in inputs]
rval = self._fn_cache[inputs](*args)
return rval
def __getstate__(self):
d = self.__dict__.copy()
d.pop("_fn", None)
d.pop("_fn_cache", None)
return d
env = property(env_getter, env_setter, env_deleter)
......
theano/gof/opt.py
浏览文件 @ 9ad99d0f
......@@ -131,6 +131,9 @@ class FromFunctionOptimizer(Optimizer):
def __call__(self, *args, **kwargs):
return self.fn(*args, **kwargs)
def __str__(self):
return self.__name__
def optimizer(f):
"""decorator for FromFunctionOptimizer"""
......@@ -626,7 +629,10 @@ class MergeOptimizer(Optimizer):
print >> stream, blanc, " replace_time", replace_time
print >> stream, blanc, " validate_time", validate_time
print >> stream, blanc, " callback_time", callback_time
print >> stream, blanc, " callback_times", callbacks_time
print >> stream, blanc, " callbacks_time"
for i in sorted(callbacks_time.iteritems(), key=lambda a: a[1]):
if i[1] > 0:
print i
print >> stream, blanc, " nb_merged", nb_merged
print >> stream, blanc, " nb_constant", nb_constant
......@@ -1490,7 +1496,6 @@ class EquilibriumOptimizer(NavigatorOptimizer):
def __init__(self,
optimizers,
failure_callback=None,
max_depth=None,
max_use_ratio=None):
"""
:param optimizers: list or set of local or global optimizations to
......@@ -1499,8 +1504,6 @@ class EquilibriumOptimizer(NavigatorOptimizer):
:param max_use_ratio: each optimizer can be applied at most
(size of graph * this number) times
:param max_depth: TODO what does this do? (EquilibriumDB sets it to 5)
"""
super(EquilibriumOptimizer, self).__init__(
......@@ -1520,7 +1523,6 @@ class EquilibriumOptimizer(NavigatorOptimizer):
self.local_optimizers_map.setdefault(c, []).append(opt)
else:
self.global_optimizers.append(opt)
self.max_depth = max_depth
self.max_use_ratio = max_use_ratio
assert self.max_use_ratio is not None, (
'max_use_ratio has to be a number')
......@@ -1723,11 +1725,13 @@ class EquilibriumOptimizer(NavigatorOptimizer):
for (t, count, opt) in count_opt[::-1]:
print >> stream, blanc, ' %.3fs - %d - %s' % (
t, count, opt)
print >> stream, blanc, ' %.3fs - in %d optimization that where not used' % (
print >> stream, blanc, ' %.3fs - in %d optimization that where not used (display only those with a runtime > 0)' % (
not_used_time, len(not_used))
not_used.sort()
for (t, opt) in not_used[::-1]:
print >> stream, blanc + " ", ' %.3fs - %s' % (t, opt)
if t > 0:
# Skip opt that have 0 times, they probably wasn't even tried.
print >> stream, blanc + " ", ' %.3fs - %s' % (t, opt)
print >> stream
@staticmethod
......@@ -1899,31 +1903,3 @@ def pre_greedy_local_optimizer(list_optimizations, out):
final_outs, optimized_nodes = local_recursive_function(
list_optimizations, out, {}, 0)
return final_outs[out_index]
############
### Misc ###
############
class InplaceOptimizer(Optimizer):
def __init__(self, inplace):
self.inplace = inplace
def apply(self, fgraph):
self.inplace(fgraph)
def add_requirements(self, fgraph):
fgraph.attach_feature(dh.DestroyHandler())
class PureThenInplaceOptimizer(Optimizer):
def __init__(self, pure, inplace):
self.pure = pure
self.inplace = inplace
def apply(self, fgraph):
self.pure(fgraph)
fgraph.attach_feature(dh.DestroyHandler())
self.inplace(fgraph)
theano/gof/optdb.py
浏览文件 @ 9ad99d0f
......@@ -194,7 +194,6 @@ class EquilibriumDB(DB):
def query(self, *tags, **kwtags):
opts = super(EquilibriumDB, self).query(*tags, **kwtags)
return opt.EquilibriumOptimizer(opts,
max_depth=5,
max_use_ratio=config.optdb.max_use_ratio,
failure_callback=opt.NavigatorOptimizer.warn_inplace)
......
theano/gof/tests/test_graph.py
浏览文件 @ 9ad99d0f
......@@ -297,14 +297,17 @@ class TestIsSameGraph(unittest.TestCase):
# eval #
################
def test_eval():
x = tensor.scalar()
y = tensor.scalar()
z = x + y
result = z.eval({x : 1., y : 2.})
assert result == 3.
# We don't want to pickle the tmp function.
assert not hasattr(pickle.loads(pickle.dumps(z)), '_fn')
class TestEval(unittest.TestCase):
def setUp(self):
self.x, self.y = tensor.scalars('x', 'y')
self.z = self.x + self.y
self.w = 2 * self.z
def test_eval(self):
self.assertEquals(self.w.eval({self.x : 1., self.y : 2.}), 6.)
self.assertEquals(self.w.eval({self.z : 3}), 6.)
self.assertTrue(hasattr(self.w, "_fn_cache"),
"variable must have cache after eval")
self.assertFalse(hasattr(pickle.loads(pickle.dumps(self.w)), '_fn_cache'),
"temporary functions must not be serialized")
theano/misc/check_blas.py
浏览文件 @ 9ad99d0f
......@@ -205,6 +205,7 @@ if __name__ == "__main__":
gpu
K20m/ECC 0.07s
K20/NOECC 0.07s
M2090 0.19s
C2075 0.25s
M2075 0.25s
M2070 0.25s 0.27s 0.32s
......
theano/sandbox/cuda/blas.py
浏览文件 @ 9ad99d0f
......@@ -671,7 +671,7 @@ class GpuConv(GpuOp):
def c_code_cache_version(self):
# raise this whenever modifying any of the support_code_files
return (0, 20)
return (0, 21)
def c_support_code_apply(self, node, nodename):
# REMEMBER TO RAISE c_code_cache_version when changing any of
......
theano/sandbox/cuda/conv.cu
浏览文件 @ 9ad99d0f
......@@ -1018,6 +1018,7 @@ CudaNdarray_conv_full(const CudaNdarray *img, const CudaNdarray * kern,
(version==3||version==4||version==5||version==-1) &&
out_wid<=max_threads_dim0 &&//Maximum of X threads by block.x
(kern_len+2*kern_len-2)*img_wid_padded*sizeof(float) + kern_size_byte<shared_avail && //their is only 16k of shared memory
(kern_len > 1 || (img_size_padded_byte+kern_size_byte)<=shared_avail) &&
!work_complete) //conv_full_patch_stack_padded
{
//version 3 without split
......
theano/sandbox/cuda/opt.py
浏览文件 @ 9ad99d0f
......@@ -14,7 +14,7 @@ import theano.ifelse
from theano.compile import optdb
from theano.gof import (local_optimizer, EquilibriumDB, SequenceDB, ProxyDB,
Optimizer, toolbox, DestroyHandler)
Optimizer, toolbox)
from theano.gof.python25 import all, any
from theano.sandbox.cuda.basic_ops import (
device_properties, gpu_eye,
......@@ -62,7 +62,7 @@ optdb.register('gpu_opt',
# inside the elemwise. When there is no float64 op, this is working.
optdb.register('gpu_after_fusion',
ProxyDB(gpu_seqopt),
optdb.__position__.get('elemwise_fusion', 71) + .1,
optdb.__position__.get('elemwise_fusion', 49) + .1,
'gpu')
......@@ -88,7 +88,6 @@ class InputToGpuOptimizer(Optimizer):
def add_requirements(self, fgraph):
fgraph.attach_feature(toolbox.ReplaceValidate())
fgraph.attach_feature(DestroyHandler())
def apply(self, fgraph):
for input in fgraph.inputs:
......@@ -1339,9 +1338,10 @@ gpu_local_elemwise_fusion = tensor.opt.local_elemwise_fusion_op(
max_inputs_to_GpuElemwise)
if config.gpu.local_elemwise_fusion:
_logger.debug("enabling optimization fusion of gpu elemwise in fast_run")
#Must be after cpu fusion at 40, gpu at 48.5 and before AddDestroyHandler at 49.5
optdb.register('gpu_elemwise_fusion',
tensor.opt.FusionOptimizer(gpu_local_elemwise_fusion),
71.00, 'fast_run', 'fusion',
49, 'fast_run', 'fusion',
'local_elemwise_fusion', 'gpu')
else:
_logger.debug(("not enabling optimization fusion of gpu elemwise in "
......
theano/sandbox/cuda/tests/test_conv_cuda_ndarray.py
浏览文件 @ 9ad99d0f
......@@ -679,6 +679,7 @@ def test_full():
#Test more than maxThreadsDim0
, ((2,4,13,1050), (3,4,10, 11), (1, 1), (1, 1), (1, 1))
, ((2,4,1050,13), (3,4,10, 11), (1, 1), (1, 1), (1, 1))
, ((1,1,44800,1), (6,1,1,1), (1, 1), (1, 1), (1, 1))#This caused crash
]
# shapes=shapes[:277]
......
theano/sandbox/cuda/tests/test_type.py
浏览文件 @ 9ad99d0f
import cPickle
import os.path
import sys
from nose.tools import assert_raises
import numpy
import os.path
from theano import config
from theano.sandbox.cuda import cuda_available
......@@ -12,10 +15,13 @@ if cuda_available:
# >>> with open('CudaNdarray.pkl', 'wb') as fp:
# >>> cPickle.dump(theano.sandbox.cuda.CudaNdarray(np.array([-42.0], dtype=np.float32)), fp)
def test_unpickle_flag_is_false_by_default():
assert not config.experimental.unpickle_gpu_on_cpu, "Config flag experimental.unpickle_gpu_on_cpu is " \
+ "set to true. Make sure the default value stays false " \
+ "and that you have not set the flag manually."
assert not config.experimental.unpickle_gpu_on_cpu, (
"Config flag experimental.unpickle_gpu_on_cpu is "
"set to true. Make sure the default value stays false "
"and that you have not set the flag manually.")
def test_unpickle_cudandarray_as_numpy_ndarray_flag0():
oldflag = config.experimental.unpickle_gpu_on_cpu
......@@ -23,7 +29,11 @@ def test_unpickle_cudandarray_as_numpy_ndarray_flag0():
try:
testfile_dir = os.path.dirname(os.path.realpath(__file__))
with open(os.path.join(testfile_dir, 'CudaNdarray.pkl')) as fp:
fname = 'CudaNdarray.pkl'
if sys.version_info[0] == 3:
fname = 'CudaNdarray_py3.pkl'
with open(os.path.join(testfile_dir, fname), 'rb') as fp:
if cuda_available:
mat = cPickle.load(fp)
else:
......
theano/sandbox/gpuarray/blas.py
浏览文件 @ 9ad99d0f
......@@ -61,7 +61,7 @@ class GpuGemv(BlasOp, Gemv):
((dtype_%(alpha)s *)PyArray_DATA(%(alpha)s))[0],
%(A)s, %(x)s,
((dtype_%(beta)s *)PyArray_DATA(%(beta)s))[0],
%(out)s) == NULL) {
%(out)s, 0) == -1) {
%(fail)s
}
""" % vars
......@@ -72,7 +72,7 @@ class GpuGemv(BlasOp, Gemv):
return code
def c_code_cache_version(self):
return (0,)
return (1,)
gpugemv_no_inplace = GpuGemv(inplace=False)
gpugemv_inplace = GpuGemv(inplace=True)
......@@ -117,7 +117,7 @@ class GpuGemm(BlasOp, Gemm):
((dtype_%(alpha)s *)PyArray_DATA(%(alpha)s))[0],
%(A)s, %(B)s,
((dtype_%(beta)s *)PyArray_DATA(%(beta)s))[0],
%(out)s) == NULL) {
%(out)s, 0) == -1) {
%(fail)s
}
""" % vars
......@@ -128,7 +128,7 @@ class GpuGemm(BlasOp, Gemm):
return code
def c_code_cache_version(self):
return (0,)
return (1,)
gpugemm_no_inplace = GpuGemm(inplace=False)
......@@ -176,7 +176,7 @@ class GpuDot22(BlasOp, Dot22):
one,
%(A)s, %(B)s,
zero,
%(out)s) == NULL) {
%(out)s, 0) == -1) {
%(fail)s
}
""" % vars
......@@ -187,7 +187,7 @@ class GpuDot22(BlasOp, Dot22):
return code
def c_code_cache_version(self):
return (0,)
return (1,)
def c_headers(self):
ret = super(GpuDot22, self).c_headers()
......
theano/sandbox/gpuarray/opt.py
浏览文件 @ 9ad99d0f
......@@ -5,7 +5,7 @@ from theano import tensor, scalar
from theano.compile import optdb
from theano.gof import (local_optimizer, EquilibriumDB,
SequenceDB, ProxyDB,
Optimizer, toolbox, DestroyHandler,
Optimizer, toolbox,
InconsistencyError, EquilibriumOptimizer)
from theano.gof.python25 import all, any
......@@ -90,7 +90,6 @@ class InputToGpuOptimizer(Optimizer):
def add_requirements(self, fgraph):
fgraph.attach_feature(toolbox.ReplaceValidate())
fgraph.attach_feature(DestroyHandler())
def apply(self, fgraph):
for input in fgraph.inputs:
......
theano/sandbox/rng_mrg.py
浏览文件 @ 9ad99d0f
......@@ -734,9 +734,11 @@ class MRG_RandomStreams(object):
:param low: Lower bound of the interval on which values are sampled.
If the ``dtype`` arg is provided, ``low`` will be cast into dtype.
This bound is excluded.
:param high: Higher bound of the interval on which values are sampled.
If the ``dtype`` arg is provided, ``high`` will be cast into dtype.
This bound is excluded.
:param size: Can be a list of integer or Theano variable
(ex: the shape of other Theano Variable)
......
theano/scalar/basic.py
浏览文件 @ 9ad99d0f
......@@ -869,7 +869,8 @@ class ScalarOp(Op):
return self.name
else:
param = [(k, v) for k, v in self.__dict__.items()
if k not in ["name", "_op_use_c_code"]]
if k not in ["name", "_op_use_c_code",
"output_types_preference"]]
if param:
return "%s{%s}" % (self.__class__.__name__,
", ".join("%s=%s" % (k, v)
......
theano/scan_module/scan_opt.py
浏览文件 @ 9ad99d0f
......@@ -1509,7 +1509,6 @@ class PushOutDot1(gof.Optimizer):
def add_requirements(self, fgraph):
fgraph.attach_feature(toolbox.ReplaceValidate())
fgraph.attach_feature(DestroyHandler())
def apply(self, fgraph):
......
theano/sparse/basic.py
浏览文件 @ 9ad99d0f
......@@ -2623,11 +2623,14 @@ class TrueDot(gof.op.Op):
self.grad_preserves_dense = grad_preserves_dense
def __eq__(self, other):
return (type(self) == type(other) and
self.grad_preserves_dense == other.grad_preserves_dense)
# The grad_preserves_dense attribute doesn't change the
# execution behavior. To let the optimizer merge nodes with
# different values of this attribute we shouldn't compare it
# here.
return type(self) == type(other)
def __hash__(self):
return hash(type(self)) ^ hash(self.grad_preserves_dense)
return hash(type(self))
def __ne__(self, other):
return not (self == other)
......@@ -2712,15 +2715,17 @@ class TrueDot(gof.op.Op):
def true_dot(x, y, grad_preserves_dense=True):
"""
Operation for efficiently calculating the dot product when
one or all operands is sparse. Supported format are CSC and CSR.
one or all operands are sparse. Supported formats are CSC and CSR.
The output of the operation is sparse.
:param x: Matrix variable.
:param y: Matrix variable.
:param grad_preserves_dense: if True and one on the input is dense,
make the output dense.
:param x: Sparse matrix or 2d tensor variable.
:param y: Sparse matrix or 2d tensor variable.
:param grad_preserves_dense: if True (default), makes the grad of
dense inputs dense. Otherwise the grad is always sparse.
:return: The dot product `x`.`y` in a sparse format.
:note: one of ``x`` or ``y`` must be sparse.
"""
# TODO
# Maybe the triple-transposition formulation
......
theano/tensor/basic.py
浏览文件 @ 9ad99d0f
......@@ -557,7 +557,7 @@ def get_scalar_constant_value(v):
data = v.data
return numpy_scalar(data)
if v.owner:
if getattr(v, 'owner', None):
if isinstance(v.owner.op, (Alloc, DimShuffle, Rebroadcast,
compile.ops.OutputGuard,
compile.DeepCopyOp)):
......@@ -590,14 +590,10 @@ def get_scalar_constant_value(v):
v.owner.op.perform(v.owner, const, ret)
return ret[0][0]
if isinstance(v.owner.op, theano.tensor.subtensor.Subtensor) and v.ndim == 0:
# This condition depends on Subtensor always embedding constant
# indices in the Op rather than making them inputs to the Apply
# node.
if isinstance(v.owner.inputs[0], TensorConstant) and \
len(v.owner.inputs) == 1:
if isinstance(v.owner.inputs[0], TensorConstant):
cdata = tuple(v.owner.op.get_constant_idx(v.owner.inputs))
try:
return v.owner.inputs[0].data.__getitem__(
tuple(v.owner.op.idx_list))
return v.owner.inputs[0].data.__getitem__(cdata)
except IndexError:
raise IndexError(
str(tuple(v.owner.op.idx_list)) +
......@@ -620,10 +616,12 @@ def get_scalar_constant_value(v):
v.owner.inputs[0].owner.inputs) and
len(v.owner.op.idx_list) == 1):
idx = v.owner.op.idx_list[0]
if isinstance(idx, gof.Type):
idx = get_scalar_constant_value(v.owner.inputs[1])
# Note the '+ 1' is because the first argument to Join is the
# axis.
ret = v.owner.inputs[0].owner.inputs[
v.owner.op.idx_list[0] + 1]
ret = v.owner.inputs[0].owner.inputs[idx + 1]
ret = get_scalar_constant_value(ret)
# join can cast implicitly its input in some case.
return theano._asarray(ret, dtype=v.type.dtype)
......@@ -635,14 +633,13 @@ def get_scalar_constant_value(v):
# We put this check in case there is change in the future
python_all(var.ndim == 0 for var in
v.owner.inputs[0].owner.inputs) and
len(v.owner.op.idx_list) == 1 and
#idx_list can contain Scalar Type object.
isinstance(v.owner.op.idx_list[0], (int, long,
numpy.integer))):
len(v.owner.op.idx_list) == 1):
idx = v.owner.op.idx_list[0]
if isinstance(idx, gof.Type):
idx = get_scalar_constant_value(v.owner.inputs[1])
# Python 2.4 does not support indexing with numpy.integer
# So we cast it.
idx = int(v.owner.op.idx_list[0])
idx = int(idx)
ret = v.owner.inputs[0].owner.inputs[idx]
ret = get_scalar_constant_value(ret)
# MakeVector can cast implicitly its input in some case.
......@@ -658,6 +655,8 @@ def get_scalar_constant_value(v):
op = owner.op
idx_list = op.idx_list
idx = idx_list[0]
if isinstance(idx, gof.Type):
idx = get_scalar_constant_value(owner.inputs[1])
grandparent = leftmost_parent.owner.inputs[0]
gp_broadcastable = grandparent.type.broadcastable
ndim = grandparent.type.ndim
......@@ -3107,6 +3106,48 @@ def batched_dot(x, y):
return result
def batched_tensordot(x, y, axes=2):
"""
:param x: A Tensor with sizes e.g.: for 3D (dim1, dim3, dim2)
:param y: A Tensor with sizes e.g.: for 3D (dim1, dim2, dim4)
:param axes: an integer or array. If an integer, the number of axes
to sum over. If an array, it must have two array
elements containing the axes to sum over in each tensor.
If an integer i, it is converted to an array containing
the last i dimensions of the first tensor and the first
i dimensions of the second tensor (excluding the first
(batch) dimension):
axes = [range(a.ndim - i, b.ndim), range(1,i+1)]
If an array, its two elements must contain compatible axes
of the two tensors. For example, [[1, 2], [2, 4]] means sum
over the 2nd and 3rd axes of a and the 3rd and 5th axes of b.
(Remember axes are zero-indexed!) The 2nd axis of a and the
3rd axis of b must have the same shape; the same is true for
the 3rd axis of a and the 5th axis of b.
:type axes: int or array-like of length 2
A hybrid of batch_dot and tensordot, this function computes the
tensordot product between the two tensors, by iterating over the
first dimension using scan to perform a sequence of tensordots.
"""
if isinstance(axes, (list, numpy.ndarray)):
if isinstance(axes, list):
axes = numpy.asarray(axes)
else:
axes = axes.copy()
assert numpy.greater(axes,0).all(), "All axes should be greater than one, as the first axis is iterated over (batch-wise scan)"
axes -= 1
result, updates = theano.scan(fn=lambda x_mat, y_mat:
theano.tensor.tensordot(x_mat, y_mat, axes),
outputs_info=None,
sequences=[x, y],
non_sequences=None)
return result
def split(x, splits_size, n_splits, axis=0):
the_split = Split(n_splits)
return the_split(x, axis, splits_size)
......
theano/tensor/blas.py
浏览文件 @ 9ad99d0f
......@@ -139,7 +139,7 @@ except ImportError:
pass
from theano.configparser import config, AddConfigVar, StrParam
from theano.gof import (utils, Op, view_roots, DestroyHandler,
from theano.gof import (utils, Op, view_roots,
local_optimizer, Optimizer,
InconsistencyError, toolbox, SequenceDB,
EquilibriumOptimizer, Apply,
......@@ -1488,7 +1488,6 @@ class GemmOptimizer(Optimizer):
def add_requirements(self, fgraph):
fgraph.attach_feature(toolbox.ReplaceValidate())
fgraph.attach_feature(DestroyHandler())
def apply(self, fgraph):
did_something = True
......@@ -1501,9 +1500,21 @@ class GemmOptimizer(Optimizer):
time_factor_can = 0
time_factor_list = 0
time_toposort = 0
if fgraph.profile:
validate_before = fgraph.profile.validate_time
callbacks_before = fgraph.execute_callbacks_times.copy()
callback_before = fgraph.execute_callbacks_time
class Updater:
def on_import(self, fgraph, new_node, reason):
if new_node is not node:
nodelist.append(new_node)
u = Updater()
fgraph.attach_feature(u)
while did_something:
nb_iter += 1
t0 = time.time()
nodelist = list(fgraph.toposort())
nodelist = theano.gof.graph.io_toposort(fgraph.inputs, fgraph.outputs)
time_toposort += time.time() - t0
did_something = False
nodelist.reverse()
......@@ -1546,16 +1557,30 @@ class GemmOptimizer(Optimizer):
except ReplacementDidntRemovedError, e:
nb_replacement_didn_t_remove += 1
self.warned = True
nb_iter += 1
fgraph.remove_feature(u)
if fgraph.profile:
validate_time = fgraph.profile.validate_time - validate_before
callback_time = fgraph.execute_callbacks_time - callback_before
callbacks_time = {}
for k, v in fgraph.execute_callbacks_times.iteritems():
if k in callbacks_before:
callbacks_time[k] = v - callbacks_before[k]
else:
callbacks_time[k] = v
else:
validate_time = None
callback_time = None
callbacks_time = {}
return (self, nb_iter, nb_replacement, nb_replacement_didn_t_remove,
nb_inconsistency_make, nb_inconsistency_replace,
time_canonicalize, time_factor_can,
time_factor_list, time_toposort)
time_factor_list, time_toposort,
validate_time, callback_time, callbacks_time,)
@staticmethod
def print_profile(stream, prof, level=0):
blanc = (' ' * level)
#1946.912556s - ('gemm_optimizer', 'GemmOptimizer', 1)
print >> stream, blanc, "GemmOptimizer"
print >> stream, blanc, " nb_iter", prof[1]
print >> stream, blanc, " nb_replacement", prof[2]
......@@ -1566,6 +1591,12 @@ class GemmOptimizer(Optimizer):
print >> stream, blanc, " time_factor_can", prof[7]
print >> stream, blanc, " time_factor_list", prof[8]
print >> stream, blanc, " time_toposort", prof[9]
print >> stream, blanc, " validate_time", prof[10]
print >> stream, blanc, " callback_time", prof[11]
print >> stream, blanc, " callbacks_time"
for i in sorted(prof[12].iteritems(), key=lambda a: a[1]):
if i[1] > 0:
print i
class Dot22(GemmRelated):
......@@ -1816,17 +1847,15 @@ blas_optdb.register('local_gemm_to_gemv',
15, 'fast_run')
# After destroyhandler is in but before we try to make elemwise things inplace
# Try to make gemm inplace
# Also, need to make the gemm optimisation(step 70) happen before the
# fusion of elemwise(step 71)
# After destroyhandler(49.5) but before we try to make elemwise things
# inplace (75)
blas_opt_inplace = in2out(local_inplace_gemm,
local_inplace_gemv,
local_inplace_ger,
name="blas_opt_inplace")
optdb.register('InplaceBlasOpt',
blas_opt_inplace,
70.0, 'fast_run', 'inplace')
blas_opt_inplace,
70.0, 'fast_run', 'inplace', 'blas_opt_inplace')
class Dot22Scalar(GemmRelated):
......
theano/tensor/nnet/Conv3D.py
浏览文件 @ 9ad99d0f
......@@ -562,9 +562,13 @@ conv3D = Conv3D()
:note: The order of dimensions does not correspond to the one in `conv2d`.
This is for optimization.
:note: The GPU implementation is very slow. You are better to use
:func:`conv3d2d <theano.tensor.nnet.conv3d2d.conv3d>` that is faster
on GPU.
:note: The GPU implementation is very slow. You should use
:func:`conv3d2d <theano.tensor.nnet.conv3d2d.conv3d>` for a GPU
graph instead.
:see: Someone made a script that shows how to swap the axes between
both 3d convolution implementations in Theano. See the last
`attachment <https://groups.google.com/d/msg/theano-users/1S9_bZgHxVw/0cQR9a4riFUJ>`_.
"""
......
theano/tensor/nnet/conv3d2d.py
浏览文件 @ 9ad99d0f
......@@ -178,6 +178,10 @@ def conv3d(signals, filters,
Another way to define signals: (batch, time, in channel, row, column)
Another way to define filters: (out channel,time,in channel, row, column)
:see: Someone made a script that shows how to swap the axes between
both 3d convolution implementations in Theano. See the last
`attachment <https://groups.google.com/d/msg/theano-users/1S9_bZgHxVw/0cQR9a4riFUJ>`_.
"""
if isinstance(border_mode, str):
......
theano/tensor/nnet/nnet.py
浏览文件 @ 9ad99d0f
......@@ -1396,8 +1396,9 @@ def _check_rows_is_arange_len_labels(rows, labels):
# ShapeOptimizer, but we keep it if ShapeOptimizer is not present
if isinstance(stop.owner.op, subtensor.Subtensor):
shape_subtensor = stop.owner
if list(shape_subtensor.op.idx_list) == [0]:
shape_var, = shape_subtensor.inputs
if shape_subtensor.op.get_constant_idx(shape_subtensor.inputs,
allow_partial=True) == [0]:
shape_var = shape_subtensor.inputs[0]
if shape_var.owner and shape_var.owner.op == tensor.shape:
return shape_var.owner.inputs[0] is labels
else:
......
theano/tensor/opt.py
浏览文件 @ 9ad99d0f
差异被折叠。
theano/tensor/raw_random.py
浏览文件 @ 9ad99d0f
......@@ -576,11 +576,11 @@ def random_integers(random_state, size=None, low=0, high=1, ndim=None,
def choice_helper(random_state, a, replace, p, size):
"""
Helper function to draw random numbers using numpy's choice function.
"""Helper function to draw random numbers using numpy's choice function.
This is a generalization of numpy.random.choice to the case where `a`,
`replace` and `p` are tensors.
This is a generalization of numpy.random.choice that coerces
`replace` to a bool and replaces `p` with None when p is a vector
of 0 elements.
"""
if a.ndim > 1:
raise ValueError('a.ndim (%i) must be 0 or 1' % a.ndim)
......@@ -622,16 +622,6 @@ def choice(random_state, size=None, a=2, replace=True, p=None, ndim=None,
broadcastable=bcast))
return op(random_state, size, a, replace, p)
def poisson_helper(random_state, lam, size):
"""
Helper function to draw random numbers using numpy's poisson function.
This is a generalization of numpy.random.poisson to the case where
`lam` is a tensor.
"""
return random_state.poisson(lam, size)
def poisson(random_state, size=None, lam=1.0, ndim=None, dtype='int64'):
"""
Draw samples from a Poisson distribution.
......@@ -652,7 +642,7 @@ def poisson(random_state, size=None, lam=1.0, ndim=None, dtype='int64'):
ndim, size, bcast = _infer_ndim_bcast(ndim, size)
op = RandomFunction(poisson_helper, tensor.TensorType(dtype=dtype,
op = RandomFunction("poisson", tensor.TensorType(dtype=dtype,
broadcastable=bcast))
return op(random_state, size, lam)
......@@ -668,6 +658,9 @@ def permutation_helper(random_state, n, shape):
If you wish to perform a permutation of the elements of an existing vector,
see shuffle_row_elements.
This is a generalization of numpy.random.permutation to tensors.
Otherwise it behaves the same.
"""
# n should be a 0-dimension array
assert n.shape == ()
......@@ -680,7 +673,7 @@ def permutation_helper(random_state, n, shape):
shape = ()
out_shape = list(shape)
out_shape.append(n)
out = numpy.zeros(out_shape, int)
out = numpy.empty(out_shape, int)
for i in numpy.ndindex(*shape):
out[i] = random_state.permutation(n)
......@@ -869,7 +862,7 @@ class RandomStreamsBase(object):
def binomial(self, size=None, n=1, p=0.5, ndim=None, dtype='int64',
prob=None):
"""
Sample n times with probability of success prob for each trial,
Sample n times with probability of success p for each trial and
return the number of successes.
If the size argument is ambiguous on the number of dimensions,
......
theano/tensor/subtensor.py
浏览文件 @ 9ad99d0f
......@@ -47,6 +47,23 @@ class AdvancedIndexingError(TypeError):
# Helpful functions to deal with Subtensor and IncSubtensor
##########
def make_constant(args):
"""
Convert python litterals to theano constants in subtensor arguments.
"""
def conv(a):
if a is None:
return a
elif isinstance(a, slice):
return slice(conv(a.start),
conv(a.stop),
conv(a.step))
elif isinstance(a, (int, long, numpy.integer)):
return scal.ScalarConstant(scal.int64, a)
else:
return a
return tuple(map(conv, args))
def get_idx_list(inputs, idx_list):
'''
Given a list of inputs to the subtensor and its idx_list reorders
......@@ -347,24 +364,56 @@ class Subtensor(Op):
slice_c = None
return slice(slice_a, slice_b, slice_c)
# There is a bug in numpy that results in isinstance(x, int) returning
# False for numpy integers.
# See <http://projects.scipy.org/numpy/ticket/2235>.
elif isinstance(entry, numpy.integer):
return entry
# On Windows 64-bit, shapes are returned as Python long, as they can
# be bigger than what a Python int can hold.
# Shapes should always fit in a numpy.int64, and we support them better
# 2) In Python3, long replaced int. So we must assert it fit in int64.
elif isinstance(entry, (int, long)):
entry64 = numpy.int64(entry)
return entry64
elif isinstance(entry, (int, long, numpy.integer)):
# Disallow the use of python scalars in idx_list
raise TypeError("Python scalar in idx_list."
"Please report this error to theano-dev.")
else:
raise AdvancedIndexingError(Subtensor.e_indextype, entry)
def get_constant_idx(self, inputs, allow_partial=False):
"""
Return the idx_list with constant inputs replaced by their
python scalar equivalent. May raise
`theano.tensor.NotScalarConstantError` if the idx contains
non-constant entries.
If allow_partial is True, then entries that are not constant
will stay as their input variable rather than raising an
exception.
None entries are always left as-is.
Example usage (where v, a are appropriately typed theano variables):
>>> b = a[v, 1:3]
>>> b.owner.op.idx_list
(Scalar(int64), slice(Scalar(int64), Scalar(int64), None))
>>> b.owner.op.get_constant_idx(b.owner.inputs, allow_partial=True)
[v, slice(1, 3, None)]
>>> b.owner.op.get_constant_idx(b.owner.inputs)
NotScalarConstantError: v
"""
real_idx = get_idx_list(inputs, self.idx_list)
def conv(val):
if val is None:
return None
elif isinstance(val, slice):
return slice(conv(val.start),
conv(val.stop),
conv(val.step))
else:
try:
return get_scalar_constant_value(val)
except theano.tensor.NotScalarConstantError:
if allow_partial:
return val
else:
raise
return map(conv, real_idx)
def __init__(self, idx_list):
self.idx_list = tuple(map(self.convert, idx_list))
self.perform_cache_cdata = None
@staticmethod
def my_as_scalar(a):
......@@ -404,31 +453,21 @@ class Subtensor(Op):
% (input.type, expected_type))
# infer the broadcasting pattern
padded = (idx_list
+ [slice(None, None, None)] * (x.type.ndim - len(idx_list)))
padded = (self.get_constant_idx((None,)+inputs, allow_partial=True)
+ [slice(None, None, None)] * (x.type.ndim - len(idx_list)))
broadcastable = []
for i, (p, bc) in enumerate(izip(padded, x.type.broadcastable)):
if isinstance(p, slice):
if bc and p.start in [None, 0]:
# No need to check step when there is only
# one element.
# We could call get_canonical_form_slice() to
# catch more broadcast case. I let this to
# later.
if p.stop is None:
broadcastable.append(bc)
start = p.start
if start is None:
start = 0
if (p.stop is None or
(isinstance(p.stop, (int, numpy.integer)) and
p.stop > start)):
broadcastable.append(True)
continue
try:
if p.start is None:
start = 0
else:
start = get_scalar_constant_value(p.start)
stop = get_scalar_constant_value(p.stop)
if stop > start:
broadcastable.append(True)
continue
except theano.tensor.NotScalarConstantError:
pass
broadcastable.append(False)
return gof.Apply(self,
......@@ -440,18 +479,9 @@ class Subtensor(Op):
out, = out_
x = inputs[0]
# The subtensor (or idx_list) does not depend on the inputs.
# (and cdata was cached on initial call)
if self.perform_cache_cdata is not None:
out[0] = numpy.asarray(x.__getitem__(self.perform_cache_cdata))
return
cdata = get_idx_list(inputs, self.idx_list)
if len(cdata) == 1:
cdata = cdata[0]
# (first call caches cdata here)
if len(inputs) == 1:
self.perform_cache_cdata = cdata
out[0] = numpy.asarray(x.__getitem__(cdata))
......
theano/tensor/tests/test_basic.py
浏览文件 @ 9ad99d0f
......@@ -2323,8 +2323,32 @@ def test_batched_dot():
result_fn = theano.function([first_mat, second_mat], output)
result = result_fn(first_mat_val, second_mat_val)
assert result.shape[0] == first_mat_val.shape[0]
def test_batched_tensordot():
first = theano.tensor.tensor4("first")
second = theano.tensor.tensor4("second")
axes = [[1,2], [3,1]]
output = theano.tensor.basic.batched_tensordot(first, second, axes)
first_val = numpy.random.rand(8, 10, 20, 3).astype(config.floatX)
second_val = numpy.random.rand(8, 20, 5, 10).astype(config.floatX)
result_fn = theano.function([first, second], output)
result = result_fn(first_val, second_val)
assert result.shape[0] == first_val.shape[0]
assert result.shape[1] == first_val.shape[3]
assert result.shape[2] == second_val.shape[2]
first_mat = theano.tensor.dmatrix("first")
second_mat = theano.tensor.dmatrix("second")
axes = 1
output = theano.tensor.basic.batched_tensordot(first_mat, second_mat, axes)
first_mat_val = numpy.random.rand(10, 4).astype(config.floatX)
second_mat_val = numpy.random.rand(10, 4).astype(config.floatX)
result_fn = theano.function([first_mat, second_mat], output)
result = result_fn(first_mat_val, second_mat_val)
print(result.shape)
assert result.shape[0] == first_mat_val.shape[0]
assert len(result.shape) == 1
def test_tensor_values_eq_approx():
#test, inf, -inf and nan equal themself
......
theano/tensor/var.py
浏览文件 @ 9ad99d0f
......@@ -348,6 +348,8 @@ class _tensor_py_operators:
def __getitem__(self, args):
if not isinstance(args, tuple):
args = args,
# Convert python literals to theano constants
args = theano.tensor.subtensor.make_constant(args)
# Determine if advanced indexing is needed or not
# The logic is already in Subtensor.convert: if it succeeds,
# standard indexing is used; if it fails with
......
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