e1b7872fc2
Make it possible to access IR from Python.
...
Also, add a new trace_fn field to attach forward IR to Variables.
Signed-off-by: Edward Z. Yang <ezyang@fb.com >
2017-09-05 17:48:55 -04:00
94b5990201
Add torch.cuda.get_device_name function ( #2540 )
2017-08-26 15:06:37 -04:00
eb58740651
add ones_like and zeros_like
2017-08-25 14:11:04 -04:00
c000d15058
Properly use Py_RETURN_True, Py_RETURN_False in back compatibility warnings. ( #2345 )
2017-08-08 21:54:20 -04:00
9d8cff9bc1
initialize aten and pytorch to share the same THCState
2017-07-11 10:35:03 -04:00
714351ff39
Officially enable process-group mode
2017-06-12 22:02:11 -04:00
4f602a52b5
Use THPUtils_assert rather than THError in torch/csrc/Module.
2017-06-11 05:37:59 -04:00
ffd808768e
Remove raiseErrors from THTensor functions, have THStorage functions take an error_buffer to return a proper error message while being able to handle memory management correctly from calling function.
2017-06-11 05:37:59 -04:00
177785eecf
explicit Ptr constructors, fast transposed copy.
2017-06-11 05:37:59 -04:00
be65f46c76
Add optional warning for backwards incompatible keepdim. Setting torch.utils.backcompat.keepdim.warning.enabled=True will cause Python warnings in the case where the default value of keepdim is used for 1-d reductions.
...
Also specify keepdim via kwargs in library so these warnings have less
noise.
2017-06-11 05:37:59 -04:00
3556d1b8a3
Add optional warning for backwards incompatible broadcast.
...
Setting torch.utils.backcompat.broadcast.warning.enabled=True
will cause Python warnings in the case where broadcast occurs
but previously 1-d view style pointwise ops occured.
2017-06-11 05:37:59 -04:00
5af46cb352
Add broadcasting support for matmul.
2017-06-11 05:37:59 -04:00
d81da41650
Make sure the number of MKL and OpenMP threads match
...
Otherwise, on many machines, the size of the OpenMP thread pool will
change between MKL and our OpenMP enabled functions. The constant thread
creation and destruction results in worse performance and leaks memory
on GCC 5.4
2017-06-07 14:53:29 -04:00
8ea7c87c29
Improve init methods
2017-06-02 23:42:11 +02:00
181d2f41bd
Add initial Python wrappers for THDTensors
2017-06-02 23:42:11 +02:00
05bc877a05
make THPPointer have explicit constructors ( #1636 )
2017-05-25 15:35:54 -04:00
d0504aa41d
Implement lgamma function.
2017-05-08 16:21:26 -07:00
4c1cdb6148
Refactor Python string utility function
2017-04-28 21:25:26 +02:00
27990fee54
Use fully qualified name as tp_name for tensors and storages ( #1379 )
2017-04-27 16:26:44 -04:00
cd3bbc9dfd
more operations and optimizations (hspmm, reorder, ...)
2017-04-18 12:46:54 -07:00
71303b8af4
Autograd deadlock for recent glibc fix ( #1243 )
2017-04-12 22:24:31 +02:00
afeeb81e79
Add support for keyword arguments in torch.cat
2017-04-11 14:48:54 -07:00
91c4ba7980
Add torch.arange and deprecate torch.range
2017-04-03 10:38:58 -04:00
dfa2d26830
* make random_ range correct when both lower and upper are specified
2017-03-31 15:37:24 -04:00
8dc5d2a22e
export current_blas_handle
2017-03-23 23:32:45 +01:00
bb353ccc17
Add batch triangular factorization and solves, add IntegerTensor to cwrap ( #903 )
2017-03-23 15:06:00 -04:00
faac0f5c25
Fix torch.cat bugs
...
Always use PySequence API and disallow catting along inexistent
dimensions.
2017-03-22 18:58:42 -04:00
379ae6d865
Refactor out dispatchStateless ( #1007 )
...
Some of the error messages were incorrect due to erroneous
'tensor == THPDefaultTensorClass' checks
2017-03-15 16:24:55 -04:00
f17cfe4293
sparse tensor operations ( #735 )
2017-03-03 18:37:03 +01:00
f366e5fc81
Support int16 numpy conversions
...
issue #891
2017-03-02 09:15:57 -05:00
fc6fcf23f7
Lock the cudaFree mutex. ( #880 )
...
Prevents NCCL calls from overlapping with cudaFree() which can lead to
deadlocks.
2017-03-01 11:29:25 -05:00
67f94557ff
Expose torch.HalfTensor
2017-02-27 19:35:47 -05:00
bd5303010d
Refactor autograd package to separate Python dependencies. ( #662 )
...
The core autograd Variable, Function, and Engine no longer depend on the
Python API. This let's us implement functions in C++. In the future, we
can also multithread engine and release the GIL for most of the
non-Python backwards.
2017-02-13 16:00:16 -08:00
712686ce91
Add cat, contiguous, squeeze, and unsqueeze to THPP
...
Use unsqueeze and view from TH/THC
2017-02-11 17:49:31 +01:00
79232c24e2
Fixes after rebase
2017-01-31 01:58:09 +01:00
76520512e7
DataChannel tests rewrite ( #42 ); DataChannel isend and irecv implementation ( #44 )
2017-01-31 01:58:09 +01:00
60d1852c7b
Major improvements to master-worker mode
...
* Fixed all undefined symbol errors
* Implemented storage interface and THStorage class
* RPC improvements
* Code refactor
2017-01-31 01:58:09 +01:00
55632d81d2
Add Python wrappers for process group mode
2017-01-31 01:58:09 +01:00
c414bf0aaf
Fix handling of unicode in torch._C._add_docstr ( #487 )
2017-01-18 17:22:30 -05:00
9302f860ae
Remove unused file TensorDocstrings.cpp ( #481 )
...
Tensor docstrings are created in _tensor_docs.py
2017-01-18 13:34:40 -05:00
8aa8f791fc
add more torch.* and Tensor docs ( #476 )
2017-01-18 08:39:33 -05:00
14d5d52789
Add placeholder tensor documentation for methods that exist in torch. ( #463 )
2017-01-17 19:37:47 -05:00
f91bb96071
Remove cmin, cmax and cinv
2017-01-16 19:07:37 -05:00
bdfef2975c
adding more docs for torch.* functions
2017-01-11 08:19:49 -08:00
59d66e6963
Sparse Library ( #333 )
2017-01-05 00:43:41 +01:00
6b4ed52f10
adding docs for some torch.* functions, removing all, any stateless methods
2017-01-03 18:29:50 -05:00
849794cd2c
Remove deprecated and unimplemented functions ( #383 )
2016-12-30 18:37:44 -05:00
ab5776449c
Add documentation for some torch.xxx functions ( #382 )
2016-12-30 17:01:47 -05:00
9b7eceddc8
Accept outputs in out argument
2016-12-29 12:25:59 +01:00
24af02154c
Use ForkingPickler for sharing tensor/storages across processes ( #344 )
...
This hooks into the (internal) ForkingPickler class in multiprocessing
to reduce tensors, storages, and CUDA events instead of our queue from
joblib. This makes it easier to use the standard multiprocessing classes
in later versions of Python.
This also exposes:
- Tensor/Storage.share_memory_()
- Module.share_memory()
These methods move the CPU tensors and storages to shared memory. If
you're using the "fork" method of multiprocessing, these objects can be
directly inherited instead of serialized through a queue.
2016-12-28 20:34:23 -05:00
