| commit | 711e5a6ceb46fcfe90dc8ca176c94c4f44dfbc17 | [log] [tgz] |
|---|---|---|
| author | Edward Z. Yang <ezyang@mit.edu> | Fri Jun 15 17:52:21 2018 -0400 |
| committer | GitHub <noreply@github.com> | Fri Jun 15 17:52:21 2018 -0400 |
| tree | 85102ca198940ccd1d9024b30334a8c0a645faf7 | |
| parent | c537fd7432b1cf2d11552e485dc00dd9eea0e03c [diff] |
Port THS to ATen. (#8409)
* Port THS to ATen.
The basic structure of the patch:
- All kernels in aten/src/THS got rewritten as native
functions in aten/src/ATen/native/sparse
I took the liberty to rename some of the kernels,
opting for a longer, more transparent names than
things like 'spaddcmul'.
- Instead of holding fields for sparse tensor in the TH
C struct THSTensor, they are now held in a C++ class
SparseTensorImpl (this explains why I had to do this
all in one go; I can't have *two* reps for sparse
tensors!)
Along the way, we change a key internal representation
invariant: an "empty" sparse tensor has dimI == 1 and
dimV == 0 (this is different from dimI == 0 and dimV == 0
we had before); this ensures that we maintain the invariant
that dim == dimI + dimV. "Scalar" sparse tensors are
made illegal, because there really is no way to properly
express them in COO format.
- Because we haven't ported THCS or any of the traditional
dense TH implementations, there is a new set of adapter
functions in native/LegacyBridge.cpp exclusively devoted
to deciding whether or not to go to the new native implementation
or back to the legacy TH binding (prefixed with th_).
The intent is that when everything gets ported, we can
delete this file.
- I've kept the stubs for all the THS functions, but they now all
error if you try to actually call them. Eventually, we should
replace these with calls to ATen so that everything keeps
working.
- I gobbled up SparseMM (SparseMM.cpp is no more). It was tasty.
There are some miscellaneous improvements which were needed for other
changes in this patch:
- There is now AT_FORALL_SCALAR_TYPES_EXCEPT_HALF, which does what
it says on the tin.
- axpy templated function moved to TH/BlasUtils.h, there's a new macro
which lets you easily forward to all of the TH functions. We also expose
THBlas_copy. I'm not terribly pleased with these functions but
they seem to serve a purpose they need.
- New method on Tensor to get TensorImpl*, unsafeGetTensorImpl
- accessor() is now this-const, since const-correctness on Tensor is a lie
- New toSparse()/toDense() methods on Type; now you can call these
directly without having to manually apply at::toSparse/toDense
on the Backend and then running toBackend yourself.
Changes to the kernels:
- Previously, the whole body of all kernels was compiled for
every supported scalar type. In our new implementation,
the scalar dispatch has been pushed into the smallest extent
which (1) is not in a type loop and (2) requires statically
knowing the scalar type. These sites all use
AT_DISPATCH_ALL_TYPES. I tried to use lambdas as much as
possible, but sometimes it was not possible when a OpenMP
pragma was used.
- Anywhere we tested if the nDimension of a tensor was zero,
we replaced with a test that numel is zero. Because, as we
known, nDimension of zero-size tensors in TH is zero, and
that's wrong wrong wrong (and not done this way in ATen).
Some subtleties:
- Places where previously fastget1d was used, I now use a
TensorAccessor. However, you have to be careful about grabbing
the accessor, because sometimes you will be accessor'ing
indices/values and they are empty, which means they will
be *1D* ("oh, aren't indices always 2D?" Nope. Nyet.)
So, essentially, it is only safe to grab an accessor *after*
you have checked that nnz != 0. All of these shenanigans
will go away when we properly support zero-size dimensions.
A few places, we test for this case just by wrapping the loop
in a conditional on nnz. Some other places this is not so easy,
so we instead short-circuit the function with a special case for
when nnz == 0 (usually, these implementations are degenerate).
- There is a very subtle but important difference between
_sparse_get_impl(self)->indices() and self._indices();
the latter may return a view! This is because nnz is
not guaranteed to match the dimensions of indices/values;
you can "truncate" a sparse tensor by setting the nnz.
Actually, I think this is not a good idea and we should
enforce a stronger invariant, but for this patch I slavishly
adhere to the old ways, and as such I have to be very
careful if I want to resize something, I had better use
the former and not the latter.
- I had to reimplement broadcasting by hand (thus the s_
and non-s_ functions in the sparse native files). There
is a very important distinction between foo_out and foo_,
so it is important that the LegacyBridge function always
call to the lower layer, and not try to avoid boilerplate
by calling to another LegacyBridge function first.
I did NOT put broadcasting in LegacyBridge (even though,
ultimately, that's where it must live), because the th_
functions which are invoked from LegacyBridge handle
broadcasting themselves, and I don't want to broadcast
twice.
- Sparse function MUST explicitly specify the Type they
dispatch from, otherwise Variable wrapping/unwrapping will
not work correctly. If you use _get_sparse_impl, that is
sufficient to levy this requirement.
- The "has native" tests in LegacyBridge.cpp are not 100%,
because some of the functions are mixed dense-sparse functions,
and so you can't just say, "Oh, if it's sparse and CPU, call
the native sparse implementation." This is handled on a
case by case basis. There is some especially complex
logic for add(), which has dense-dense, sparse-sparse
and dense-sparse implementations.
- I added some uses of SparseTensorRef in native_functions.yaml,
but you will notice that these are all on native_* functions,
and not the actual, top-level functions. So the SparseTensorRef
is purely documentary (helping you not call the wrong overload)
but there is no magic; we do the wrapping ourselves the hard
way. (This is in constrast to the TH binding code which is magical.)
Except for _sparse_mask; _sparse_mask is magical.
- There is a raw_copy_sparse_ method, which is really my way of
getting around the fact that copy_ has never been implemented
for sparse tensors (even before this patch), but there IS a
super secret, internal way of doing these copies that the THS
code used, and which I needed to get my hands on when I did this
port. We should refactor so that either (a) copy_ does support
sparse-sparse copy natively, or (b) we do this other ways.
- Irritatingly, I must explicitly resize_as_ before copy_ into
a tensor. This was not the case with THTensor_(copy) but I don't
have any direct binding that doesn't have this requirement.
- For some reason, the sparse tensor constructor accepts a scalar
tensor for the values tensor. This is kind of weird because
you always need an nnz-dimension. However, the old code supported
this and just expanded it into a 1D size 0 tensor; so we need some
explicit code to do this.
There are maybe a bit more AT_ASSERTs in some of the kernels
than is wise. I added them all when I was debugging and was
loathe to remove them.
Some last mile fixes after this commit went into PR
- Move expand outside of dispatch so autograd works (it used to be inside and then we lost all of the recorded broadcasts).
- Hack to duplicate the derivatives for our now two definitions TH and native. Mercifully the derivatives are short.
- Apparently, TH has a special case to make foo_ functions method only, and if you don't do this the Python arg parsing is wrong. We carefully work around this in the native bindings
- Apply DCE to a test_jit case, fixes wobbling due to DCE trick in tracing
- Update test_function's output
- Some last mile fixes for dispatch confusion in sparse_coo_tensor functions.
- New simplified regression test based on failures I saw in ONNX
- Increase tolerance on super resolution test
- More robust dynamic_type normalization, fixes ONNX bug.
The dynamic_type situation is very delicate; probably need
to stop having both Scalar and real.
- Make new_with_tensor_sparse more CUDA safe
- Note about CUDA-safety in SparseTensorImpl
- Rename dimI/dimV to sparseDims/denseDims.
- Make localScalar on SparseTensorImpl work.
- Make numel uniformly supported on all types, not just dense
types
- Add tests for is_nonzero() method (which exercises localScalar)
- Disable constant JIT autogenerated tests, which are fragile and broken
by this change, but being fixed in a parallel track.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
PyTorch is a Python package that provides two high-level features:
You can reuse your favorite Python packages such as NumPy, SciPy and Cython to extend PyTorch when needed.
We are in an early-release beta. Expect some adventures and rough edges.
| System | 2.7 | 3.5 |
|---|---|---|
| Linux CPU | ||
| Linux GPU | ||
| Windows GPU | — |
See also the ci.pytorch.org HUD
At a granular level, PyTorch is a library that consists of the following components:
Usually one uses PyTorch either as:
Elaborating further:
If you use NumPy, then you have used Tensors (a.k.a ndarray).
PyTorch provides Tensors that can live either on the CPU or the GPU, and accelerate compute by a huge amount.
We provide a wide variety of tensor routines to accelerate and fit your scientific computation needs such as slicing, indexing, math operations, linear algebra, reductions. And they are fast!
PyTorch has a unique way of building neural networks: using and replaying a tape recorder.
Most frameworks such as TensorFlow, Theano, Caffe and CNTK have a static view of the world. One has to build a neural network, and reuse the same structure again and again. Changing the way the network behaves means that one has to start from scratch.
With PyTorch, we use a technique called reverse-mode auto-differentiation, which allows you to change the way your network behaves arbitrarily with zero lag or overhead. Our inspiration comes from several research papers on this topic, as well as current and past work such as torch-autograd, autograd, Chainer, etc.
While this technique is not unique to PyTorch, it's one of the fastest implementations of it to date. You get the best of speed and flexibility for your crazy research.
PyTorch is not a Python binding into a monolithic C++ framework. It is built to be deeply integrated into Python. You can use it naturally like you would use NumPy / SciPy / scikit-learn etc. You can write your new neural network layers in Python itself, using your favorite libraries and use packages such as Cython and Numba. Our goal is to not reinvent the wheel where appropriate.
PyTorch is designed to be intuitive, linear in thought and easy to use. When you execute a line of code, it gets executed. There isn't an asynchronous view of the world. When you drop into a debugger, or receive error messages and stack traces, understanding them is straightforward. The stack trace points to exactly where your code was defined. We hope you never spend hours debugging your code because of bad stack traces or asynchronous and opaque execution engines.
PyTorch has minimal framework overhead. We integrate acceleration libraries such as Intel MKL and NVIDIA (cuDNN, NCCL) to maximize speed. At the core, its CPU and GPU Tensor and neural network backends (TH, THC, THNN, THCUNN) are written as independent libraries with a C99 API. They are mature and have been tested for years.
Hence, PyTorch is quite fast – whether you run small or large neural networks.
The memory usage in PyTorch is extremely efficient compared to Torch or some of the alternatives. We've written custom memory allocators for the GPU to make sure that your deep learning models are maximally memory efficient. This enables you to train bigger deep learning models than before.
Writing new neural network modules, or interfacing with PyTorch's Tensor API was designed to be straightforward and with minimal abstractions.
You can write new neural network layers in Python using the torch API or your favorite NumPy-based libraries such as SciPy.
If you want to write your layers in C/C++, we provide an extension API based on cffi that is efficient and with minimal boilerplate. There is no wrapper code that needs to be written. You can see a tutorial here and an example here.
Commands to install from binaries via Conda or pip wheels are on our website:
If you are installing from source, we highly recommend installing an Anaconda environment. You will get a high-quality BLAS library (MKL) and you get a controlled compiler version regardless of your Linux distro.
Once you have Anaconda installed, here are the instructions.
If you want to compile with CUDA support, install
If you want to disable CUDA support, export environment variable NO_CUDA=1. Other potentially useful environment variables may be found in setup.py.
If you want to build on Windows, Visual Studio 2017 and NVTX are also needed.
On Linux
export CMAKE_PREFIX_PATH="$(dirname $(which conda))/../" # [anaconda root directory] # Install basic dependencies conda install numpy pyyaml mkl mkl-include setuptools cmake cffi typing conda install -c mingfeima mkldnn # Add LAPACK support for the GPU conda install -c pytorch magma-cuda80 # or magma-cuda90 if CUDA 9
On macOS
export CMAKE_PREFIX_PATH=[anaconda root directory] conda install numpy pyyaml mkl mkl-include setuptools cmake cffi typing
On Windows
conda install numpy pyyaml mkl mkl-include setuptools cmake cffi typing
git clone --recursive https://github.com/pytorch/pytorch cd pytorch
On Linux
python setup.py install
On macOS
MACOSX_DEPLOYMENT_TARGET=10.9 CC=clang CXX=clang++ python setup.py install
On Windows
set "VS150COMNTOOLS=C:\Program Files (x86)\Microsoft Visual Studio\2017\Enterprise\VC\Auxiliary\Build" set CMAKE_GENERATOR=Visual Studio 15 2017 Win64 set DISTUTILS_USE_SDK=1 REM The following line is needed for Python 2.7, but the support for it is very experimental. set MSSdk=1 call "%VS150COMNTOOLS%\vcvarsall.bat" x64 -vcvars_ver=14.11 python setup.py install
Dockerfile is supplied to build images with cuda support and cudnn v7. Build as usual
docker build -t pytorch -f docker/pytorch/Dockerfile .
You can also pull a pre-built docker image from Docker Hub and run with nvidia-docker, but this is not currently maintained and will pull PyTorch 0.2.
nvidia-docker run --rm -ti --ipc=host pytorch/pytorch:latest
Please note that PyTorch uses shared memory to share data between processes, so if torch multiprocessing is used (e.g. for multithreaded data loaders) the default shared memory segment size that container runs with is not enough, and you should increase shared memory size either with --ipc=host or --shm-size command line options to nvidia-docker run.
Installation instructions and binaries for previous PyTorch versions may be found on our website.
Three pointers to get you started:
PyTorch has a 90 day release cycle (major releases). Its current state is Beta, we expect no obvious bugs. Please let us know if you encounter a bug by filing an issue.
We appreciate all contributions. If you are planning to contribute back bug-fixes, please do so without any further discussion.
If you plan to contribute new features, utility functions or extensions to the core, please first open an issue and discuss the feature with us. Sending a PR without discussion might end up resulting in a rejected PR, because we might be taking the core in a different direction than you might be aware of.
PyTorch is a community driven project with several skillful engineers and researchers contributing to it.
PyTorch is currently maintained by Adam Paszke, Sam Gross, Soumith Chintala and Gregory Chanan with major contributions coming from 10s of talented individuals in various forms and means. A non-exhaustive but growing list needs to mention: Trevor Killeen, Sasank Chilamkurthy, Sergey Zagoruyko, Adam Lerer, Francisco Massa, Alykhan Tejani, Luca Antiga, Alban Desmaison, Andreas Kopf, James Bradbury, Zeming Lin, Yuandong Tian, Guillaume Lample, Marat Dukhan, Natalia Gimelshein, Christian Sarofeen, Martin Raison, Edward Yang, Zachary Devito.
Note: this project is unrelated to hughperkins/pytorch with the same name. Hugh is a valuable contributor in the Torch community and has helped with many things Torch and PyTorch.