diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml index 1b6b588e..a9450588 100644 --- a/.github/workflows/ci.yml +++ b/.github/workflows/ci.yml @@ -29,7 +29,6 @@ jobs: uses: actions/checkout@v2 - name: Checkout submodules run: | - apt update&&apt install ninja-build ./third_party/prepare.sh ./third_party/install-mkl.sh - name: Build MegEngine @@ -58,7 +57,6 @@ jobs: uses: actions/checkout@v2 - name: Checkout submodules run: | - apt update&&apt install ninja-build ./third_party/prepare.sh ./third_party/install-mkl.sh - name: Build MegEngine diff --git a/README.md b/README.md index 1008b517..bf7d4619 100644 --- a/README.md +++ b/README.md @@ -12,7 +12,7 @@ MegEngine is a fast, scalable and easy-to-use deep learning framework, with auto ## Installation -**NOTE:** MegEngine now supports Python installation on Linux-64bit/Windows-64bit/MacOS(CPU-Only)-10.14+/Android 7+(CPU-Only) platforms with Python from 3.5 to 3.8. On Windows 10 you can either install the Linux distribution through [Windows Subsystem for Linux (WSL)](https://docs.microsoft.com/en-us/windows/wsl) or install the Windows distribution directly. Many other platforms are supported for inference. +**NOTE:** MegEngine now supports Python installation on Linux-64bit/Windows-64bit/MacOS(CPU-Only)-10.14+ platforms with Python from 3.5 to 3.8. On Windows 10 you can either install the Linux distribution through [Windows Subsystem for Linux (WSL)](https://docs.microsoft.com/en-us/windows/wsl) or install the Windows distribution directly. Many other platforms are supported for inference. ### Binaries diff --git a/README_CN.md b/README_CN.md index c72a0d8d..3a2eb426 100644 --- a/README_CN.md +++ b/README_CN.md @@ -13,7 +13,7 @@ MegEngine 是一个快速、可拓展、易于使用且支持自动求导的深 ## 安装说明 -**注意:** MegEngine 现在支持在 Linux-64bit/Windows-64bit/macos-10.14/Android 7+ 及其以上 (MacOS/Android只支持cpu) 等平台上安装 Python 包，支持Python3.5 到 Python3.8。对于 Windows 10 用户，可以通过安装 [Windows Subsystem for Linux (WSL)](https://docs.microsoft.com/en-us/windows/wsl) 进行体验，同时我们也原生支持Windows。MegEngine 也支持在很多其它平台上进行推理运算。 +**注意:** MegEngine 现在支持在 Linux-64bit/Windows-64bit/macos-10.14及其以上 (MacOS只支持cpu) 等平台上安装 Python 包，支持Python3.5 到 Python3.8。对于 Windows 10 用户，可以通过安装 [Windows Subsystem for Linux (WSL)](https://docs.microsoft.com/en-us/windows/wsl) 进行体验，同时我们也原生支持Windows。MegEngine 也支持在很多其它平台上进行推理运算。 ### 通过包管理器安装 @@ -26,8 +26,8 @@ python3 -m pip install megengine -f https://megengine.org.cn/whl/mge.html ## 通过源码编译安装 -* CMake 编译细节请参考 [BUILD_README.md](scripts/cmake-build/BUILD_README.md) -* Python 绑定编译细节请参考 [BUILD_PYTHON_WHL_README.md](scripts/whl/BUILD_PYTHON_WHL_README.md) +* CMake编译细节请参考 [BUILD_README.md](scripts/cmake-build/BUILD_README.md) +* Python绑定编译细节请参考 [BUILD_PYTHON_WHL_README.md](scripts/whl/BUILD_PYTHON_WHL_README.md) ## 如何参与贡献 diff --git a/ci/cmake.sh b/ci/cmake.sh index 4808e63e..8d8c55bf 100755 --- a/ci/cmake.sh +++ b/ci/cmake.sh @@ -27,8 +27,7 @@ function build() { -DMGE_WITH_DISTRIBUTED=${DMGE_WITH_DISTRIBUTED} \ -DMGE_WITH_CUDA=${DMGE_WITH_CUDA} \ -DMGE_WITH_TEST=ON \ - -DCMAKE_BUILD_TYPE=RelWithDebInfo \ - -DMGE_WITH_CUSTOM_OP=ON + -DCMAKE_BUILD_TYPE=RelWithDebInfo make -j$(($(nproc) * 2)) -I ${build_dir} make develop popd >/dev/null diff --git a/imperative/python/megengine/functional/math.py b/imperative/python/megengine/functional/math.py index 3690f562..b0fcc7ce 100644 --- a/imperative/python/megengine/functional/math.py +++ b/imperative/python/megengine/functional/math.py @@ -1153,35 +1153,39 @@ def dot(inp1: Tensor, inp2: Tensor) -> Tensor: def svd(inp: Tensor, full_matrices=False, compute_uv=True) -> Tensor: - r"""Computes the singular value decompositions of input matrix. + r"""Returns a singular value decomposition ``A = USVh`` of a matrix (or a stack of matrices) ``x`` , where ``U`` is a matrix (or a stack of matrices) with orthonormal columns, ``S`` is a vector of non-negative numbers (or stack of vectors), and ``Vh`` is a matrix (or a stack of matrices) with orthonormal rows. Args: - inp: input matrix, must has shape `[..., M, N]`. + x (Tensor): A input real tensor having the shape ``(..., M, N)`` with ``x.ndim >= 2`` . + full_matrices (bool, optional): If ``False`` , ``U`` and ``Vh`` have the shapes ``(..., M, K)`` and ``(..., K, N)`` , respectively, where ``K = min(M, N)`` . If ``True`` , the shapes are ``(..., M, M)`` and ``(..., N, N)`` , respectively. Default: ``False`` . + compute_uv (bool, optional): Whether or not to compute ``U`` and ``Vh`` in addition to ``S`` . Default: ``True`` . + + Note: + * naive does not support ``full_matrices`` and ``compute_uv`` as ``True`` . Returns: - output matrices, `(U, sigma, V)`. + Returns a tuple ( ``U`` , ``S`` , ``Vh`` ), which are SVD factors ``U`` , ``S``, ``Vh`` of input matrix ``x``. ( ``U`` , ``Vh`` only returned when ``compute_uv`` is True). + ``U`` contains matrices orthonormal columns (i.e., the columns are left singular vectors). If ``full_matrices`` is ``True`` , the array must have shape ``(..., M, M)`` . If ``full_matrices`` is ``False`` , the array must have shape ``(..., M, K)`` , where ``K = min(M, N)`` . Examples: - .. testcode:: - - import numpy as np - from megengine import tensor - import megengine.functional as F - - x = tensor(np.arange(0, 6, dtype=np.float32).reshape(2,3)) - _, y, _ = F.svd(x) - print(y.numpy().round(decimals=3)) + >>> import numpy as np + >>> x = Tensor(np.random.randn(9, 6)) + >>> y = Tensor(np.random.randn(2, 7, 8, 3)) - Outputs: - - .. testoutput:: + Reconstruction based on reduced SVD, 2D case: + >>> U, S, Vh = F.svd(x, full_matrices=False) + >>> print(U._tuple_shape, S._tuple_shape, Vh._tuple_shape) + (9, 6) (6,) (6, 6) - [7.348 1. ] + Reconsturction based on reduced SVD, 4D case: + >>> u, s, vh = F.svd(y, full_matrices=False) + >>> print(u._tuple_shape, s._tuple_shape, vh._tuple_shape) + (2, 7, 8, 3) (2, 7, 3) (2, 7, 3, 3) """ op = builtin.SVD(full_matrices=full_matrices, compute_uv=compute_uv) - U, sigma, V = apply(op, inp) - return U, sigma, V + U, S, Vh = apply(op, inp) + return U, S, Vh def _check_non_finite(inps: Iterable[Tensor], scale=1.0) -> Tensor: