.. _comparison: ======================= 与其它框架 API 进行对比 ======================= .. toctree:: :glob: :hidden: :maxdepth: 1 comparison/* .. warning:: MegEngine 的 API 设计遵循 :ref:`mep-0003` , 向《数组 API 标准》靠齐。 * 在同其它框架进行对比时，同样的命名不意味着用法也完全一致； * 如果有新的 API 支持需求，可在 GitHub 创建相应的 Issue 或 Pull Request. .. note:: * 你可以利用浏览器的查找功能在当前页面查询对应的 API. * 当前页面并非自动生成，如发现有缺失/过时内容，欢迎编辑当前页面。 .. seealso:: 当前页面更多用于检索，具有其它框架使用经验的用户还可以参考 :ref:`transfer-to-megengine` 。 Data Structure -------------- .. list-table:: :header-rows: 1 :widths: 20 20 20 40 * - NumPy - Pytorch - MegEngine - Comment * - :py:class:`~numpy.ndarray` - :py:class:`~torch.Tensor` - :py:class:`~megengine.Tensor` - :ref:`tensor-guide` General tensor operations ------------------------- Creation Functions ~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:func:`~numpy.arange` - :py:func:`~torch.arange` - :py:func:`~megengine.functional.arange` - * - :py:func:`~numpy.linspace` - :py:func:`~torch.linspace` - :py:func:`~megengine.functional.linspace` - * - :py:func:`~numpy.eye` - :py:func:`~torch.eye` - :py:func:`~megengine.functional.eye` - * - :py:func:`~numpy.zeros` - :py:func:`~torch.zeros` - :py:func:`~megengine.functional.zeros` - * - :py:func:`~numpy.zeros_like` - :py:func:`~torch.zeros_like` - :py:func:`~megengine.functional.zeros_like` - * - :py:func:`~numpy.ones` - :py:func:`~torch.ones` - :py:func:`~megengine.functional.ones` - * - :py:func:`~numpy.ones_like` - :py:func:`~torch.ones_like` - :py:func:`~megengine.functional.ones_like` - * - :py:func:`~numpy.full` - :py:func:`~torch.full` - :py:func:`~megengine.functional.full` - * - :py:func:`~numpy.full_like` - :py:func:`~torch.full_like` - :py:func:`~megengine.functional.full_like` - Manipulation Functions ~~~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:func:`~numpy.reshape` - :py:func:`~torch.reshape` - :py:func:`~megengine.functional.reshape` - * - :py:meth:`~numpy.ndarray.flatten` - :py:func:`~torch.flatten` - :py:func:`~megengine.functional.flatten` - * - :py:func:`~numpy.broadcast_to` - :py:func:`~torch.broadcast_to` / :py:meth:`~torch.Tensor.expand` - :py:func:`~megengine.functional.broadcast_to` - * - :py:func:`~numpy.expand_dims` - :py:func:`~torch.unsqueeze` - :py:func:`~megengine.functional.expand_dims` - * - :py:func:`~numpy.squeeze` - :py:func:`~torch.squeeze` - :py:func:`~megengine.functional.squeeze` - * - :py:func:`~numpy.concatenate` - :py:func:`~torch.cat` - :py:func:`~megengine.functional.concat` - * - :py:func:`~numpy.stack` - :py:func:`~torch.stack` - :py:func:`~megengine.functional.stack` - * - :py:func:`~numpy.split` - :py:func:`~torch.split` - :py:func:`~megengine.functional.split` - * - :py:func:`~numpy.tile` - :py:func:`~torch.tile` - :py:func:`~megengine.functional.tile` - * - :py:func:`~numpy.repeat` - :py:func:`~torch.repeat_interleave` - :py:func:`~megengine.functional.repeat` - * - :py:func:`~numpy.roll` - :py:func:`~torch.roll` - :py:func:`~megengine.functional.roll` - Arithmetic operations ~~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:data:`~numpy.add` - :py:func:`~torch.add` - :py:func:`~megengine.functional.add` - ``+`` operator * - :py:data:`~numpy.subtract` - :py:func:`~torch.sub` - :py:func:`~megengine.functional.sub` - ``-`` operator * - :py:data:`numpy.multiply` - :py:func:`~torch.mul` - :py:func:`~megengine.functional.mul` - ``*`` operator * - :py:data:`~numpy.divide` - :py:func:`~torch.div` - :py:func:`~megengine.functional.div` - ``/`` operator * - :py:data:`~numpy.floor_divide` - :py:func:`~torch.floor_divide` - :py:func:`~megengine.functional.floor_div` - ``//`` operator * - :py:data:`~numpy.negative` - :py:func:`~torch.neg` - :py:func:`~megengine.functional.neg` - * - :py:data:`~numpy.absolute` - :py:func:`~torch.abs` - :py:func:`~megengine.functional.abs` - * - :py:data:`~numpy.power` - :py:func:`~torch.pow` - :py:func:`~megengine.functional.pow` - ``**`` operator * - :py:data:`~numpy.mod` - :py:func:`~torch.remainder` - :py:func:`~megengine.functional.mod` - ``%`` operator * - :py:data:`~numpy.sqrt` - :py:func:`~torch.sqrt` - :py:func:`~megengine.functional.sqrt` - * - :py:data:`~numpy.square` - :py:func:`~torch.square` - :py:func:`~megengine.functional.square` - * - :py:data:`~numpy.sign` - :py:func:`~torch.sign` - :py:func:`~megengine.functional.sign` - * - :py:data:`~numpy.maximum` - :py:func:`~torch.maximum` - :py:func:`~megengine.functional.maximum` - * - :py:data:`~numpy.minimum` - :py:func:`~torch.minimum` - :py:func:`~megengine.functional.minimum` - * - :py:meth:`~numpy.ndarray.round` - :py:func:`~torch.round` - :py:func:`~megengine.functional.round` - * - :py:data:`~numpy.ceil` - :py:func:`~torch.ceil` - :py:func:`~megengine.functional.ceil` - * - :py:data:`~numpy.floor` - :py:func:`~torch.floor` - :py:func:`~megengine.functional.floor` - * - :py:func:`~numpy.clip` - :py:func:`~torch.clamp` - :py:func:`~megengine.functional.clip` - * - :py:data:`~numpy.exp` - :py:func:`~torch.exp` - :py:func:`~megengine.functional.exp` - * - :py:data:`~numpy.expm1` - :py:func:`~torch.expm1` - :py:func:`~megengine.functional.expm1` - * - :py:data:`~numpy.log` - :py:func:`~torch.log` - :py:func:`~megengine.functional.log` - * - :py:data:`~numpy.log1p` - :py:func:`~torch.log1p` - :py:func:`~megengine.functional.log1p` - * - :py:data:`~numpy.logaddexp` - :py:func:`~torch.logaddexp` - :py:func:`~megengine.functional.logaddexp` - Trigonometric functions ~~~~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:data:`~numpy.sin` - :py:func:`~torch.sin` - :py:func:`~megengine.functional.sin` - * - :py:data:`~numpy.cos` - :py:func:`~torch.cos` - :py:func:`~megengine.functional.cos` - * - :py:data:`~numpy.tan` - :py:func:`~torch.tan` - :py:func:`~megengine.functional.tan` - * - :py:data:`~numpy.arcsin` - :py:func:`~torch.asin` - :py:func:`~megengine.functional.asin` - * - :py:data:`~numpy.arccos` - :py:func:`~torch.acos` - :py:func:`~megengine.functional.acos` - * - :py:data:`~numpy.arctan` - :py:func:`~torch.atan` - :py:func:`~megengine.functional.atan` - Hyperbolic functions ~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:data:`~numpy.sinh` - :py:func:`~torch.sinh` - :py:func:`~megengine.functional.sinh` - * - :py:data:`~numpy.cosh` - :py:func:`~torch.cosh` - :py:func:`~megengine.functional.cosh` - * - :py:data:`~numpy.tanh` - :py:func:`~torch.tanh` - :py:func:`~megengine.functional.tanh` - * - :py:data:`~numpy.arcsinh` - :py:func:`~torch.asinh` - :py:func:`~megengine.functional.asinh` - * - :py:data:`~numpy.arccosh` - :py:func:`~torch.acosh` - :py:func:`~megengine.functional.acosh` - * - :py:data:`~numpy.arctanh` - :py:func:`~torch.atanh` - :py:func:`~megengine.functional.atanh` - Bit operations ~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:data:`~numpy.left_shift` - Not Found - :py:func:`~megengine.functional.left_shift` - ``<<`` operator * - :py:data:`~numpy.right_shift` - Not Found - :py:func:`~megengine.functional.right_shift` - ``>>`` operator Logic functions ~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:data:`~numpy.isnan` - :py:func:`~torch.isnan` - :py:func:`~megengine.functional.isnan` - * - :py:data:`~numpy.isinf` - :py:func:`~torch.isinf` - :py:func:`~megengine.functional.isinf` - * - :py:data:`~numpy.logical_and` - Not Found - :py:func:`~megengine.functional.logical_and` - ``&`` operator * - :py:data:`~numpy.logical_not` - Not Found - :py:func:`~megengine.functional.logical_not` - ``~`` operator * - :py:data:`~numpy.logical_or` - Not Found - :py:func:`~megengine.functional.logical_or` - ``|`` operator * - :py:data:`~numpy.logical_xor` - Not Found - :py:func:`~megengine.functional.logical_xor` - ``^`` operator * - :py:data:`~numpy.equal` - :py:func:`~torch.equal` - :py:func:`~megengine.functional.equal` - * - :py:data:`~numpy.not_equal` - :py:func:`~torch.not_equal` - :py:func:`~megengine.functional.not_equal` - * - :py:data:`~numpy.less` - :py:func:`~torch.less` - :py:func:`~megengine.functional.less` - * - :py:data:`~numpy.less_equal` - :py:func:`~torch.less_equal` - :py:func:`~megengine.functional.less_equal` - * - :py:data:`~numpy.greater` - :py:func:`~torch.greater` - :py:func:`~megengine.functional.greater` - * - :py:data:`~numpy.greater_equal` - :py:func:`~torch.greater_equal` - :py:func:`~megengine.functional.greater_equal` - Statistical Functions ~~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:func:`~numpy.sum` - :py:func:`~torch.sum` - :py:func:`~megengine.functional.sum` - * - :py:func:`~numpy.prod` - :py:func:`~torch.prod` - :py:func:`~megengine.functional.prod` - * - :py:func:`~numpy.mean` - :py:func:`~torch.mean` - :py:func:`~megengine.functional.mean` - * - :py:meth:`~numpy.ndarray.min` - :py:func:`~torch.min` - :py:func:`~megengine.functional.min` - * - :py:meth:`~numpy.ndarray.max` - :py:func:`~torch.max` - :py:func:`~megengine.functional.max` - * - :py:func:`~numpy.var` - :py:func:`~torch.var` - :py:func:`~megengine.functional.var` - * - :py:func:`~numpy.std` - :py:func:`~torch.std` - :py:func:`~megengine.functional.std` - * - :py:func:`~numpy.cumsum` - :py:func:`~torch.cumsum` - :py:func:`~megengine.functional.cumsum` - Linear Algebra Functions ~~~~~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:func:`~numpy.transpose` - :py:func:`~torch.transpose` - :py:func:`~megengine.functional.transpose` - * - :py:func:`~numpy.dot` - :py:func:`~torch.dot` - :py:func:`~megengine.functional.dot` - * - :py:func:`~numpy.linalg.inv` - :py:func:`~torch.linalg.inv` - :py:func:`~megengine.functional.matinv` - * - :py:data:`~numpy.matmul` - :py:func:`~torch.matmul` - :py:func:`~megengine.functional.matmul` - * - :py:func:`~numpy.linalg.svd` - :py:func:`~torch.linalg.svd` - :py:func:`~megengine.functional.svd` - * - :py:func:`~numpy.linalg.norm` - :py:func:`~torch.norm` - :py:func:`~megengine.functional.norm` - Indexing Functions ~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:func:`~numpy.take_along_axis` - :py:func:`~torch.gather` - :py:func:`~megengine.functional.gather` - * - :py:func:`~numpy.put_along_axis` - :py:func:`~torch.scatter` - :py:func:`~megengine.functional.scatter` - * - :py:func:`~numpy.where` - :py:func:`~torch.where` - :py:func:`~megengine.functional.where` / :py:func:`~megengine.functional.cond_take` - 取决于传参情况 Searching Functions ~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:func:`~numpy.argmin` - :py:func:`~torch.argmin` - :py:func:`~megengine.functional.argmin` - * - :py:func:`~numpy.argmax` - :py:func:`~torch.argmax` - :py:func:`~megengine.functional.argmax` - Sorting Functions ~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - NumPy - Pytorch - MegEngine - Comment * - :py:func:`~numpy.argsort` - :py:func:`~torch.argsort` - :py:func:`~megengine.functional.argsort` - * - :py:func:`~numpy.sort` - :py:func:`~torch.sort` - :py:func:`~megengine.functional.sort` - NN Funtional Operations ----------------------- Convolution functions ~~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:func:`~torch.nn.functional.conv1d` - :py:func:`~megengine.functional.nn.conv1d` - * - :py:func:`~torch.nn.functional.conv2d` - :py:func:`~megengine.functional.nn.conv2d` - * - :py:func:`~torch.nn.functional.conv3d` - :py:func:`~megengine.functional.nn.conv3d` - * - :py:func:`~torch.nn.functional.conv_transpose1d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.conv_transpose2d` - :py:func:`~megengine.functional.nn.conv_transpose2d` - * - :py:func:`~torch.nn.functional.conv_transpose3d` - :py:func:`~megengine.functional.nn.conv_transpose3d` - * - local_conv2d - :py:func:`~megengine.functional.nn.local_conv2d` - * - deformable_conv2d - :py:func:`~megengine.functional.nn.deformable_conv2d` - * - :py:func:`~torch.nn.functional.unfold` - :py:func:`~megengine.functional.nn.sliding_window` - * - :py:func:`~torch.nn.functional.fold` - :py:func:`~megengine.functional.nn.sliding_window_transpose` - Pooling functions ~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:func:`~torch.nn.functional.avg_pool1d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.avg_pool2d` - :py:func:`~megengine.functional.nn.avg_pool2d` - * - :py:func:`~torch.nn.functional.avg_pool3d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.max_pool1d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.max_pool2d` - :py:func:`~megengine.functional.nn.max_pool2d` - * - :py:func:`~torch.nn.functional.max_pool3d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.max_unpool1d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.max_unpool2d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.max_unpool3d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.lp_pool1d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.lp_pool2d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.adaptive_max_pool1d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.adaptive_max_pool2d` - :py:func:`~megengine.functional.nn.adaptive_max_pool2d` - * - :py:func:`~torch.nn.functional.adaptive_max_pool3d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.adaptive_avg_pool1d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.adaptive_avg_pool2d` - :py:func:`~megengine.functional.nn.adaptive_avg_pool2d` - * - :py:func:`~torch.nn.functional.adaptive_avg_pool3d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.fractional_max_pool2d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.fractional_max_pool3d` - :ref:`not-implemented` - Non-linear activation functions ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:func:`~torch.nn.functional.threshold` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.relu` - :py:func:`~megengine.functional.nn.relu` - * - :py:func:`~torch.nn.functional.hardtanh` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.hardswish` - :py:func:`~megengine.functional.nn.hswish` - * - :py:func:`~torch.nn.functional.relu6` - :py:func:`~megengine.functional.nn.relu6` - * - :py:func:`~torch.nn.functional.elu` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.selu` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.celu` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.leaky_relu` - :py:func:`~megengine.functional.nn.leaky_relu` - * - :py:func:`~torch.nn.functional.prelu` - :py:func:`~megengine.functional.nn.prelu` - * - :py:func:`~torch.nn.functional.rrelu` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.glu` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.gelu` - :py:func:`~megengine.functional.nn.gelu` - * - :py:func:`~torch.nn.functional.logsigmoid` - :py:func:`~megengine.functional.nn.logsigmoid` - * - :py:func:`~torch.nn.functional.hardshrink` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.tanhshrink` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.softsign` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.softplus` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.softmin` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.softmax` - :py:func:`~megengine.functional.nn.softmax` - * - :py:func:`~torch.nn.functional.softshrink` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.gumbel_softmax` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.log_softmax` - :py:func:`~megengine.functional.nn.logsoftmax` - * - :py:func:`~torch.nn.functional.sigmoid` - :py:func:`~megengine.functional.nn.sigmoid` - * - :py:func:`~torch.nn.functional.hardsigmoid` - :py:func:`~megengine.functional.nn.hsigmoid` - * - :py:func:`~torch.nn.functional.silu` - :py:func:`~megengine.functional.nn.silu` - Normalization functions ~~~~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:func:`~torch.nn.functional.batch_norm` - :py:func:`~megengine.functional.nn.batch_norm` - * - :py:func:`~torch.nn.functional.group_norm` - :ref:`not-implemented` - See :py:func:`~megengine.module.GroupNorm` * - :py:func:`~torch.nn.functional.instance_norm` - :ref:`not-implemented` - See :py:func:`~megengine.module.InstanceNorm` * - :py:func:`~torch.nn.functional.layer_norm` - :ref:`not-implemented` - See :py:func:`~megengine.module.LayerNorm` * - :py:func:`~torch.nn.functional.local_response_norm` - :ref:`not-implemented` - See :py:func:`~megengine.module.LocalResponseNorm` * - :py:func:`~torch.nn.functional.normalize` - :py:func:`~megengine.functional.normalize` - Linear functions ~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:func:`~torch.nn.functional.linear` - :py:func:`~megengine.functional.nn.linear` - * - :py:func:`~torch.nn.functional.bilinear` - :ref:`not-implemented` - Dropout functions ~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:func:`~torch.nn.functional.dropout` - :py:func:`~megengine.functional.nn.dropout` - * - :py:func:`~torch.nn.functional.alpha_dropout` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.feature_alpha_dropout` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.dropout2d` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.dropout3d` - :ref:`not-implemented` - Sparse functions ~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:func:`~torch.nn.functional.embedding` - :py:func:`~megengine.functional.nn.embedding` - * - :py:func:`~torch.nn.functional.embedding_bag` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.one_hot` - :py:func:`~megengine.functional.nn.one_hot` - Metric functions ~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:func:`~torch.nn.functional.pairwise_distance` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.cosine_similarity` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.pdist` - :ref:`not-implemented` - Loss functions ~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:func:`~torch.nn.functional.binary_cross_entropy` - :py:func:`~megengine.functional.nn.binary_cross_entropy` - * - :py:func:`~torch.nn.functional.binary_cross_entropy_with_logits` - :py:func:`~megengine.functional.nn.binary_cross_entropy` - * - :py:func:`~torch.nn.functional.poisson_nll_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.cosine_embedding_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.cross_entropy` - :py:func:`~megengine.functional.nn.cross_entropy` - * - :py:func:`~torch.nn.functional.ctc_loss` - :py:func:`~megengine.functional.nn.ctc_loss` - * - :py:func:`~torch.nn.functional.gaussian_nll_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.hinge_embedding_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.kl_div` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.l1_loss` - :py:func:`~megengine.functional.nn.l1_loss` - * - :py:func:`~torch.nn.functional.mse_loss` - :py:func:`~megengine.functional.nn.square_loss` - * - :py:func:`~torch.nn.functional.margin_ranking_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.multilabel_margin_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.multilabel_soft_margin_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.multi_margin_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.nll_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.huber_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.smooth_l1_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.soft_margin_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.triplet_margin_loss` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.triplet_margin_with_distance_loss` - :ref:`not-implemented` - NN Module --------- .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.parameter.Parameter` - :py:class:`~megengine.Parameter` - Containers ~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.Module` - :py:class:`~megengine.module.Module` - * - :py:class:`~torch.nn.Sequential` - :py:class:`~megengine.module.Sequential` - * - :py:class:`~torch.nn.ModuleList` - MegEngine 原生支持 - * - :py:class:`~torch.nn.ModuleDict` - MegEngine 原生支持 - * - :py:class:`~torch.nn.ParameterList` - MegEngine 原生支持 - * - :py:class:`~torch.nn.ParameterDict` - MegEngine 原生支持 - Initialization ~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:func:`~torch.nn.init.calculate_gain` - :py:class:`~megengine.module.init.calculate_gain` - * - _calculate_fan_in_and_fan_out - :py:class:`~megengine.module.init.calculate_fan_in_and_fan_out` - * - _calculate_correct_fan - :py:class:`~megengine.module.init.calculate_correct_fan` - * - :py:func:`~torch.nn.init.uniform_` - :py:class:`~megengine.module.init.uniform_` - * - :py:func:`~torch.nn.init.normal_` - :py:class:`~megengine.module.init.normal_` - * - :py:func:`~torch.nn.init.constant_` - :py:class:`~megengine.module.init.fill_` - * - :py:func:`~torch.nn.init.ones_` - :py:class:`~megengine.module.init.ones_` - * - :py:func:`~torch.nn.init.zeros_` - :py:class:`~megengine.module.init.zeros_` - * - :py:func:`~torch.nn.init.eye_` - :ref:`not-implemented` - * - :py:func:`~torch.nn.init.dirac_` - :ref:`not-implemented` - * - :py:func:`~torch.nn.init.xavier_uniform_` - :py:class:`~megengine.module.init.xavier_uniform_` - * - :py:func:`~torch.nn.init.xavier_normal_` - :py:class:`~megengine.module.init.xavier_normal_` - * - :py:func:`~torch.nn.init.kaiming_uniform_` - :py:class:`~megengine.module.init.msra_uniform_` - * - :py:func:`~torch.nn.init.kaiming_normal_` - :py:class:`~megengine.module.init.msra_normal_` - * - :py:func:`~torch.nn.init.orthogonal_` - :ref:`not-implemented` - * - :py:func:`~torch.nn.init.sparse_` - :ref:`not-implemented` - Convolution Layers ~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.Conv1d` - :py:class:`~megengine.module.Conv1d` - :ref:`差异对比 ` * - :py:class:`~torch.nn.Conv2d` - :py:class:`~megengine.module.Conv2d` - :ref:`差异对比 ` * - :py:class:`~torch.nn.Conv3d` - :py:class:`~megengine.module.Conv3d` - :ref:`差异对比 ` * - :py:class:`~torch.nn.ConvTranspose1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.ConvTranspose2d` - :py:class:`~megengine.module.ConvTranspose2d` - :ref:`差异对比 ` * - :py:class:`~torch.nn.ConvTranspose3d` - :py:class:`~megengine.module.ConvTranspose3d` - :ref:`差异对比 ` * - :py:class:`~torch.nn.LazyConv1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LazyConv2d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LazyConv3d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LazyConvTranspose1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LazyConvTranspose2d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LazyConvTranspose3d` - :ref:`not-implemented` - * - LocalConv2d - :py:class:`~megengine.module.LocalConv2d` - * - DeformableConv2d - :py:class:`~megengine.module.DeformableConv2d` - * - :py:class:`~torch.nn.Conv1d` - :py:class:`~megengine.module.Conv1d` - * - :py:class:`~torch.nn.Unfold` - :py:class:`~megengine.module.SlidingWindowTranspose` - * - :py:class:`~torch.nn.Fold` - :py:class:`~megengine.module.SlidingWindow` - Pooling layers ~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.MaxPool1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.MaxPool2d` - :py:class:`~megengine.module.MaxPool2d` - :ref:`差异对比 ` * - :py:class:`~torch.nn.MaxPool3d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.MaxUnpool1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.MaxUnpool2d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.MaxUnpool3d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.AvgPool1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.AvgPool2d` - :py:class:`~megengine.module.AvgPool2d` - :ref:`差异对比 ` * - :py:class:`~torch.nn.AvgPool3d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.FractionalMaxPool2d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.FractionalMaxPool3d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LPPool1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LPPool2d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.AdaptiveMaxPool1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.AdaptiveMaxPool2d` - :py:class:`~megengine.module.AdaptiveMaxPool2d` - :ref:`差异对比 ` * - :py:class:`~torch.nn.AdaptiveMaxPool3d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.AdaptiveAvgPool1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.AdaptiveAvgPool2d` - :py:class:`~megengine.module.AdaptiveAvgPool2d` - :ref:`差异对比 ` * - :py:class:`~torch.nn.AdaptiveAvgPool3d` - :ref:`not-implemented` - Padding Layers ~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.ReflectionPad1d` - :py:class:`~megengine.module.Pad` - mode = REFLECT * - :py:class:`~torch.nn.ReflectionPad2d` - :py:class:`~megengine.module.Pad` - mode = REFLECT * - :py:class:`~torch.nn.ReflectionPad3d` - :py:class:`~megengine.module.Pad` - mode = REFLECT * - :py:class:`~torch.nn.ReplicationPad1d` - :py:class:`~megengine.module.Pad` - mode = EDGE * - :py:class:`~torch.nn.ReplicationPad2d` - :py:class:`~megengine.module.Pad` - mode = EDGE * - :py:class:`~torch.nn.ReplicationPad3d` - :py:class:`~megengine.module.Pad` - mode = EDGE * - :py:class:`~torch.nn.ZeroPad2d` - :py:class:`~megengine.module.Pad` - mode = CONSTANT * - :py:class:`~torch.nn.ConstantPad1d` - :py:class:`~megengine.module.Pad` - mode = CONSTANT * - :py:class:`~torch.nn.ConstantPad2d` - :py:class:`~megengine.module.Pad` - mode = CONSTANT * - :py:class:`~torch.nn.ConstantPad3d` - :py:class:`~megengine.module.Pad` - mode = CONSTANT Non-linear Activations ~~~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.ELU` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Hardshrink` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Hardsigmoid` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Hardtanh` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Hardswish` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LeakyReLU` - :py:class:`~megengine.module.LeakyReLU` - :ref:`差异对比 ` * - :py:class:`~torch.nn.LogSigmoid` - :ref:`not-implemented` - * - :py:class:`~torch.nn.MultiheadAttention` - :ref:`not-implemented` - * - :py:class:`~torch.nn.PReLU` - :py:class:`~megengine.module.PReLU` - * - :py:class:`~torch.nn.ReLU` - :py:class:`~megengine.module.ReLU` - :ref:`差异对比 ` * - :py:class:`~torch.nn.ReLU6` - :ref:`not-implemented` - * - :py:class:`~torch.nn.RReLU` - :ref:`not-implemented` - * - :py:class:`~torch.nn.SELU` - :ref:`not-implemented` - * - :py:class:`~torch.nn.CELU` - :ref:`not-implemented` - * - :py:class:`~torch.nn.GELU` - :py:class:`~megengine.module.GELU` - * - :py:class:`~torch.nn.Sigmoid` - :py:class:`~megengine.module.Sigmoid` - :ref:`差异对比 ` * - :py:class:`~torch.nn.SiLU` - :py:class:`~megengine.module.SiLU` - * - :py:class:`~torch.nn.Softplus` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Softshrink` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Softsign` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Tanh` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Tanhshrink` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Threshold` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Softmin` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Softmax` - :py:class:`~megengine.module.Softmax` - :ref:`差异对比 ` * - :py:class:`~torch.nn.Softmax2d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LogSoftmax` - :ref:`not-implemented` - * - :py:class:`~torch.nn.AdaptiveLogSoftmaxWithLoss` - :ref:`not-implemented` - Normalization Layers ~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.BatchNorm1d` - :py:class:`~megengine.module.BatchNorm1d` - :ref:`差异对比 ` * - :py:class:`~torch.nn.BatchNorm2d` - :py:class:`~megengine.module.BatchNorm2d` - :ref:`差异对比 ` * - :py:class:`~torch.nn.BatchNorm3d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LazyBatchNorm1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LazyBatchNorm2d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LazyBatchNorm3d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.GroupNorm` - :py:class:`~megengine.module.GroupNorm` - * - :py:class:`~torch.nn.SyncBatchNorm` - :py:class:`~megengine.module.SyncBatchNorm` - * - :py:class:`~torch.nn.InstanceNorm1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.InstanceNorm2d` - :py:class:`~megengine.module.InstanceNorm` - :ref:`差异对比 ` * - :py:class:`~torch.nn.InstanceNorm3d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LayerNorm` - :py:class:`~megengine.module.LayerNorm` - :ref:`差异对比 ` * - :py:class:`~torch.nn.LocalResponseNorm` - :ref:`not-implemented` - Recurrent Layers ~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.RNNBase` - :py:class:`~megengine.module.RNNBase` - * - :py:class:`~torch.nn.RNN` - :py:class:`~megengine.module.RNN` - :ref:`差异对比 ` * - :py:class:`~torch.nn.LSTM` - :py:class:`~megengine.module.LSTM` - :ref:`差异对比 ` * - :py:class:`~torch.nn.GRU` - :ref:`not-implemented` - * - :py:class:`~torch.nn.RNNCell` - :py:class:`~megengine.module.RNNCell` - :ref:`差异对比 ` * - :py:class:`~torch.nn.LSTMCell` - :py:class:`~megengine.module.LSTMCell` - :ref:`差异对比 ` * - :py:class:`~torch.nn.GRUCell` - :ref:`not-implemented` - Transformer Layers ~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.Transformer` - :ref:`not-implemented` - * - :py:class:`~torch.nn.TransformerEncoder` - :ref:`not-implemented` - * - :py:class:`~torch.nn.TransformerDecoder` - :ref:`not-implemented` - * - :py:class:`~torch.nn.TransformerEncoderLayer` - :ref:`not-implemented` - * - :py:class:`~torch.nn.TransformerDecoderLayer` - :ref:`not-implemented` - Linear Layers ~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.Identity` - :py:class:`~megengine.module.Identity` - * - :py:class:`~torch.nn.Linear` - :py:class:`~megengine.module.Linear` - :ref:`差异对比 ` * - :py:class:`~torch.nn.Bilinear` - :ref:`not-implemented` - * - :py:class:`~torch.nn.LazyLinear` - :ref:`not-implemented` - Dropout Layers ~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.Dropout` - :py:class:`~megengine.module.Dropout` - :ref:`差异对比 ` * - :py:class:`~torch.nn.Dropout1d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Dropout2d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.Dropout3d` - :ref:`not-implemented` - * - :py:class:`~torch.nn.AlphaDropout` - :ref:`not-implemented` - * - :py:class:`~torch.nn.FeatureAlphaDropout` - :ref:`not-implemented` - Sparse Layers ~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.Embedding` - :py:class:`~megengine.module.Embedding` - :ref:`差异对比 ` * - :py:class:`~torch.nn.EmbeddingBag` - :ref:`not-implemented` - Distance Functions ~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.CosineSimilarity` - :ref:`not-implemented` - * - :py:class:`~torch.nn.PairwiseDistance` - :ref:`not-implemented` - Loss Functions ~~~~~~~~~~~~~~ .. seealso:: 请参考 loss function 的 functional 实现。 .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:class:`~torch.nn.L1Loss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.MSELoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.CrossEntropyLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.CTCLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.NLLLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.PoissonNLLLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.KLDivLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.BCELoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.BCEWithLogitsLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.MarginRankingLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.HingeEmbeddingLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.MultiLabelMarginLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.SmoothL1Loss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.SoftMarginLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.MultiLabelSoftMarginLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.CosineEmbeddingLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.MultiMarginLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.TripletMarginLoss` - :ref:`not-implemented` - * - :py:class:`~torch.nn.TripletMarginWithDistanceLoss` - :ref:`not-implemented` - Vision functions ---------------- .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - :py:func:`~torch.nn.functional.pixel_shuffle` - :py:func:`~megengine.functional.nn.pixel_shuffle` - :ref:`差异对比 ` * - :py:func:`~torch.nn.functional.pixel_unshuffle` - :ref:`not-implemented` - * - :py:func:`~torch.nn.functional.pad` - :py:func:`~megengine.functional.nn.pad` - * - :py:func:`~torch.nn.functional.interpolate` - :py:func:`~megengine.functional.nn.interpolate` - * - :py:func:`~torch.nn.functional.upsample` - :py:func:`~megengine.functional.nn.interpolate` - * - :py:func:`~torch.nn.functional.upsample_nearest` - :py:func:`~megengine.functional.nn.interpolate` - * - :py:func:`~torch.nn.functional.upsample_bilinear` - :py:func:`~megengine.functional.nn.interpolate` - * - :py:func:`~torch.nn.functional.grid_sample` - :py:func:`~megengine.functional.nn.remap` - * - :py:func:`~torch.nn.functional.affine_grid` - :py:func:`~megengine.functional.nn.warp_affine` - * - :py:func:`~torchvision.ops.nms` - :py:func:`~megengine.functional.nn.nms` - * - :py:func:`~torchvision.ops.roi_align` - :py:func:`~megengine.functional.nn.roi_align` - * - :py:func:`~torchvision.ops.roi_pool` - :py:func:`~megengine.functional.nn.roi_pooling` - OpenCV Python Package ~~~~~~~~~~~~~~~~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - cvtColor - :py:func:`~megengine.functional.nn.cvt_color` - * - resize - :py:func:`~megengine.functional.nn.interpolate` - * - remap - :py:func:`~megengine.functional.nn.remap` - * - warpAffine - :py:func:`~megengine.functional.nn.warp_affine` - * - warpPerspective - :py:func:`~megengine.functional.nn.warp_perspective` - NVIDIA ~~~~~~ .. list-table:: :header-rows: 1 * - Pytorch - MegEngine - Comment * - correlation - :py:func:`~megengine.functional.nn.correlation` - * - nvof - :py:func:`~megengine.functional.nn.nvof` - .. _not-implemented: Not Implemeted -------------- .. note:: 一些 API 在 MegEngine 中可能还没有实现，但所有的 API 并不是一开始就被设计出来的。 我们可以像搭积木一样，利用已经存在的基础 API 来组合出 MegEngine 中尚未提供的接口。 比如 “如何实现 :py:func:`~torch.roll` ” 这个问题，可以使用 :py:func:`~.functional.split` 和 :py:func:`~.functional.concat` 拼接出来： .. code-block:: python import megengine.functional as F def roll(x, shifts, axis): shp = x.shape dim = len(shp) if isinstance(shifts, int): assert isinstance(axis, int) shifts = [shifts] axis = [axis] assert len(shifts) == len(axis) y = x for i in range(len(shifts)): axis_ = axis[i] shift_ = shifts[i] axis_t_ = axis_ + dim if axis_ < 0 else axis_ assert ( dim > axis_t_ >= 0 ), "axis out of range (expected to be in range of [{}, {}], but got {})".format( -dim, dim - 1, axis_ ) if shift_ == 0: continue size = shp[axis_t_] if shift_ > 0: a, b = F.split(y, [size - shift_,], axis=axis_t_) else: a, b = F.split(y, [-shift_,], axis=axis_t_) y = F.concat((b, a), axis=axis_t_) return y 除此之外，你可以尝试在 GitHub Issues 或论坛中针对 API 问题发起求助。 我们也欢迎你将自己实现的 API 以 Pull Request 的形式提交到 MegEngine 代码库中来～ .. note:: 对于缺失的 Loss Funtions 算子，大都可自行设计实现。