Skip to content

使用 autograd.Function 扩展 torch.func

译者:片刻小哥哥

项目地址:https://pytorch.apachecn.org/2.0/docs/notes/extending.func

原始地址:https://pytorch.org/docs/stable/notes/extending.func.html

所以你想将 torch.autograd.Functiontorch.func 转换为 torch.vmap()torch.func.grad()

有两个主要用例:

  • 您希望调用不包含 PyTorch 操作的代码并使其与函数转换一起使用。也就是说, torch.autograd.Function 的向前/向后等调用来自其他系统(如 C++、CUDA)的函数, numpy.
  • 您希望指定自定义渐变规则,例如 JAX 的 custom_vjp/custom_jvp

PyTorch 将这两个概念结合到 torch.autograd.Function 中。

基本用法

本指南假设您熟悉 扩展 torch.autograd ,它解释了如何使用 torch.autograd.Function

torch.autograd.Function 可以有一个 forward() 接受 ctx 对象,或者它可以有单独的 forward() (不接受 ctx )和修改 ctxsetup_context() 静态方法目的。

函数转换仅支持后者:

  • forward() 是执行该操作的代码,它不应接受 ctx 对象。
  • setup_context(ctx,inputs,output) 是您可以在 ctx 上调用方法的代码。您应该在此处保存向后tensor(通过调用 ctx.save_for_backward(*tensors) ),或保存非tensor(通过将它们分配给 ctx 对象)。

因为 setup_context() 只接受 inputsoutput ,所以可以保存的唯一数量是输入或输出中的对象(例如 Tensors)或从它们派生的数量(例如 Tensor.shape ).如果您希望从 Function.forward() 向后,那么您需要将其作为 forward() 以便将其传递给 setup_context()

根据变换,

为了使 torch.autograd.Function 可以与函数转换任意组合,我们建议除 forward()setup_context() 必须是可转换的:也就是说,它们必须仅包含 PyTorchoperators 或调用其他 torch.autograd.Function (可能会调用 C++/CUDA/等)。

让我们看一些常见用例的示例。

示例 1:autograd.Function 调用另一个系统

常见的情况是 torch.autograd.Function 具有forward() 和backward() 调用另一个系统(如C++、CUDA、numpy、triton)。

import torch
import numpy as np

def to_numpy(tensor):
    return tensor.cpu().numpy()

class NumpySort(torch.autograd.Function):
    # Note that forward does not take ctx
    @staticmethod
    def forward(x, dim):
        device = x.device
        x = to_numpy(x)
        ind = np.argsort(x, axis=dim)
        ind_inv = np.argsort(ind, axis=dim)
        result = np.take_along_axis(x, ind, axis=dim)
        # Any intermediates to be saved in backward must be returned as
        # outputs.
        return (
            # The desired output
            torch.tensor(result, device=device),
            # intermediate to save for backward
            torch.tensor(ind, device=device),
            # intermediate to save for backward
            torch.tensor(ind_inv, device=device),
        )

    # setup_context is responsible for calling methods and/or assigning to
    # the ctx object. Please do not do additional compute (e.g. add
    # Tensors together) in setup_context.
    @staticmethod
    def setup_context(ctx, inputs, output):
        x, dim = inputs
        # Note that output is whatever you returned from forward.
        # If you returned multiple values, then output is a Tuple of multiple values.
        # If you returned a single Tensor, then output is a Tensor.
        # If you returned a Tuple with a single Tensor, then output is a
        # Tuple with a single Tensor.
        _, ind, ind_inv = output
        ctx.mark_non_differentiable(ind, ind_inv)
        # Tensors must be saved via ctx.save_for_backward. Please do not
        # assign them directly onto the ctx object.
        ctx.save_for_backward(ind, ind_inv)
        # Non-tensors may be saved by assigning them as attributes on the ctx object.
        ctx.dim = dim

    @staticmethod
    def backward(ctx, grad_output, _0, _1):
        # For the autograd.Function to be arbitrarily composable with function
        # transforms, all staticmethod other than forward and setup_context
        # must be implemented in a "transformable" way; that is, they must
        # only consist of PyTorch operations or autograd.Function.
        #
        # For example, this allows us to do double backwards and/or compute
        # second order gradients.
        #
        # We've written the backward pass of NumpySort in terms of another
        # autograd.Function, NumpyTake.
        ind, ind_inv = ctx.saved_tensors
        return NumpyTake.apply(grad_output, ind_inv, ind, ctx.dim), None

class NumpyTake(torch.autograd.Function):
    @staticmethod
    def forward(x, ind, ind_inv, dim):
        device = x.device
        x = to_numpy(x)
        ind = to_numpy(ind)
        return torch.tensor(np.take_along_axis(x, ind, dim), device=device)

    @staticmethod
    def setup_context(ctx, inputs, output):
        x, ind, ind_inv, dim = inputs
        ctx.save_for_backward(ind, ind_inv)
        ctx.dim = dim

    @staticmethod
    def backward(ctx, grad_output):
        ind, ind_inv = ctx.saved_tensors
        result = NumpyTake.apply(grad_output, ind_inv, ind, ctx.dim)
        return result, None, None, None

现在,为了更容易使用“NumpySort”(隐藏中间结果作为输出,并允许默认参数和 kwargs),我们创建一个调用它的新函数:

def numpy_sort(x, dim=-1):
    result, _, _ = NumpySort.apply(x, dim)
    return result

这是一个健全性检查:

x = torch.randn(2, 3)
grad_x = torch.func.grad(lambda x: numpy_sort(x).sum())(x)
assert torch.allclose(grad_x, torch.ones_like(x))

示例 2:autograd.Function 指定自定义渐变规则

另一个常见的情况是使用 PyTorchoperations 实现的 torch.autograd.Function。 PyTorch 能够自动计算 PyTorch 操作的梯度,但也许我们希望自定义梯度的计算方式。我们可能想要与 PyTorch 提供的不同的自定义向后的一些原因是:

  • 提高数值稳定性
  • 改变后向的性能特征
  • 改变边缘情况的处理方式(例如 nans、inf)
  • 修改梯度(例如梯度裁剪)

这是函数 y = x ** 3torch.autograd.Function 示例,其中我们更改性能特征(通常在后向传递过程中发生的一些计算,即计算 dx,发生在前向传递中)。

class MyCube(torch.autograd.Function):
    @staticmethod
    def forward(x):
        result = x ** 3
        # In regular PyTorch, if we had just run y = x ** 3, then the backward
        # pass computes dx = 3 * x ** 2. In this autograd.Function, we've done
        # that computation here in the forward pass instead.
        dx = 3 * x ** 2
        return result, dx

    @staticmethod
    def setup_context(ctx, inputs, output):
        x, = inputs
        result, dx = output
        ctx.save_for_backward(x, dx)

    @staticmethod
    def backward(ctx, grad_output, grad_dx):
        x, dx = ctx.saved_tensors
        # In order for the autograd.Function to work with higher-order
        # gradients, we must add the gradient contribution of `dx`.
        result = grad_output * dx + grad_dx * 6 * x
        return result

现在,为了更容易使用“NumpySort”(并隐藏中间结果作为输出),我们创建一个调用它的新函数:

def my_cube(x):
    result, _ = MyCube.apply(x)
    return result

这是计算二阶梯度的健全性检查:

x = torch.randn([])
ggx = torch.func.grad(torch.func.grad(my_cube))(x)
assert torch.allclose(ggx, 6 * x)

限制和陷阱

警告

请仔细阅读 torch.autograd.Function 与 torch.func 转换的这些限制。我们无法捕获许多这样的情况和错误,因此它们会导致未定义的行为。

请不要将正在转换的tensor、haverequires_grad=True 或双tensor捕获到 torch.autograd.Function 。完全安全的方法是确保在 torch.autograd.Function 的任何方法中使用的 onlyTensors 必须直接作为输入传递(或通过 ctx 对象),而不是来自 torch.autograd.Function 外部。

torch.autograd.Function 不处理 pytree 中的tensor(任意嵌套的 Python 数据结构,可能包含也可能不包含tensor)。对于要由 autograd 跟踪的tensor,必须将它们作为参数直接传递给 torch.autograd.Function 。这与 jax.{custom_vjp, custom_jvp} 形成对比,jax.{custom_vjp, custom_jvp} 接受 pytree。

请仅使用 save_for_backward()save_for_forward() 来保存tensor。请不要将tensor或tensor集合直接分配到 ctx 对象上 - 这些tensor将不会被跟踪

torch.vmap() 支持

要将 torch.autograd.Functiontorch.vmap() 一起使用,您必须:

自动生成 vmap 规则

如果您的 torch.autograd.Function 满足以下附加约束,那么我们就能够为其生成 vmap 规则。如果它不满足约束或者您想要在 vmap 下自定义行为,请手动定义 vmap 静态方法(请参阅下一节)。

警告

我们不容易优雅地检查以下约束和错误。违反约束可能会导致未定义的行为。

例子:

class MyCube(torch.autograd.Function):
    # Set generate_vmap_rule to True to ask PyTorch to automatically generate
    # a vmap rule.
    generate_vmap_rule = True

    @staticmethod
    def forward(x):
        result = x ** 3
        dx = 3 * x ** 2
        return result, dx

    @staticmethod
    def setup_context(ctx, inputs, output):
        x, = inputs
        result, dx = output
        ctx.save_for_backward(x, dx)

    @staticmethod
    def backward(ctx, grad_output, grad_dx):
        x, dx = ctx.saved_tensors
        result = grad_output * dx + grad_dx * 6 * x
        return result

def my_cube(x):
    result, dx = MyCube.apply(x)
    return result

x = torch.randn(3)
result = torch.vmap(my_cube)(x)
assert torch.allclose(result, x ** 3)

定义 vmap 静态方法

如果你的 torch.autograd.Function 调用另一个系统(如 NumPy、C++、CUDA、triton),那么得到它要与 torch.vmap() 或使用它的转换一起使用,您需要手动定义 vmap () 静态方法。

根据您要使用的转换和您的用例,您可能不需要添加 vmap() 静态方法到你的所有 torch.autograd.Function

我们建议确保您的所有 torch.autograd.Function 都支持 torch.vmap() 不过,特别是如果您正在编写第三方库并且您想要 torch.autograd.Functiontorch.func()转变。

从概念上讲, vmap staticmethod 负责定义 forward() 应该在 torch.vmap() 下运行。也就是说,它定义了如何转换 forward() 来运行在具有附加维度的输入上(被映射的维度)。这类似于在 PyTorch 操作上实现 torch.vmap() 的方式:对于每个操作,我们定义一个 vmap 规则(有时也称为“批处理规则”)。

以下是如何定义 vmap() 静态方法:

  • 签名是 vmap(info, in_dims: Tuple[Optional[int]], *args) ,其中 *argsforward().
  • vmap 静态方法负责定义 forward() 应该在 torch.vmap() 。也就是说,给定具有附加维度的输入(由 in_dims 指定),我们如何计算 forward() ?
  • 对于 args 中的每个 arg,in_dims 都有一个对应的 Optional[int] 。如果 arg 则为 None不是一个 Tensor,或者如果 arg 没有被 vmapped,否则,它是一个整数,指定正在 vmappedover 的 Tensor 的维度。
  • info 是可能有用的附加元数据的集合:info.batch_size 指定被 vmapped 的维度的大小,而 info.randomness 是传递给 torch.vmap() 的随机性选项。
  • vmap 静态方法的返回是 (output, out_dims) 的元组。与“in_dims”类似,“out_dims”应该与“output”具有相同的结构,并且每个输出包含一个“out_dim”,用于指定输出是否具有 vmappeddimension 以及它所在的索引。

例子:

def to_numpy(tensor):
    return tensor.cpu().numpy()

class NumpySort(torch.autograd.Function):
    @staticmethod
    def forward(x, dim):
        device = x.device
        x = to_numpy(x)
        ind = np.argsort(x, axis=dim)
        ind_inv = np.argsort(ind, axis=dim)
        result = np.take_along_axis(x, ind, axis=dim)
        return (
            torch.tensor(result, device=device),
            torch.tensor(ind, device=device),
            torch.tensor(ind_inv, device=device),
        )

    @staticmethod
    def setup_context(ctx, inputs, output):
        x, dim = inputs
        _, ind, ind_inv = output
        ctx.mark_non_differentiable(ind, ind_inv)
        ctx.save_for_backward(ind, ind_inv)
        ctx.dim = dim

    @staticmethod
    def backward(ctx, grad_output, _0, _1):
        return NumpyTake.apply(grad_output, ind_inv, ind, ctx.dim), None

    # The signature of the vmap staticmethod is:
    # vmap(info, in_dims: Tuple[Optional[int]], *args)
    # where *args is the same as the arguments to `forward`.
    @staticmethod
    def vmap(info, in_dims, x, dim):
        # For every input (x and dim), in_dims stores an Optional[int]
        # that is:
        # - None if the input is not being vmapped over or if the input
        # is not a Tensor
        # - an integer if the input is being vmapped over that represents
        # the index of the dimension being vmapped over.
        x_bdim, _ = in_dims

        # A "vmap rule" is the logic of how to perform the operation given
        # inputs with one additional dimension. In NumpySort, x has an
        # additional dimension (x_bdim). The vmap rule is simply
        # to call NumpySort again but pass it a different `dim`.
        x = x.movedim(x_bdim, 0)
        # Handle negative dims correctly
        dim = dim if dim >= 0 else dim + x.dim() - 1
        result = NumpySort.apply(x, dim + 1)

        # The vmap rule must return a tuple of two things
        # 1. the output. Should be the same amount of things
        # as returned by the forward().
        # 2. one Optional[int] for each output specifying if each output
        # is being vmapped over, and if so, the index of the
        # dimension being vmapped over.
        #
        # NumpySort.forward returns a Tuple of 3 Tensors. Since we moved the
        # dimension being vmapped over to the front of `x`, that appears at
        # dimension 0 of all outputs.
        # The return is (output, out_dims) -- output is a tuple of 3 Tensors
        # and out_dims is a Tuple of 3 Optional[int]
        return NumpySort.apply(x, dim + 1), (0, 0, 0)

class NumpyTake(torch.autograd.Function):
    @staticmethod
    def forward(x, ind, ind_inv, dim):
        device = x.device
        x = to_numpy(x)
        ind = to_numpy(ind)
        return torch.tensor(np.take_along_axis(x, ind, dim), device=device)

    @staticmethod
    def setup_context(ctx, inputs, output):
        x, ind, ind_inv, dim = inputs
        ctx.save_for_backward(ind, ind_inv)
        ctx.dim = dim

    @staticmethod
    def backward(ctx, grad_output):
        ind, ind_inv = ctx.saved_tensors
        result = NumpyTake.apply(grad_output, ind_inv, ind, ctx.dim)
        return result, None, None, None

    @staticmethod
    def vmap(info, in_dims, x, ind, ind_inv, dim):
        x_bdim, ind_bdim, ind_inv_bdim, _ = in_dims

        # The strategy is: expand {x, ind, ind_inv} to all have the dimension
        # being vmapped over.
        # Then, call back into NumpyTake(expanded_x, expanded_ind, expanded_ind_inv, new_dim).

        # Handle negative dims by wrapping them to be positive
        logical_dim = x.dim() if x_bdim is None else x_bdim - 1
        dim = dim if dim >= 0 else dim + logical_dim

        def maybe_expand_bdim_at_front(x, x_bdim):
            if x_bdim is None:
                return x.expand(info.batch_size, *x.shape)
            return x.movedim(x_bdim, 0)

        # If the Tensor doesn't have the dimension being vmapped over,
        # expand it out. Otherwise, move it to the front of the Tensor
        x = maybe_expand_bdim_at_front(x, x_bdim)
        ind = maybe_expand_bdim_at_front(ind, ind_bdim)
        ind_inv = maybe_expand_bdim_at_front(ind_inv, ind_inv_bdim)

        # The return is a tuple (output, out_dims). Since output is a Tensor,
        # then out_dims is an Optional[int](instead of being a Tuple).
        return NumpyTake.apply(x, ind, ind_inv, dim + 1), 0

def numpy_sort(x, dim=-1):
    result, _, _ = NumpySort.apply(x, dim)
    return result

x = torch.randn(2, 3)
result = torch.vmap(numpy_sort)(x)
assert torch.allclose(result, numpy_sort(result, 1))

笔记

vmap 静态方法应该旨在保留整个 Function 的语义。也就是说,(伪代码) grad(vmap(MyFunc)) 应该可以替换为 grad(map(MyFunc))

如果您的 autograd.Function 在向后传递中有任何自定义行为,请记住这一点。

笔记

为 PyTorch 能够通过以下方式生成 vmaprule 的 Function 编写自定义 vmap 静态方法是一个合法的用例generate_vmap_rule=True 。如果生成的 vmap 规则不具有您正在寻找的语义,您可能希望这样做。

torch.func.jvp() 支持

为了支持正向模式 AD,torch.autograd.Function 必须有一个 jvp() staticmethod。请参阅Forward mode AD了解详细信息。


我们一直在努力

apachecn/AiLearning

【布客】中文翻译组