{ "cells": [ { "cell_type": "markdown", "id": "903e2f76", "metadata": {}, "source": [ "# Whirlwind Tour\n", "\n", "\n", " \"Open\n", "\n", "\n", "## What is functorch?\n", "\n", "functorch is a library for [JAX](https://github.com/google/jax)-like composable function transforms in PyTorch.\n", "- A \"function transform\" is a higher-order function that accepts a numerical function and returns a new function that computes a different quantity.\n", "- functorch has auto-differentiation transforms (`grad(f)` returns a function that computes the gradient of `f`), a vectorization/batching transform (`vmap(f)` returns a function that computes `f` over batches of inputs), and others.\n", "- These function transforms can compose with each other arbitrarily. For example, composing `vmap(grad(f))` computes a quantity called per-sample-gradients that stock PyTorch cannot efficiently compute today.\n", "\n", "Furthermore, we also provide an experimental compilation transform in the `functorch.compile` namespace. Our compilation transform, named AOT (ahead-of-time) Autograd, returns to you an [FX graph](https://pytorch.org/docs/stable/fx.html) (that optionally contains a backward pass), of which compilation via various backends is one path you can take.\n", "\n", "\n", "## Why composable function transforms?\n", "There are a number of use cases that are tricky to do in PyTorch today:\n", "- computing per-sample-gradients (or other per-sample quantities)\n", "- running ensembles of models on a single machine\n", "- efficiently batching together tasks in the inner-loop of MAML\n", "- efficiently computing Jacobians and Hessians\n", "- efficiently computing batched Jacobians and Hessians\n", "\n", "Composing `vmap`, `grad`, `vjp`, and `jvp` transforms allows us to express the above without designing a separate subsystem for each.\n", "\n", "## What are the transforms?\n", "\n", "### grad (gradient computation)\n", "\n", "`grad(func)` is our gradient computation transform. It returns a new function that computes the gradients of `func`. It assumes `func` returns a single-element Tensor and by default it computes the gradients of the output of `func` w.r.t. to the first input." ] }, { "cell_type": "code", "execution_count": null, "id": "f920b923", "metadata": {}, "outputs": [], "source": [ "import torch\n", "from functorch import grad\n", "x = torch.randn([])\n", "cos_x = grad(lambda x: torch.sin(x))(x)\n", "assert torch.allclose(cos_x, x.cos())\n", "\n", "# Second-order gradients\n", "neg_sin_x = grad(grad(lambda x: torch.sin(x)))(x)\n", "assert torch.allclose(neg_sin_x, -x.sin())" ] }, { "cell_type": "markdown", "id": "ef3b2d85", "metadata": {}, "source": [ "### vmap (auto-vectorization)\n", "\n", "Note: vmap imposes restrictions on the code that it can be used on. For more details, please read its docstring.\n", "\n", "`vmap(func)(*inputs)` is a transform that adds a dimension to all Tensor operations in `func`. `vmap(func)` returns a new function that maps `func` over some dimension (default: 0) of each Tensor in inputs.\n", "\n", "vmap is useful for hiding batch dimensions: one can write a function func that runs on examples and then lift it to a function that can take batches of examples with `vmap(func)`, leading to a simpler modeling experience:" ] }, { "cell_type": "code", "execution_count": null, "id": "6ebac649", "metadata": {}, "outputs": [], "source": [ "import torch\n", "from functorch import vmap\n", "batch_size, feature_size = 3, 5\n", "weights = torch.randn(feature_size, requires_grad=True)\n", "\n", "def model(feature_vec):\n", " # Very simple linear model with activation\n", " assert feature_vec.dim() == 1\n", " return feature_vec.dot(weights).relu()\n", "\n", "examples = torch.randn(batch_size, feature_size)\n", "result = vmap(model)(examples)" ] }, { "cell_type": "markdown", "id": "5161e6d2", "metadata": {}, "source": [ "When composed with `grad`, `vmap` can be used to compute per-sample-gradients:" ] }, { "cell_type": "code", "execution_count": null, "id": "ffb2fcb1", "metadata": {}, "outputs": [], "source": [ "from functorch import vmap\n", "batch_size, feature_size = 3, 5\n", "\n", "def model(weights,feature_vec):\n", " # Very simple linear model with activation\n", " assert feature_vec.dim() == 1\n", " return feature_vec.dot(weights).relu()\n", "\n", "def compute_loss(weights, example, target):\n", " y = model(weights, example)\n", " return ((y - target) ** 2).mean() # MSELoss\n", "\n", "weights = torch.randn(feature_size, requires_grad=True)\n", "examples = torch.randn(batch_size, feature_size)\n", "targets = torch.randn(batch_size)\n", "inputs = (weights,examples, targets)\n", "grad_weight_per_example = vmap(grad(compute_loss), in_dims=(None, 0, 0))(*inputs)" ] }, { "cell_type": "markdown", "id": "11d711af", "metadata": {}, "source": [ "### vjp (vector-Jacobian product)\n", "\n", "The `vjp` transform applies `func` to `inputs` and returns a new function that computes the vector-Jacobian product (vjp) given some `cotangents` Tensors." ] }, { "cell_type": "code", "execution_count": null, "id": "ad48f9d4", "metadata": {}, "outputs": [], "source": [ "from functorch import vjp\n", "\n", "inputs = torch.randn(3)\n", "func = torch.sin\n", "cotangents = (torch.randn(3),)\n", "\n", "outputs, vjp_fn = vjp(func, inputs); vjps = vjp_fn(*cotangents)" ] }, { "cell_type": "markdown", "id": "e0221270", "metadata": {}, "source": [ "### jvp (Jacobian-vector product)\n", "\n", "The `jvp` transforms computes Jacobian-vector-products and is also known as \"forward-mode AD\". It is not a higher-order function unlike most other transforms, but it returns the outputs of `func(inputs)` as well as the jvps." ] }, { "cell_type": "code", "execution_count": null, "id": "f3772f43", "metadata": {}, "outputs": [], "source": [ "from functorch import jvp\n", "x = torch.randn(5)\n", "y = torch.randn(5)\n", "f = lambda x, y: (x * y)\n", "_, output = jvp(f, (x, y), (torch.ones(5), torch.ones(5)))\n", "assert torch.allclose(output, x + y)" ] }, { "cell_type": "markdown", "id": "7b00953b", "metadata": {}, "source": [ "### jacrev, jacfwd, and hessian\n", "\n", "The `jacrev` transform returns a new function that takes in `x` and returns the Jacobian of the function\n", "with respect to `x` using reverse-mode AD." ] }, { "cell_type": "code", "execution_count": null, "id": "20f53be2", "metadata": {}, "outputs": [], "source": [ "from functorch import jacrev\n", "x = torch.randn(5)\n", "jacobian = jacrev(torch.sin)(x)\n", "expected = torch.diag(torch.cos(x))\n", "assert torch.allclose(jacobian, expected)" ] }, { "cell_type": "markdown", "id": "b9007c88", "metadata": {}, "source": [ "Use `jacrev` to compute the jacobian. This can be composed with `vmap` to produce batched jacobians:" ] }, { "cell_type": "code", "execution_count": null, "id": "97d6c382", "metadata": {}, "outputs": [], "source": [ "x = torch.randn(64, 5)\n", "jacobian = vmap(jacrev(torch.sin))(x)\n", "assert jacobian.shape == (64, 5, 5)" ] }, { "cell_type": "markdown", "id": "cda642ec", "metadata": {}, "source": [ "`jacfwd` is a drop-in replacement for `jacrev` that computes Jacobians using forward-mode AD:" ] }, { "cell_type": "code", "execution_count": null, "id": "a8c1dedb", "metadata": {}, "outputs": [], "source": [ "from functorch import jacfwd\n", "x = torch.randn(5)\n", "jacobian = jacfwd(torch.sin)(x)\n", "expected = torch.diag(torch.cos(x))\n", "assert torch.allclose(jacobian, expected)" ] }, { "cell_type": "markdown", "id": "39f85b50", "metadata": {}, "source": [ "Composing `jacrev` with itself or `jacfwd` can produce hessians:" ] }, { "cell_type": "code", "execution_count": null, "id": "1e511139", "metadata": {}, "outputs": [], "source": [ "def f(x):\n", " return x.sin().sum()\n", "\n", "x = torch.randn(5)\n", "hessian0 = jacrev(jacrev(f))(x)\n", "hessian1 = jacfwd(jacrev(f))(x)" ] }, { "cell_type": "markdown", "id": "18efdc65", "metadata": {}, "source": [ "The `hessian` is a convenience function that combines `jacfwd` and `jacrev`:" ] }, { "cell_type": "code", "execution_count": null, "id": "fd1765df", "metadata": {}, "outputs": [], "source": [ "from functorch import hessian\n", "\n", "def f(x):\n", " return x.sin().sum()\n", "\n", "x = torch.randn(5)\n", "hess = hessian(f)(x)" ] }, { "cell_type": "markdown", "id": "b597d7ad", "metadata": {}, "source": [ "## Conclusion\n", "\n", "Check out our other tutorials (in the left bar) for more detailed explanations of how to apply functorch transforms for various use cases. `functorch` is very much a work in progress and we'd love to hear how you're using it -- we encourage you to start a conversation at our [issues tracker](https://github.com/pytorch/functorch) to discuss your use case." ] } ], "metadata": { "kernelspec": { "display_name": "Python 3 (ipykernel)", "language": "python", "name": "python3" }, "language_info": { "codemirror_mode": { "name": "ipython", "version": 3 }, "file_extension": ".py", "mimetype": "text/x-python", "name": "python", "nbconvert_exporter": "python", "pygments_lexer": "ipython3", "version": "3.7.4" } }, "nbformat": 4, "nbformat_minor": 5 }