Jax Usage Demo

jax_helloworld.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import jax\n",
    "import jax.numpy as jnp"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "class BaseLayer:\n",
    "    _parameters = {}\n",
    "\n",
    "\n",
    "class Dense2Layer(BaseLayer):\n",
    "    def __init__(self, input_dims: int, output_dims: int):\n",
    "        super().__init__()\n",
    "        assert (isinstance(input_dims, int) and isinstance(output_dims, int))\n",
    "        self._parameters['u'] = jax.random.normal(jax.random.PRNGKey(0), (input_dims, output_dims))\n",
    "        self._parameters['v'] = jax.random.normal(jax.random.PRNGKey(0), (input_dims, output_dims))\n",
    "\n",
    "        self._parameters['b'] = jax.random.normal(jax.random.PRNGKey(0), (output_dims, ))\n",
    "        self._parameters['x'] = None\n",
    "        self.vjp_fn = None\n",
    "\n",
    "    @classmethod\n",
    "    def forward_fn(cls, x, u, v, b):\n",
    "        return x**2 @ u + x @ v + b\n",
    "\n",
    "    def __call__(self, x: jax.numpy.DeviceArray):\n",
    "        self._parameters['x'] = x\n",
    "        out, self.vjp_fn = jax.vjp(self.forward_fn, layer._parameters['x'], layer._parameters['u'], layer._parameters['v'],\n",
    "                                   layer._parameters['b'])\n",
    "        return out\n",
    "\n",
    "\n",
    "class MSELoss:\n",
    "    def __call__(self, x, y):\n",
    "        return jnp.mean((x - y)**2)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "layer = Dense2Layer(20, 3)\n",
    "loss_fn = MSELoss()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "x = jax.random.normal(jax.random.PRNGKey(0), (2, 20))\n",
    "y = jax.numpy.array([[1, 1, 1], [0, 0, 0]], dtype=jax.numpy.float32)\n",
    "pred = layer(x)\n",
    "loss = loss_fn(pred, y)\n",
    "loss_grad = jax.grad(loss_fn, argnums=0)(pred, y)\n",
    "print(loss)\n",
    "print(loss_grad)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(layer.vjp_fn(loss_grad)[0].shape)\n",
    "print(layer.vjp_fn(loss_grad)[1].shape)\n",
    "print(layer.vjp_fn(loss_grad)[2].shape)\n",
    "print(layer.vjp_fn(loss_grad)[3].shape)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": []
  }
 ],
 "metadata": {
  "interpreter": {
   "hash": "aedd5e66c633617a2b6367c11c0534f47dbb53f58d311e2fdabbf5bcacebad6a"
  },
  "kernelspec": {
   "display_name": "Python 3.8.8 ('default')",
   "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": "3.8.8"
  },
  "orig_nbformat": 4
 },
 "nbformat": 4,
 "nbformat_minor": 2
}