GRU tuning

GRU tuning.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "import torch.nn as nn\n",
    "from torch import Tensor\n",
    "from typing import List"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "inputs = [torch.randn(64, 256, 256), torch.randn(128, 128, 256), torch.randn(256, 64, 256), torch.randn(512, 32, 256)]\n",
    "inputs = [x.cuda() for x in inputs]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "gru_cudnn = nn.GRU(256, 256).cuda()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "2.93 ms ± 36.1 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "out, _ = gru_cudnn(inputs[1])\n",
    "s = out.sum()\n",
    "s.sum().detach().cpu().numpy()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "8.91 ms ± 796 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "out, _ = gru_cudnn(inputs[1])\n",
    "s = out.sum()\n",
    "s.backward()\n",
    "gru_cudnn.bias_ih_l0.grad.sum().detach().cpu().numpy()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "class GRU(nn.Module):\n",
    "    def __init__(self, ks=256):\n",
    "        super(GRU, self).__init__()\n",
    "        \n",
    "        self.weight_ih = nn.Parameter(torch.randn(ks, ks * 3))\n",
    "        self.weight_hh = nn.Parameter(torch.randn(ks, ks * 3))\n",
    "        self.bias_ih = nn.Parameter(torch.randn(3 * ks))\n",
    "        self.bias_hh = nn.Parameter(torch.randn(3 * ks))\n",
    "        \n",
    "        self.ks = ks\n",
    "        \n",
    "    def forward(self, x: Tensor):\n",
    "        xparts = torch.unbind(x, 0)\n",
    "        \n",
    "        outs: List[Tensor] = []\n",
    "        last = torch.zeros_like((x[0]))\n",
    "        for i in range(len(xparts)):\n",
    "            part = xparts[i]\n",
    "            ip = torch.mm(part, self.weight_ih) + self.bias_ih\n",
    "            hp = torch.mm(last, self.weight_hh) + self.bias_hh\n",
    "            i_r, i_z, i_n = ip.chunk(3, 1)\n",
    "            h_r, h_z, h_n = hp.chunk(3, 1)\n",
    "            r = torch.sigmoid(i_r + h_r)\n",
    "            z = torch.sigmoid(i_z + h_z)\n",
    "            \n",
    "            n = torch.tanh(i_n + r * h_n)\n",
    "            last = (1 - z) * n + z * last\n",
    "            \n",
    "            outs += [last]\n",
    "            \n",
    "        return torch.stack(outs)\n",
    "    \n",
    "gru_raw = GRU(256).cuda()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "68.9 ms ± 1.52 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "out = gru_raw(inputs[1])\n",
    "s = out.sum()\n",
    "s.sum().detach().cpu().numpy()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "161 ms ± 2.04 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "out = gru_raw(inputs[1])\n",
    "s = out.sum()\n",
    "s.backward()\n",
    "gru_raw.bias_ih.grad.sum().detach().cpu().numpy()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "gru_jit = torch.jit.script(GRU(256).cuda())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "22.7 ms ± 23.4 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "out = gru_jit(inputs[1])\n",
    "s = out.sum()\n",
    "s.sum().detach().cpu().numpy()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "58.1 ms ± 682 µs per loop (mean ± std. dev. of 7 runs, 1 loop each)\n"
     ]
    }
   ],
   "source": [
    "%%timeit\n",
    "out = gru_jit(inputs[1])\n",
    "s = out.sum()\n",
    "s.backward()\n",
    "gru_jit.bias_ih.grad.sum().detach().cpu().numpy()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "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": "3.6.10"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}