computational-graph-is-only-built-once.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "electoral-admission",
   "metadata": {},
   "source": [
    "# computational-graph-is-only-built-once/113923"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "id": "fallen-isolation",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1.8.0\n"
     ]
    }
   ],
   "source": [
    "import torch\n",
    "import torch.nn as nn\n",
    "import torch.nn.functional as F\n",
    "\n",
    "print(torch.__version__)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "id": "tracked-virtue",
   "metadata": {},
   "outputs": [],
   "source": [
    "m = nn.Linear(4, 2)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "id": "according-polls",
   "metadata": {},
   "outputs": [],
   "source": [
    "m.zero_grad()\n",
    "loss = None\n",
    "grads = []\n",
    "inputs = []\n",
    "for _ in range(5):\n",
    "    x = torch.rand(4, 4)\n",
    "    inputs.append(x.clone().detach())\n",
    "    cur_loss = m(x).sum()\n",
    "    grads.append(torch.autograd.grad(cur_loss, m.weight)[0])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "id": "higher-arlington",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[tensor([[1.5253, 1.6025, 2.1368, 1.9085],\n",
       "         [1.5253, 1.6025, 2.1368, 1.9085]]),\n",
       " tensor([[1.9054, 1.6825, 2.2009, 2.3773],\n",
       "         [1.9054, 1.6825, 2.2009, 2.3773]]),\n",
       " tensor([[2.5990, 2.4845, 2.4768, 2.2761],\n",
       "         [2.5990, 2.4845, 2.4768, 2.2761]]),\n",
       " tensor([[2.4386, 1.6646, 1.7932, 1.1445],\n",
       "         [2.4386, 1.6646, 1.7932, 1.1445]]),\n",
       " tensor([[2.8388, 2.5705, 1.8727, 2.1668],\n",
       "         [2.8388, 2.5705, 1.8727, 2.1668]])]"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "grads"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "id": "aboriginal-failing",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([[11.3070, 10.0046, 10.4804,  9.8733],\n",
       "        [11.3070, 10.0046, 10.4804,  9.8733]])"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "sum(grads)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "id": "light-regulation",
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "tensor([[11.3070, 10.0046, 10.4804,  9.8733],\n",
       "        [11.3070, 10.0046, 10.4804,  9.8733]])"
      ]
     },
     "execution_count": 15,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "m.zero_grad()\n",
    "loss = None\n",
    "for x in inputs:\n",
    "    cur_loss = m(x).sum()\n",
    "    if loss is None:\n",
    "        loss = cur_loss\n",
    "    else:\n",
    "        loss += cur_loss\n",
    "loss.backward()\n",
    "m.weight.grad"
   ]
  }
 ],
 "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": "3.7.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}