{ "cells": [ { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [], "source": [ "from utils import Scalar, draw_graph\n", "from linear_model import Linear, mse" ] }, { "cell_type": "code", "execution_count": 2, "metadata": {}, "outputs": [ { "data": { "image/svg+xml": [ "\n", "\n", "\n", "\n", "\n", "\n", "%3\n", "\n", "\n", "140289705437696backward\n", "\n", "x1=1.50\n", "\n", "\n", "140289705438224backward\n", "\n", "grad=-4.50\n", "\n", "value=0.00\n", "\n", "*\n", "\n", "\n", "140289705437744backward\n", "\n", "x2=2.00\n", "\n", "\n", "140289705438224backward->140289705437744backward\n", "\n", "\n", "\n", "140289705437888backward\n", "\n", "grad=-10.50\n", "\n", "value=0.00\n", "\n", "a\n", "\n", "\n", "140289705438224backward->140289705437888backward\n", "\n", "\n", "-9.00\n", "\n", "\n", "140289705438272backward\n", "\n", "grad=-4.50\n", "\n", "value=0.00\n", "\n", "+\n", "\n", "\n", "140289705438272backward->140289705438224backward\n", "\n", "\n", "-4.50\n", "\n", "\n", "140289705438032backward\n", "\n", "grad=-5.50\n", "\n", "value=0.00\n", "\n", "b\n", "\n", "\n", "140289705438272backward->140289705438032backward\n", "\n", "\n", "-4.50\n", "\n", "\n", "140289705438320backward\n", "\n", "grad=4.50\n", "\n", "value=4.50\n", "\n", "-\n", "\n", "\n", "140289705438320backward->140289705438272backward\n", "\n", "\n", "-4.50\n", "\n", "\n", "140289705437648backward\n", "\n", "y2=4.50\n", "\n", "\n", "140289705438320backward->140289705437648backward\n", "\n", "\n", "\n", "140289705437840backward\n", "\n", "grad=-1.00\n", "\n", "value=0.00\n", "\n", "*\n", "\n", "\n", "140289705437840backward->140289705437696backward\n", "\n", "\n", "\n", "140289705437840backward->140289705437888backward\n", "\n", "\n", "-1.50\n", "\n", "\n", "140289705438368backward\n", "\n", "grad=1.00\n", "\n", "value=10.62\n", "\n", "mse\n", "\n", "\n", "140289705438368backward->140289705438320backward\n", "\n", "\n", "4.50\n", "\n", "\n", "140289705437936backward\n", "\n", "grad=1.00\n", "\n", "value=1.00\n", "\n", "-\n", "\n", "\n", "140289705438368backward->140289705437936backward\n", "\n", "\n", "1.00\n", "\n", "\n", "140289705438080backward\n", "\n", "grad=-1.00\n", "\n", "value=0.00\n", "\n", "+\n", "\n", "\n", "140289705437936backward->140289705438080backward\n", "\n", "\n", "-1.00\n", "\n", "\n", "140289705437600backward\n", "\n", "y1=1.00\n", "\n", "\n", "140289705437936backward->140289705437600backward\n", "\n", "\n", "\n", "140289705438080backward->140289705437840backward\n", "\n", "\n", "-1.00\n", "\n", "\n", "140289705438080backward->140289705438032backward\n", "\n", "\n", "-1.00\n", "\n", "\n", "\n" ], "text/plain": [ "" ] }, "execution_count": 2, "metadata": {}, "output_type": "execute_result" } ], "source": [ "# 定义训练数据\n", "x1 = Scalar(1.5, label='x1', requires_grad=False)\n", "y1 = Scalar(1.0, label='y1', requires_grad=False)\n", "x2 = Scalar(2.0, label='x2', requires_grad=False)\n", "y2 = Scalar(4.5, label='y2', requires_grad=False)\n", "# 反向传播\n", "model = Linear()\n", "loss = mse([model.error(x1, y1), model.error(x2, y2)])\n", "loss.backward()\n", "draw_graph(loss, 'backward')" ] }, { "cell_type": "code", "execution_count": 3, "metadata": {}, "outputs": [ { "data": { "image/svg+xml": [ "\n", "\n", "\n", "\n", "\n", "\n", "%3\n", "\n", "\n", "140289705437696backward\n", "\n", "x1=1.50\n", "\n", "\n", "140289705436688backward\n", "\n", "grad=-1.00\n", "\n", "value=0.00\n", "\n", "+\n", "\n", "\n", "140289705437792backward\n", "\n", "grad=-1.00\n", "\n", "value=0.00\n", "\n", "*\n", "\n", "\n", "140289705436688backward->140289705437792backward\n", "\n", "\n", "\n", "140289705436448backward\n", "\n", "grad=-5.50\n", "\n", "value=0.00\n", "\n", "b\n", "\n", "\n", "140289705436688backward->140289705436448backward\n", "\n", "\n", "-1.00\n", "\n", "\n", "140289705436112backward\n", "\n", "grad=-4.50\n", "\n", "value=0.00\n", "\n", "*\n", "\n", "\n", "140289705437744backward\n", "\n", "x2=2.00\n", "\n", "\n", "140289705436112backward->140289705437744backward\n", "\n", "\n", "\n", "140289705435536backward\n", "\n", "a=0.00\n", "\n", "\n", "140289705436112backward->140289705435536backward\n", "\n", "\n", "\n", "140289705437792backward->140289705437696backward\n", "\n", "\n", "\n", "140289705437792backward->140289705435536backward\n", "\n", "\n", "\n", "140289705435872backward\n", "\n", "grad=-4.50\n", "\n", "value=0.00\n", "\n", "+\n", "\n", "\n", "140289705435872backward->140289705436112backward\n", "\n", "\n", "\n", "140289705435872backward->140289705436448backward\n", "\n", "\n", "-4.50\n", "\n", "\n", "140289705437984backward\n", "\n", "grad=1.00\n", "\n", "value=1.00\n", "\n", "-\n", "\n", "\n", "140289705437984backward->140289705436688backward\n", "\n", "\n", "-1.00\n", "\n", "\n", "140289705437600backward\n", "\n", "y1=1.00\n", "\n", "\n", "140289705437984backward->140289705437600backward\n", "\n", "\n", "\n", "140289705436496backward\n", "\n", "grad=4.50\n", "\n", "value=4.50\n", "\n", "-\n", "\n", "\n", "140289705436496backward->140289705435872backward\n", "\n", "\n", "-4.50\n", "\n", "\n", "140289705437648backward\n", "\n", "y2=4.50\n", "\n", "\n", "140289705436496backward->140289705437648backward\n", "\n", "\n", "\n", "140289705436544backward\n", "\n", "grad=1.00\n", "\n", "value=10.62\n", "\n", "mse\n", "\n", "\n", "140289705436544backward->140289705437984backward\n", "\n", "\n", "1.00\n", "\n", "\n", "140289705436544backward->140289705436496backward\n", "\n", "\n", "4.50\n", "\n", "\n", "\n" ], "text/plain": [ "" ] }, "execution_count": 3, "metadata": {}, "output_type": "execute_result" } ], "source": [ "# 反向传播\n", "model = Linear()\n", "# 冻结参数a\n", "model.a.requires_grad = False\n", "loss = mse([model.error(x1, y1), model.error(x2, y2)])\n", "loss.backward()\n", "draw_graph(loss, 'backward')" ] }, { "cell_type": "code", "execution_count": 4, "metadata": {}, "outputs": [ { "data": { "image/svg+xml": [ "\n", "\n", "\n", "\n", "\n", "\n", "%3\n", "\n", "\n", "140289705437696backward\n", "\n", "x1=1.50\n", "\n", "\n", "140289705377312backward\n", "\n", "grad=-1.00\n", "\n", "value=0.00\n", "\n", "+\n", "\n", "\n", "140289705437312backward\n", "\n", "grad=-1.00\n", "\n", "value=0.00\n", "\n", "*\n", "\n", "\n", "140289705377312backward->140289705437312backward\n", "\n", "\n", "\n", "140289705377504backward\n", "\n", "b=0.00\n", "\n", "\n", "140289705377312backward->140289705377504backward\n", "\n", "\n", "\n", "140289705437744backward\n", "\n", "x2=2.00\n", "\n", "\n", "140289705437312backward->140289705437696backward\n", "\n", "\n", "\n", "140289705436064backward\n", "\n", "a=0.00\n", "\n", "\n", "140289705437312backward->140289705436064backward\n", "\n", "\n", "\n", "140289705377456backward\n", "\n", "grad=-4.50\n", "\n", "value=0.00\n", "\n", "+\n", "\n", "\n", "140289705377456backward->140289705377504backward\n", "\n", "\n", "\n", "140289705377072backward\n", "\n", "grad=-4.50\n", "\n", "value=0.00\n", "\n", "*\n", "\n", "\n", "140289705377456backward->140289705377072backward\n", "\n", "\n", "\n", "140289705376976backward\n", "\n", "grad=1.00\n", "\n", "value=1.00\n", "\n", "-\n", "\n", "\n", "140289705376976backward->140289705377312backward\n", "\n", "\n", "\n", "140289705437600backward\n", "\n", "y1=1.00\n", "\n", "\n", "140289705376976backward->140289705437600backward\n", "\n", "\n", "\n", "140289705377024backward\n", "\n", "grad=4.50\n", "\n", "value=4.50\n", "\n", "-\n", "\n", "\n", "140289705377024backward->140289705377456backward\n", "\n", "\n", "\n", "140289705437648backward\n", "\n", "y2=4.50\n", "\n", "\n", "140289705377024backward->140289705437648backward\n", "\n", "\n", "\n", "140289705377072backward->140289705437744backward\n", "\n", "\n", "\n", "140289705377072backward->140289705436064backward\n", "\n", "\n", "\n", "140289705377648backward\n", "\n", "grad=1.00\n", "\n", "value=10.62\n", "\n", "mse\n", "\n", "\n", "140289705377648backward->140289705376976backward\n", "\n", "\n", "\n", "140289705377648backward->140289705377024backward\n", "\n", "\n", "\n", "\n" ], "text/plain": [ "" ] }, "execution_count": 4, "metadata": {}, "output_type": "execute_result" } ], "source": [ "# 反向传播\n", "model = Linear()\n", "# 冻结参数a和参数b\n", "model.a.requires_grad = False\n", "model.b.requires_grad = False\n", "loss = mse([model.error(x1, y1), model.error(x2, y2)])\n", "loss.backward()\n", "draw_graph(loss, 'backward')" ] } ], "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": "3.8.5" } }, "nbformat": 4, "nbformat_minor": 4 }