fastai_DL_nbs/lesson07_cnn_define.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "%matplotlib inline\n",
    "%reload_ext autoreload\n",
    "%autoreload 2\n",
    "# 多行输出\n",
    "from IPython.core.interactiveshell import InteractiveShell\n",
    "InteractiveShell.ast_node_interactivity = \"all\" "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 自定义模型的使用"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "from fastai.vision import *"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "[PosixPath('/home/lyc/.fastai/data/mnist_png/training'),\n",
       " PosixPath('/home/lyc/.fastai/data/mnist_png/models'),\n",
       " PosixPath('/home/lyc/.fastai/data/mnist_png/testing')]"
      ]
     },
     "execution_count": 40,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "path = untar_data(URLs.MNIST)\n",
    "path.ls()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {},
   "outputs": [],
   "source": [
    "defaults.cmap='binary'"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- 训练集、验证集 tfms\n",
    "- 这里验证集不做变换"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 42,
   "metadata": {},
   "outputs": [],
   "source": [
    "tfms = ([*rand_pad(padding=3, size=28, mode='zeros')], [])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 43,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "([RandTransform(tfm=TfmPixel (pad), kwargs={'padding': 3, 'mode': 'zeros'}, p=1.0, resolved={}, do_run=True, is_random=True, use_on_y=True),\n",
       "  RandTransform(tfm=TfmPixel (crop), kwargs={'size': 28, 'row_pct': (0, 1), 'col_pct': (0, 1)}, p=1.0, resolved={}, do_run=True, is_random=True, use_on_y=True)],\n",
       " [])"
      ]
     },
     "execution_count": 43,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tfms"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- padding\n",
    "- random crop"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [],
   "source": [
    "data = (ImageList.from_folder(path, convert_mode='L')\n",
    "        .split_by_folder(train='training', valid='testing')\n",
    "        .label_from_folder()\n",
    "        .transform(tfms)\n",
    "        .databunch(bs=64)\n",
    "        .normalize()  # 使用自身的统计信息，因为不是要预训练模型\n",
    "       )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 504x504 with 9 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "data.show_batch(rows=3, figsize=(7,7))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 自定义模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "def conv(ni,nf):\n",
    "    return nn.Conv2d(ni, nf, kernel_size=3, stride=2, padding=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<h2 id=\"Flatten\" class=\"doc_header\"><code>class</code> <code>Flatten</code><a href=\"https://github.com/fastai/fastai/blob/master/fastai/layers.py#L37\" class=\"source_link\" style=\"float:right\">[source]</a><a class=\"source_link\" data-toggle=\"collapse\" data-target=\"#Flatten-pytest\" style=\"float:right; padding-right:10px\">[test]</a></h2><blockquote><p><code>Flatten</code>(<strong><code>full</code></strong>:<code>bool</code>=<strong><em><code>False</code></em></strong>) :: <a href=\"https://pytorch.org/docs/stable/nn.html#torch.nn.Module\"><code>Module</code></a></p>\n",
       "</blockquote>\n",
       "<div class=\"collapse\" id=\"Flatten-pytest\"><div class=\"card card-body pytest_card\"><a type=\"button\" data-toggle=\"collapse\" data-target=\"#Flatten-pytest\" class=\"close\" aria-label=\"Close\"><span aria-hidden=\"true\">&times;</span></a><p>No tests found for <code>Flatten</code>. To contribute a test please refer to <a href=\"/dev/test.html\">this guide</a> and <a href=\"https://forums.fast.ai/t/improving-expanding-functional-tests/32929\">this discussion</a>.</p></div></div><p>Flatten <code>x</code> to a single dimension, often used at the end of a model. <code>full</code> for rank-1 tensor</p>\n",
       "<p><a href=\"https://docs.fast.ai/layers.html#Flatten\" target=\"_blank\" rel=\"noreferrer noopener\">Show in docs</a></p>\n"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "doc(Flatten)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = nn.Sequential(\n",
    "    conv(1, 8), # 14\n",
    "    nn.BatchNorm2d(8),\n",
    "    nn.ReLU(),\n",
    "    conv(8, 16), # 7\n",
    "    nn.BatchNorm2d(16),\n",
    "    nn.ReLU(),\n",
    "    conv(16, 32), # 4\n",
    "    nn.BatchNorm2d(32),\n",
    "    nn.ReLU(),\n",
    "    conv(32, 16), # 2\n",
    "    nn.BatchNorm2d(16),\n",
    "    nn.ReLU(),\n",
    "    conv(16, 10), # 1\n",
    "    nn.BatchNorm2d(10),\n",
    "    Flatten()     # remove (1,1) grid\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<h2 id=\"Learner\" class=\"doc_header\"><code>class</code> <code>Learner</code><a href=\"https://github.com/fastai/fastai/blob/master/fastai/basic_train.py#L144\" class=\"source_link\" style=\"float:right\">[source]</a><a class=\"source_link\" data-toggle=\"collapse\" data-target=\"#Learner-pytest\" style=\"float:right; padding-right:10px\">[test]</a></h2><blockquote><p><code>Learner</code>(<strong><code>data</code></strong>:<a href=\"https://docs.fast.ai/basic_data.html#DataBunch\"><code>DataBunch</code></a>, <strong><code>model</code></strong>:<a href=\"https://pytorch.org/docs/stable/nn.html#torch.nn.Module\"><code>Module</code></a>, <strong><code>opt_func</code></strong>:<code>Callable</code>=<strong><em><code>'Adam'</code></em></strong>, <strong><code>loss_func</code></strong>:<code>Callable</code>=<strong><em><code>None</code></em></strong>, <strong><code>metrics</code></strong>:<code>Collection</code>[<code>Callable</code>]=<strong><em><code>None</code></em></strong>, <strong><code>true_wd</code></strong>:<code>bool</code>=<strong><em><code>True</code></em></strong>, <strong><code>bn_wd</code></strong>:<code>bool</code>=<strong><em><code>True</code></em></strong>, <strong><code>wd</code></strong>:<code>Floats</code>=<strong><em><code>0.01</code></em></strong>, <strong><code>train_bn</code></strong>:<code>bool</code>=<strong><em><code>True</code></em></strong>, <strong><code>path</code></strong>:<code>str</code>=<strong><em><code>None</code></em></strong>, <strong><code>model_dir</code></strong>:<code>PathOrStr</code>=<strong><em><code>'models'</code></em></strong>, <strong><code>callback_fns</code></strong>:<code>Collection</code>[<code>Callable</code>]=<strong><em><code>None</code></em></strong>, <strong><code>callbacks</code></strong>:<code>Collection</code>[<a href=\"https://docs.fast.ai/callback.html#Callback\"><code>Callback</code></a>]=<strong><em><code>&lt;factory&gt;</code></em></strong>, <strong><code>layer_groups</code></strong>:<code>ModuleList</code>=<strong><em><code>None</code></em></strong>, <strong><code>add_time</code></strong>:<code>bool</code>=<strong><em><code>True</code></em></strong>, <strong><code>silent</code></strong>:<code>bool</code>=<strong><em><code>None</code></em></strong>)</p>\n",
       "</blockquote>\n",
       "<div class=\"collapse\" id=\"Learner-pytest\"><div class=\"card card-body pytest_card\"><a type=\"button\" data-toggle=\"collapse\" data-target=\"#Learner-pytest\" class=\"close\" aria-label=\"Close\"><span aria-hidden=\"true\">&times;</span></a><p>No tests found for <code>Learner</code>. To contribute a test please refer to <a href=\"/dev/test.html\">this guide</a> and <a href=\"https://forums.fast.ai/t/improving-expanding-functional-tests/32929\">this discussion</a>.</p></div></div><p>Trainer for <code>model</code> using <code>data</code> to minimize <code>loss_func</code> with optimizer <code>opt_func</code>.</p>\n",
       "<p><a href=\"https://docs.fast.ai/basic_train.html#Learner\" target=\"_blank\" rel=\"noreferrer noopener\">Show in docs</a></p>\n"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "doc(Learner)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "learn = Learner(data, model, loss_func = nn.CrossEntropyLoss(), metrics=accuracy).to_fp16()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "======================================================================\n",
       "Layer (type)         Output Shape         Param #    Trainable \n",
       "======================================================================\n",
       "Conv2d               [8, 14, 14]          80         True      \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [8, 14, 14]          16         True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [8, 14, 14]          0          False     \n",
       "______________________________________________________________________\n",
       "Conv2d               [16, 7, 7]           1,168      True      \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [16, 7, 7]           32         True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [16, 7, 7]           0          False     \n",
       "______________________________________________________________________\n",
       "Conv2d               [32, 4, 4]           4,640      True      \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [32, 4, 4]           64         True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [32, 4, 4]           0          False     \n",
       "______________________________________________________________________\n",
       "Conv2d               [16, 2, 2]           4,624      True      \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [16, 2, 2]           32         True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [16, 2, 2]           0          False     \n",
       "______________________________________________________________________\n",
       "Conv2d               [10, 1, 1]           1,450      True      \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [10, 1, 1]           20         True      \n",
       "______________________________________________________________________\n",
       "Flatten              [10]                 0          False     \n",
       "______________________________________________________________________\n",
       "\n",
       "Total params: 12,126\n",
       "Total trainable params: 12,126\n",
       "Total non-trainable params: 0"
      ]
     },
     "execution_count": 13,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "learn.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(Image (1, 28, 28), Category 7)"
      ]
     },
     "execution_count": 14,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "data.train_ds[0]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "7\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 216x216 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "x, y = data.train_ds[0]\n",
    "x.show()\n",
    "print(y)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "torch.Size([64, 1, 28, 28])"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    },
    {
     "data": {
      "text/plain": [
       "torch.Size([64])"
      ]
     },
     "execution_count": 16,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "xb, yb = data.one_batch()  # bs=64\n",
    "xb.shape\n",
    "yb.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.\n",
      "Min numerical gradient: 8.32E-02\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "learn.lr_find()\n",
    "learn.recorder.plot(suggestion=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "lr = 2e-1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.286687</td>\n",
       "      <td>0.265215</td>\n",
       "      <td>0.916800</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.256713</td>\n",
       "      <td>0.173404</td>\n",
       "      <td>0.944400</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.178962</td>\n",
       "      <td>0.119607</td>\n",
       "      <td>0.963400</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>0.125316</td>\n",
       "      <td>0.058706</td>\n",
       "      <td>0.983100</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>0.076931</td>\n",
       "      <td>0.037404</td>\n",
       "      <td>0.988700</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "learn.fit_one_cycle(5, lr)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 重建模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [],
   "source": [
    "def conv2(ni,nf):\n",
    "    return conv_layer(ni,nf,stride=2)  # 库函数\n",
    "# conv_layer which lets you create conv, batch norm, ReLU combinations"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = nn.Sequential(\n",
    "    conv2(1, 8),   # 14\n",
    "    conv2(8, 16),  # 7\n",
    "    conv2(16, 32), # 4\n",
    "    conv2(32, 16), # 2\n",
    "    conv2(16, 10), # 1\n",
    "    Flatten()      # remove (1,1) grid\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "learn = Learner(data, model, loss_func = nn.CrossEntropyLoss(), metrics=accuracy).to_fp16()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "======================================================================\n",
       "Layer (type)         Output Shape         Param #    Trainable \n",
       "======================================================================\n",
       "Conv2d               [8, 14, 14]          72         True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [8, 14, 14]          0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [8, 14, 14]          16         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [16, 7, 7]           1,152      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [16, 7, 7]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [16, 7, 7]           32         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [32, 4, 4]           4,608      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [32, 4, 4]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [32, 4, 4]           64         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [16, 2, 2]           4,608      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [16, 2, 2]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [16, 2, 2]           32         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [10, 1, 1]           1,440      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [10, 1, 1]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [10, 1, 1]           20         True      \n",
       "______________________________________________________________________\n",
       "Flatten              [10]                 0          False     \n",
       "______________________________________________________________________\n",
       "\n",
       "Total params: 12,044\n",
       "Total trainable params: 12,044\n",
       "Total non-trainable params: 0"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "learn.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.\n",
      "Min numerical gradient: 6.31E-01\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "learn.lr_find()\n",
    "learn.recorder.plot(suggestion=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "lr=1e-1"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.268805</td>\n",
       "      <td>0.132822</td>\n",
       "      <td>0.956900</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.227748</td>\n",
       "      <td>0.255854</td>\n",
       "      <td>0.915400</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.209142</td>\n",
       "      <td>0.381966</td>\n",
       "      <td>0.880400</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>0.165852</td>\n",
       "      <td>0.260037</td>\n",
       "      <td>0.921600</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>0.160655</td>\n",
       "      <td>0.102029</td>\n",
       "      <td>0.968300</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>5</td>\n",
       "      <td>0.121476</td>\n",
       "      <td>0.129365</td>\n",
       "      <td>0.959700</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>6</td>\n",
       "      <td>0.112839</td>\n",
       "      <td>0.049855</td>\n",
       "      <td>0.984300</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>7</td>\n",
       "      <td>0.066354</td>\n",
       "      <td>0.038637</td>\n",
       "      <td>0.987300</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>8</td>\n",
       "      <td>0.067430</td>\n",
       "      <td>0.034191</td>\n",
       "      <td>0.989400</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>9</td>\n",
       "      <td>0.059260</td>\n",
       "      <td>0.030368</td>\n",
       "      <td>0.990100</td>\n",
       "      <td>00:08</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "learn.fit_one_cycle(10, lr)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 残差结构"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<h4 id=\"conv_layer\" class=\"doc_header\"><code>conv_layer</code><a href=\"https://github.com/fastai/fastai/blob/master/fastai/layers.py#L128\" class=\"source_link\" style=\"float:right\">[source]</a><a class=\"source_link\" data-toggle=\"collapse\" data-target=\"#conv_layer-pytest\" style=\"float:right; padding-right:10px\">[test]</a></h4><blockquote><p><code>conv_layer</code>(<strong><code>ni</code></strong>:<code>int</code>, <strong><code>nf</code></strong>:<code>int</code>, <strong><code>ks</code></strong>:<code>int</code>=<strong><em><code>3</code></em></strong>, <strong><code>stride</code></strong>:<code>int</code>=<strong><em><code>1</code></em></strong>, <strong><code>padding</code></strong>:<code>int</code>=<strong><em><code>None</code></em></strong>, <strong><code>bias</code></strong>:<code>bool</code>=<strong><em><code>None</code></em></strong>, <strong><code>is_1d</code></strong>:<code>bool</code>=<strong><em><code>False</code></em></strong>, <strong><code>norm_type</code></strong>:<code>Optional</code>[<a href=\"https://docs.fast.ai/layers.html#NormType\"><code>NormType</code></a>]=<strong><em><code>&lt;NormType.Batch: 1&gt;</code></em></strong>, <strong><code>use_activ</code></strong>:<code>bool</code>=<strong><em><code>True</code></em></strong>, <strong><code>leaky</code></strong>:<code>float</code>=<strong><em><code>None</code></em></strong>, <strong><code>transpose</code></strong>:<code>bool</code>=<strong><em><code>False</code></em></strong>, <strong><code>init</code></strong>:<code>Callable</code>=<strong><em><code>'kaiming_normal_'</code></em></strong>, <strong><code>self_attention</code></strong>:<code>bool</code>=<strong><em><code>False</code></em></strong>)</p>\n",
       "</blockquote>\n",
       "<div class=\"collapse\" id=\"conv_layer-pytest\"><div class=\"card card-body pytest_card\"><a type=\"button\" data-toggle=\"collapse\" data-target=\"#conv_layer-pytest\" class=\"close\" aria-label=\"Close\"><span aria-hidden=\"true\">&times;</span></a><p>No tests found for <code>conv_layer</code>. To contribute a test please refer to <a href=\"/dev/test.html\">this guide</a> and <a href=\"https://forums.fast.ai/t/improving-expanding-functional-tests/32929\">this discussion</a>.</p></div></div><p>Create a sequence of convolutional (<code>ni</code> to <code>nf</code>), ReLU (if <code>use_activ</code>) and batchnorm (if <code>bn</code>) layers.</p>\n",
       "<p><a href=\"https://docs.fast.ai/layers.html#conv_layer\" target=\"_blank\" rel=\"noreferrer noopener\">Show in docs</a></p>\n"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "doc(conv_layer)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [],
   "source": [
    "class ResBlock(nn.Module):\n",
    "    def __init__(self, nf):\n",
    "        super().__init__()\n",
    "        self.conv1 = conv_layer(nf,nf)\n",
    "        self.conv2 = conv_layer(nf,nf)\n",
    "        \n",
    "    def forward(self, x):\n",
    "        return x + self.conv2(self.conv1(x))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<h4 id=\"res_block\" class=\"doc_header\"><code>res_block</code><a href=\"https://github.com/fastai/fastai/blob/master/fastai/layers.py#L174\" class=\"source_link\" style=\"float:right\">[source]</a><a class=\"source_link\" data-toggle=\"collapse\" data-target=\"#res_block-pytest\" style=\"float:right; padding-right:10px\">[test]</a></h4><blockquote><p><code>res_block</code>(<strong><code>nf</code></strong>, <strong><code>dense</code></strong>:<code>bool</code>=<strong><em><code>False</code></em></strong>, <strong><code>norm_type</code></strong>:<code>Optional</code>[<a href=\"https://docs.fast.ai/layers.html#NormType\"><code>NormType</code></a>]=<strong><em><code>&lt;NormType.Batch: 1&gt;</code></em></strong>, <strong><code>bottle</code></strong>:<code>bool</code>=<strong><em><code>False</code></em></strong>, <strong>**<code>conv_kwargs</code></strong>)</p>\n",
       "</blockquote>\n",
       "<div class=\"collapse\" id=\"res_block-pytest\"><div class=\"card card-body pytest_card\"><a type=\"button\" data-toggle=\"collapse\" data-target=\"#res_block-pytest\" class=\"close\" aria-label=\"Close\"><span aria-hidden=\"true\">&times;</span></a><p>No tests found for <code>res_block</code>. To contribute a test please refer to <a href=\"/dev/test.html\">this guide</a> and <a href=\"https://forums.fast.ai/t/improving-expanding-functional-tests/32929\">this discussion</a>.</p></div></div><p>Resnet block of <code>nf</code> features. <code>conv_kwargs</code> are passed to <a href=\"https://docs.fast.ai/layers.html#conv_layer\"><code>conv_layer</code></a>.</p>\n",
       "<p><a href=\"https://docs.fast.ai/layers.html#res_block\" target=\"_blank\" rel=\"noreferrer noopener\">Show in docs</a></p>\n"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "doc(res_block)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = nn.Sequential(\n",
    "    conv2(1, 8),\n",
    "    res_block(8),\n",
    "    conv2(8, 16),\n",
    "    res_block(16),\n",
    "    conv2(16, 32),\n",
    "    res_block(32),\n",
    "    conv2(32, 16),\n",
    "    res_block(16),\n",
    "    conv2(16, 10),\n",
    "    Flatten()\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {},
   "outputs": [],
   "source": [
    "def conv_and_res(ni,nf):\n",
    "    return nn.Sequential(conv2(ni, nf), res_block(nf))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 32,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = nn.Sequential(\n",
    "    conv_and_res(1, 8),\n",
    "    conv_and_res(8, 16),\n",
    "    conv_and_res(16, 32),\n",
    "    conv_and_res(32, 16),\n",
    "    conv2(16, 10),\n",
    "    Flatten()\n",
    ")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {},
   "outputs": [],
   "source": [
    "learn = Learner(data, model, loss_func = nn.CrossEntropyLoss(), metrics=accuracy).to_fp16()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 34,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "======================================================================\n",
       "Layer (type)         Output Shape         Param #    Trainable \n",
       "======================================================================\n",
       "Conv2d               [8, 14, 14]          72         True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [8, 14, 14]          0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [8, 14, 14]          16         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [8, 14, 14]          576        True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [8, 14, 14]          0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [8, 14, 14]          16         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [8, 14, 14]          576        True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [8, 14, 14]          0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [8, 14, 14]          16         True      \n",
       "______________________________________________________________________\n",
       "MergeLayer           [8, 14, 14]          0          False     \n",
       "______________________________________________________________________\n",
       "Conv2d               [16, 7, 7]           1,152      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [16, 7, 7]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [16, 7, 7]           32         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [16, 7, 7]           2,304      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [16, 7, 7]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [16, 7, 7]           32         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [16, 7, 7]           2,304      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [16, 7, 7]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [16, 7, 7]           32         True      \n",
       "______________________________________________________________________\n",
       "MergeLayer           [16, 7, 7]           0          False     \n",
       "______________________________________________________________________\n",
       "Conv2d               [32, 4, 4]           4,608      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [32, 4, 4]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [32, 4, 4]           64         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [32, 4, 4]           9,216      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [32, 4, 4]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [32, 4, 4]           64         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [32, 4, 4]           9,216      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [32, 4, 4]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [32, 4, 4]           64         True      \n",
       "______________________________________________________________________\n",
       "MergeLayer           [32, 4, 4]           0          False     \n",
       "______________________________________________________________________\n",
       "Conv2d               [16, 2, 2]           4,608      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [16, 2, 2]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [16, 2, 2]           32         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [16, 2, 2]           2,304      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [16, 2, 2]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [16, 2, 2]           32         True      \n",
       "______________________________________________________________________\n",
       "Conv2d               [16, 2, 2]           2,304      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [16, 2, 2]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [16, 2, 2]           32         True      \n",
       "______________________________________________________________________\n",
       "MergeLayer           [16, 2, 2]           0          False     \n",
       "______________________________________________________________________\n",
       "Conv2d               [10, 1, 1]           1,440      True      \n",
       "______________________________________________________________________\n",
       "ReLU                 [10, 1, 1]           0          False     \n",
       "______________________________________________________________________\n",
       "BatchNorm2d          [10, 1, 1]           20         True      \n",
       "______________________________________________________________________\n",
       "Flatten              [10]                 0          False     \n",
       "______________________________________________________________________\n",
       "\n",
       "Total params: 41,132\n",
       "Total trainable params: 41,132\n",
       "Total non-trainable params: 0"
      ]
     },
     "execution_count": 34,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "learn.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<h2 id=\"MergeLayer\" class=\"doc_header\"><code>class</code> <code>MergeLayer</code><a href=\"https://github.com/fastai/fastai/blob/master/fastai/layers.py#L166\" class=\"source_link\" style=\"float:right\">[source]</a><a class=\"source_link\" data-toggle=\"collapse\" data-target=\"#MergeLayer-pytest\" style=\"float:right; padding-right:10px\">[test]</a></h2><blockquote><p><code>MergeLayer</code>(<strong><code>dense</code></strong>:<code>bool</code>=<strong><em><code>False</code></em></strong>) :: <a href=\"https://pytorch.org/docs/stable/nn.html#torch.nn.Module\"><code>Module</code></a></p>\n",
       "</blockquote>\n",
       "<div class=\"collapse\" id=\"MergeLayer-pytest\"><div class=\"card card-body pytest_card\"><a type=\"button\" data-toggle=\"collapse\" data-target=\"#MergeLayer-pytest\" class=\"close\" aria-label=\"Close\"><span aria-hidden=\"true\">&times;</span></a><p>No tests found for <code>MergeLayer</code>. To contribute a test please refer to <a href=\"/dev/test.html\">this guide</a> and <a href=\"https://forums.fast.ai/t/improving-expanding-functional-tests/32929\">this discussion</a>.</p></div></div><p>Merge a shortcut with the result of the module by adding them or concatenating thme if <code>dense=True</code>.</p>\n",
       "<p><a href=\"https://docs.fast.ai/layers.html#MergeLayer\" target=\"_blank\" rel=\"noreferrer noopener\">Show in docs</a></p>\n"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "doc(MergeLayer)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 36,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "LR Finder is complete, type {learner_name}.recorder.plot() to see the graph.\n",
      "Min numerical gradient: 4.37E-01\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "learn.lr_find()\n",
    "learn.recorder.plot(suggestion=True)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 37,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: left;\">\n",
       "      <th>epoch</th>\n",
       "      <th>train_loss</th>\n",
       "      <th>valid_loss</th>\n",
       "      <th>accuracy</th>\n",
       "      <th>time</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <td>0</td>\n",
       "      <td>0.194276</td>\n",
       "      <td>0.152102</td>\n",
       "      <td>0.957800</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>1</td>\n",
       "      <td>0.152755</td>\n",
       "      <td>0.170926</td>\n",
       "      <td>0.942700</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>2</td>\n",
       "      <td>0.127692</td>\n",
       "      <td>0.124472</td>\n",
       "      <td>0.963100</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>3</td>\n",
       "      <td>0.106741</td>\n",
       "      <td>0.061006</td>\n",
       "      <td>0.980800</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>4</td>\n",
       "      <td>0.095490</td>\n",
       "      <td>0.099073</td>\n",
       "      <td>0.973600</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>5</td>\n",
       "      <td>0.085740</td>\n",
       "      <td>0.061804</td>\n",
       "      <td>0.980900</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>6</td>\n",
       "      <td>0.067191</td>\n",
       "      <td>0.193605</td>\n",
       "      <td>0.940600</td>\n",
       "      <td>00:14</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>7</td>\n",
       "      <td>0.074187</td>\n",
       "      <td>0.094971</td>\n",
       "      <td>0.969900</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>8</td>\n",
       "      <td>0.063273</td>\n",
       "      <td>0.041282</td>\n",
       "      <td>0.987900</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>9</td>\n",
       "      <td>0.052292</td>\n",
       "      <td>0.045675</td>\n",
       "      <td>0.986000</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>10</td>\n",
       "      <td>0.048365</td>\n",
       "      <td>0.149492</td>\n",
       "      <td>0.988500</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>11</td>\n",
       "      <td>0.034946</td>\n",
       "      <td>0.020052</td>\n",
       "      <td>0.992800</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>12</td>\n",
       "      <td>0.027740</td>\n",
       "      <td>0.017906</td>\n",
       "      <td>0.994300</td>\n",
       "      <td>00:14</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>13</td>\n",
       "      <td>0.022770</td>\n",
       "      <td>0.016615</td>\n",
       "      <td>0.994700</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <td>14</td>\n",
       "      <td>0.014335</td>\n",
       "      <td>0.016833</td>\n",
       "      <td>0.994700</td>\n",
       "      <td>00:13</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>"
      ],
      "text/plain": [
       "<IPython.core.display.HTML object>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "learn.fit_one_cycle(15, 0.05)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![](../imgs/res.png)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- 使用更简洁的代码构建模型，避免犯错\n",
    "- 减少重复，重复性工作使用函数完成，使用编程技巧"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<h2 id=\"MergeLayer\" class=\"doc_header\"><code>class</code> <code>MergeLayer</code><a href=\"https://github.com/fastai/fastai/blob/master/fastai/layers.py#L166\" class=\"source_link\" style=\"float:right\">[source]</a><a class=\"source_link\" data-toggle=\"collapse\" data-target=\"#MergeLayer-pytest\" style=\"float:right; padding-right:10px\">[test]</a></h2><blockquote><p><code>MergeLayer</code>(<strong><code>dense</code></strong>:<code>bool</code>=<strong><em><code>False</code></em></strong>) :: <a href=\"https://pytorch.org/docs/stable/nn.html#torch.nn.Module\"><code>Module</code></a></p>\n",
       "</blockquote>\n",
       "<div class=\"collapse\" id=\"MergeLayer-pytest\"><div class=\"card card-body pytest_card\"><a type=\"button\" data-toggle=\"collapse\" data-target=\"#MergeLayer-pytest\" class=\"close\" aria-label=\"Close\"><span aria-hidden=\"true\">&times;</span></a><p>No tests found for <code>MergeLayer</code>. To contribute a test please refer to <a href=\"/dev/test.html\">this guide</a> and <a href=\"https://forums.fast.ai/t/improving-expanding-functional-tests/32929\">this discussion</a>.</p></div></div><p>Merge a shortcut with the result of the module by adding them or concatenating thme if <code>dense=True</code>.</p>\n",
       "<p><a href=\"https://docs.fast.ai/layers.html#MergeLayer\" target=\"_blank\" rel=\"noreferrer noopener\">Show in docs</a></p>\n"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "doc(MergeLayer)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "```python\n",
    "class MergeLayer(nn.Module):\n",
    "    \"Merge a shortcut with the result of the module by adding them or concatenating thme if `dense=True`.\"\n",
    "    def __init__(self, dense:bool=False):\n",
    "        super().__init__()\n",
    "        self.dense=dense\n",
    "\n",
    "    def forward(self, x): return torch.cat([x,x.orig], dim=1) if self.dense else (x+x.orig)\n",
    "```"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<h2 id=\"SequentialEx\" class=\"doc_header\"><code>class</code> <code>SequentialEx</code><a href=\"https://github.com/fastai/fastai/blob/master/fastai/layers.py#L145\" class=\"source_link\" style=\"float:right\">[source]</a><a class=\"source_link\" data-toggle=\"collapse\" data-target=\"#SequentialEx-pytest\" style=\"float:right; padding-right:10px\">[test]</a></h2><blockquote><p><code>SequentialEx</code>(<strong>*<code>layers</code></strong>) :: <a href=\"https://pytorch.org/docs/stable/nn.html#torch.nn.Module\"><code>Module</code></a></p>\n",
       "</blockquote>\n",
       "<div class=\"collapse\" id=\"SequentialEx-pytest\"><div class=\"card card-body pytest_card\"><a type=\"button\" data-toggle=\"collapse\" data-target=\"#SequentialEx-pytest\" class=\"close\" aria-label=\"Close\"><span aria-hidden=\"true\">&times;</span></a><p>No tests found for <code>SequentialEx</code>. To contribute a test please refer to <a href=\"/dev/test.html\">this guide</a> and <a href=\"https://forums.fast.ai/t/improving-expanding-functional-tests/32929\">this discussion</a>.</p></div></div><p>Like <a href=\"https://pytorch.org/docs/stable/nn.html#torch.nn.Sequential\"><code>nn.Sequential</code></a>, but with ModuleList semantics, and can access module input</p>\n",
       "<p><a href=\"https://docs.fast.ai/layers.html#SequentialEx\" target=\"_blank\" rel=\"noreferrer noopener\">Show in docs</a></p>\n"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "doc(SequentialEx)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "```python\n",
    "class SequentialEx(nn.Module):\n",
    "    \"Like `nn.Sequential`, but with ModuleList semantics, and can access module input\"\n",
    "    def __init__(self, *layers):\n",
    "        super().__init__()\n",
    "        self.layers = nn.ModuleList(layers)\n",
    "\n",
    "    def forward(self, x):\n",
    "        res = x\n",
    "        for l in self.layers:\n",
    "            res.orig = x\n",
    "            nres = l(res)\n",
    "            # We have to remove res.orig to avoid hanging refs and therefore memory leaks\n",
    "            res.orig = None\n",
    "            res = nres\n",
    "        return res\n",
    "\n",
    "    def __getitem__(self,i): return self.layers[i]\n",
    "    def append(self,l): return self.layers.append(l)\n",
    "    def extend(self,l): return self.layers.extend(l)\n",
    "    def insert(self,i,l): return self.layers.insert(i,l)\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- 使用 SequentialEx 和 MergeLayer可以轻松地创建残差结构"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "```python\n",
    "def res_block(nf, dense:bool=False, norm_type:Optional[NormType]=NormType.Batch, bottle:bool=False, **conv_kwargs):\n",
    "    \"Resnet block of `nf` features. `conv_kwargs` are passed to `conv_layer`.\"\n",
    "    norm2 = norm_type\n",
    "    if not dense and (norm_type==NormType.Batch): norm2 = NormType.BatchZero\n",
    "    nf_inner = nf//2 if bottle else nf\n",
    "    return SequentialEx(conv_layer(nf, nf_inner, norm_type=norm_type, **conv_kwargs),\n",
    "                      conv_layer(nf_inner, nf, norm_type=norm2, **conv_kwargs),\n",
    "                      MergeLayer(dense))\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "- DensNet在小型数据集上表现往往较好\n",
    "- 对分割任务也比较友好"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### UNET"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![](../imgs/unet.png)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "in this case, this is from the U-Net paper. Their images, they started with one channel by 572 by 572. This is for medical imaging segmentation. After your stride 2 conv, they're doubling the number of channels to 128, and they're halving the size so they're now down to 280 by 280. In this original unit paper, they didn't add any padding. So they lost a pixel on each side each time they did a conv. That's why you are losing these two. But basically half the size, and then half the size, and then half the size, and then half the size, until they're down to 28 by 28 with 1024 channels."
   ]
  },
  {
   "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": "3.6.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}