打开股票量化的黑箱-自己动手实现一个印钞机4.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# 打开股票量化的黑箱(自己动手实现一个‘印钞机’) 第四章\n",
    "\n",
    "### 作者：阿布🐶\n",
    "\n",
    "### 未经本人允许禁止转载"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "##  ‘非均衡胜负收益’带来的必然’非均衡胜负比例‘，目标由’因子‘的能力解决一部分，’模式识别‘提升关键的一部分"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "上一章使用机器学习的方法，想要提取特征，指导交易，提高胜率，但是发现，除了最后那种把交易结果分成100份的方式外，其它\n",
    "机器学习方法基本都是瞎猜，是不是使用深度学习就能解决问题呢？\n",
    "\n",
    "本章主要通过使用卷积神经网络模型alex_net,  与google_lenet对股票\n",
    "进行模式识别"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import ZEnv\n",
    "import ZLog\n",
    "import ZCommonUtil\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "import seaborn as sns\n",
    "import matplotlib.pyplot as plt\n",
    "pd.options.display.max_columns = 100\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 加载缓存交易数据"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(array([[  0.13997608,   0.07045403,   0.06829068],\n",
       "        [ -0.40946745,  -0.38829941,   0.75027609],\n",
       "        [ 11.84144398,  -0.84690618,   2.73882144],\n",
       "        [-17.26610011,  -3.57896598,  -3.86127002],\n",
       "        [ -9.14670563,   4.14514014,   2.85483688]]),\n",
       " array([ 0.,  0.,  0.,  0.,  1.]))"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 从之前跑的结果hdf5中加载缓存\n",
    "from MlFiterDegPd import MlFiterDegPdClass\n",
    "orders_pd_train_snap = ZCommonUtil.load_hdf5('orders_pd_train_snap', 'orders_pd_train_snap')\n",
    "orders_pd_test_snap = ZCommonUtil.load_hdf5('orders_pd_test_snap', 'orders_pd_test_snap')\n",
    "orders_pd_test_snap.shape\n",
    "deg = MlFiterDegPdClass(orderPd=orders_pd_train_snap)\n",
    "deg.x[:5], deg.y[:5]"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 首先使用多层神经网络，明显也是不行"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch: 0001 cost= 13.496154785\n"
     ]
    }
   ],
   "source": [
    "from MlFiterTensorFlow import MnnTF\n",
    "from sklearn.cross_validation import KFold\n",
    "\n",
    "kf = KFold(len(deg.y), n_folds=10, shuffle=True)\n",
    "for i, (train_index, test_index) in enumerate(kf):\n",
    "    x_train, x_test = deg.x[train_index], deg.x[test_index]\n",
    "    y_train, y_test = deg.y[train_index], deg.y[test_index]\n",
    "    mnn = MnnTF(x_train, y_train, x_test, y_test, n_hidden_1=10, n_hidden_2=10, batch_size=len(deg.x)/2,\n",
    "                 learning_rate=0.01, training_epochs=10000, display_step=1000)\n",
    "    ac = mnn.fit()\n",
    "    break # only once"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'/Users/Bailey/Desktop/my_work/abu/data/save_png/2016-10-09/2016_10_09_12_06_59_042764_4241_lyQvysYoAwVoFylreOwxNXNXvvncdyyjDEjSORFMEUAfJBpLBHjXFbwQBUoZRtkO.png'"
      ]
     },
     "execution_count": 5,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "orders_pd_train_snap.ix[0].snap"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "之前把所有交易前的snap都保留了(snap为交易前60日的k图)，这回不自己传入特征，直接使用卷积神经网络，自己学习特征\n",
    "![img](./Snip20161014_5.png)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 首先使用TensorFlow的高级api tflearn来实现alex_net的卷积神经网络"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# 准备工作\n",
    "import TrainValSplit\n",
    "import os\n",
    "import shutil\n",
    "\n",
    "# 把结果cut 10份\n",
    "def cut_order_profit_classes(orders_pd, test=False, n_classes=10):\n",
    "    orders_pd.profit_cg.fillna(0, inplace=True)\n",
    "    order_has_ret = orders_pd[(orders_pd.result <> 0)]\n",
    "    order_has_ret['class'] = pd.qcut(order_has_ret.profit_cg, n_classes, labels=range(0, n_classes))\n",
    "    order_has_ret['class'] = order_has_ret['class'].astype(int)\n",
    "    snap_pd = order_has_ret.filter(['snap', 'class'])\n",
    "    img_root_folder = os.path.dirname(snap_pd.ix[0].snap)\n",
    "    img_root_folder = img_root_folder + '/'\n",
    "    if test:\n",
    "        img_root_folder += 'test/'  \n",
    "    return snap_pd, img_root_folder"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 83,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'/Users/Bailey/Desktop/my_work/abu/data/save_png/2016-10-09/'"
      ]
     },
     "execution_count": 83,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "snap_pd, img_root_train_folder = cut_order_profit_classes(orders_pd_train_snap, test=False)\n",
    "img_root_train_folder"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 84,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "'/Users/Bailey/Desktop/my_work/abu/data/save_png/2016-10-09/test/'"
      ]
     },
     "execution_count": 84,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "snap_pd_test, img_root_test_folder = cut_order_profit_classes(orders_pd_test_snap, test=True)\n",
    "img_root_test_folder"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# 做子文件夹放分类图只是为了使用tflearn的build_image_dataset_from_dir方便\n",
    "def make_sub_dir_mv(snap_pd, img_root_folder):\n",
    "    target_dir = img_root_folder \n",
    "    map(lambda cls: os.makedirs(target_dir + str(cls) + '/') \n",
    "        if not os.path.exists(target_dir + str(cls) + '/') else 1, np.arange(0, 10))\n",
    "    \n",
    "    for ind in np.arange(0, len(snap_pd)):\n",
    "        path = snap_pd.ix[ind]['snap']\n",
    "        cls  = snap_pd.ix[ind]['class']\n",
    "        \n",
    "        target_path = target_dir + str(cls) + '/' + os.path.basename(path)\n",
    "        if os.path.exists(target_path):\n",
    "            continue\n",
    "        \n",
    "        shutil.move(path, target_path)\n",
    "    snap_pd['snap'] = snap_pd.apply(lambda snap: str(snap['class']) + '/' + os.path.basename(snap.snap), axis=1)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 65,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>snap</th>\n",
       "      <th>class</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>2015-09-11</th>\n",
       "      <td>1/2016_10_09_18_15_39_583096_6110_JSNTMWgcqKcT...</td>\n",
       "      <td>1</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2015-09-11</th>\n",
       "      <td>5/2016_10_09_18_06_02_707471_6110_oBSQVzUkhtPd...</td>\n",
       "      <td>5</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2015-09-11</th>\n",
       "      <td>5/2016_10_09_18_12_20_586317_6110_JLshhttghRuO...</td>\n",
       "      <td>5</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2015-09-11</th>\n",
       "      <td>0/2016_10_09_17_59_44_188534_6110_IcMNqzQqybsW...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>2015-09-11</th>\n",
       "      <td>0/2016_10_09_18_18_28_393821_6110_wqmNPNBFlTmU...</td>\n",
       "      <td>0</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "</div>"
      ],
      "text/plain": [
       "                                                         snap  class\n",
       "2015-09-11  1/2016_10_09_18_15_39_583096_6110_JSNTMWgcqKcT...      1\n",
       "2015-09-11  5/2016_10_09_18_06_02_707471_6110_oBSQVzUkhtPd...      5\n",
       "2015-09-11  5/2016_10_09_18_12_20_586317_6110_JLshhttghRuO...      5\n",
       "2015-09-11  0/2016_10_09_17_59_44_188534_6110_IcMNqzQqybsW...      0\n",
       "2015-09-11  0/2016_10_09_18_18_28_393821_6110_wqmNPNBFlTmU...      0"
      ]
     },
     "execution_count": 65,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "# 移动训练集snap\n",
    "make_sub_dir_mv(snap_pd, img_root_train_folder)\n",
    "# 移动测试集snap\n",
    "make_sub_dir_mv(snap_pd_test, img_root_test_folder)\n",
    "snap_pd.head()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 44,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# 将训练集和测试集的png都转换为rgb的jpeg\n",
    "import ImgStdHelper\n",
    "to = img_root_train_folder\n",
    "for ind in np.arange(0, 10):\n",
    "    std_dir = to + str(ind) + \"/\"\n",
    "    ImgStdHelper.std_img_from_root_dir(std_dir, 'png')\n",
    "to = img_root_test_folder\n",
    "for ind in np.arange(0, 10):\n",
    "    std_dir = to + str(ind) + \"/\"\n",
    "    ImgStdHelper.std_img_from_root_dir(std_dir, 'png')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 85,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "snap_pd.snap = snap_pd.snap.apply(lambda snap: snap[:-4] + '.jpeg')\n",
    "snap_pd_test.snap = snap_pd_test.snap.apply(lambda snap: snap[:-4] + '.jpeg')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 获取训练数据"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "from tflearn.data_utils import build_image_dataset_from_dir\n",
    "\n",
    "dataset_file = ZEnv.g_project_root + '/Caffe/gen/stock_tensor.pkl'\n",
    "one_hot=True, \n",
    "resize_pics=(227, 227)\n",
    "x, y = build_image_dataset_from_dir(img_root_train_folder,\n",
    "                             dataset_file=dataset_file,\n",
    "                             resize=resize_pics,\n",
    "                             filetypes=['.jpg', '.jpeg'],\n",
    "                             convert_gray=False,\n",
    "                             shuffle_data=True,\n",
    "                             categorical_Y=one_hot)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 使用tflear开始训练"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 35,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import tflearn\n",
    "from tflearn.layers.core import input_data, dropout, fully_connected\n",
    "from tflearn.layers.conv import conv_2d, max_pool_2d\n",
    "from tflearn.layers.normalization import local_response_normalization\n",
    "from tflearn.layers.estimator import regression\n",
    "network = input_data(shape=[None, 227, 227, 3])\n",
    "network = conv_2d(network, 96, 11, strides=4, activation='relu')\n",
    "network = max_pool_2d(network, 3, strides=2)\n",
    "network = local_response_normalization(network)\n",
    "network = conv_2d(network, 256, 5, activation='relu')\n",
    "network = max_pool_2d(network, 3, strides=2)\n",
    "network = local_response_normalization(network)\n",
    "network = conv_2d(network, 384, 3, activation='relu')\n",
    "network = conv_2d(network, 384, 3, activation='relu')\n",
    "network = conv_2d(network, 256, 3, activation='relu')\n",
    "network = max_pool_2d(network, 3, strides=2)\n",
    "network = local_response_normalization(network)\n",
    "network = fully_connected(network, 4096, activation='tanh')\n",
    "network = dropout(network, 0.5)\n",
    "network = fully_connected(network, 4096, activation='tanh')\n",
    "network = dropout(network, 0.5)\n",
    "network = fully_connected(network, 10, activation='softmax')\n",
    "network = regression(network, optimizer='momentum',\n",
    "                     loss='categorical_crossentropy',\n",
    "                     learning_rate=0.001)\n",
    "model = tflearn.DNN(network, checkpoint_path='model_alexnet',\n",
    "                    max_checkpoints=1, tensorboard_verbose=2)\n",
    "model.fit(x, y, n_epoch=1000, validation_set=0.1, shuffle=True,\n",
    "          show_metric=True, batch_size=64, snapshot_step=200,\n",
    "          snapshot_epoch=False, run_id='alexnet_stock')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "/Users/Bailey/Desktop/my_work/abu/Caffe/gen/stock_tensor_test.pkl\n"
     ]
    }
   ],
   "source": [
    "# 抽取100张作为测试，每个class抽取10张\n",
    "fr = img_root_test_folder\n",
    "to = img_root_test_folder + 'choice_test/'\n",
    "\n",
    "for ind in np.arange(0, 10):\n",
    "    ff = fr + str(ind) + \"/*.jpeg\"\n",
    "    imglist = glob.glob(ff) \n",
    "    imglist = np.random.choice(imglist, 10, replace=False)\n",
    "    for img in imglist:\n",
    "        img_to = to + str(ind) + \"/\" + os.path.basename(img)\n",
    "        ZCommonUtil.ensure_dir(img_to)\n",
    "        shutil.copy(img, img_to)\n",
    "        \n",
    "# 加载测试数据        \n",
    "dataset_file_test = ZEnv.g_project_root + '/Caffe/gen/stock_tensor_test.pkl'\n",
    "print(dataset_file_test)\n",
    "one_hot=True, \n",
    "resize_pics=(227, 227)\n",
    "test_folder = to\n",
    "x_test, y_test = build_image_dataset_from_dir(test_folder,\n",
    "                             dataset_file=dataset_file_test,\n",
    "                             resize=resize_pics,\n",
    "                             filetypes=['.jpg', '.jpeg', '.png'],\n",
    "                             convert_gray=False,\n",
    "                             shuffle_data=True,\n",
    "                             categorical_Y=one_hot)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 33,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "((100, 227, 227, 3), (100, 10))"
      ]
     },
     "execution_count": 33,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "x_test.shape, y_test.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 41,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(58, 100)"
      ]
     },
     "execution_count": 41,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "test_ret = model.predict(x_test)\n",
    "test_cnt = len(test_ret)\n",
    "rt_cnt = 0\n",
    "for pred_y, real_y in zip(test_ret, y_test):\n",
    "    real_y = np.argmax(real_y)\n",
    "    f = np.argmax(pred_y)\n",
    "    if f == real_y:\n",
    "        rt_cnt += 1\n",
    "        continue\n",
    "        \n",
    "    pred_y[f] = 0\n",
    "    t = np.argmax(pred_y)\n",
    "    if t == real_y:\n",
    "        rt_cnt += 1\n",
    "        continue\n",
    "        \n",
    "    pred_y[t] = 0\n",
    "    ttt = np.argmax(pred_y)\n",
    "    if ttt == real_y:\n",
    "        rt_cnt += 1\n",
    "        continue\n",
    "rt_cnt, test_cnt"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## top3的准确率达到 58%，稍为加工一下就可以放到程序里作为优化器了，使用方式类似:单子买入之前对below 3和top3进行判断，如果在below3中且不top3在中进行拦截，当然也可以有更多变种，关于这方面的实例将会在后续章节hmm-gmm中演示，更会延伸到裁判机制"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "______\n",
    "\n",
    "## 接下来使用caffe进行训练，优化\n",
    "## 使用google_lenet看看能不能有再次提升"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 75,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(4612, 2)"
      ]
     },
     "execution_count": 75,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "snap_pd.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 95,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def caffe_train_sh(snap_pd, img_root_folder):\n",
    "    train_path = ZEnv.g_project_root + '/caffe/gen/train.txt'\n",
    "    ZCommonUtil.ensure_dir(train_path)\n",
    "    snap_pd.to_csv(train_path, sep=' ', header=None, index=False)\n",
    "    TrainValSplit.train_val_split(train_path, n_folds=10)\n",
    "    lmdb = ZEnv.g_project_root + '/Caffe/sh/Lmdb.sh'\n",
    "    out_put_dir = ZEnv.g_project_root + '/Caffe/gen'\n",
    "    sh_cmd = '{} ${} ${}'.format(lmdb, img_root_folder, out_put_dir) \n",
    "    print(sh_cmd)\n",
    "    os.system(sh_cmd)\n",
    "    return sh_cmd"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 生成训练测试集，之后生成lmdb 及 mean pb"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 96,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "/Users/Bailey/Desktop/my_work/abu/Caffe/sh/Lmdb.sh $/Users/Bailey/Desktop/my_work/abu/data/save_png/2016-10-09/ $/Users/Bailey/Desktop/my_work/abu/Caffe/gen\n"
     ]
    }
   ],
   "source": [
    "sh_cmd = caffe_train_sh(snap_pd, img_root_train_folder)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 关于其它配置，请自己查看代码pb文件，训练完模型之后先拿一个测试看看效果"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 152,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "('/Users/Bailey/Desktop/my_work/abu/data/save_png/2016-10-09/test/1/2016_10_09_18_15_39_583096_6110_JSNTMWgcqKcTwDJeRrHlYvCsqSlRjeLwZdJpJvVzsATDYMLKkUFOhqraqoZGHqkm.jpeg',\n",
       " 1)"
      ]
     },
     "execution_count": 152,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "path = img_root_test_folder + snap_pd_test.ix[0].snap\n",
    "classes = snap_pd_test.ix[0]['class']\n",
    "path, classes"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### 加载训练好了的模型"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 153,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "mu = [ 243.24985442  238.91917917  239.75766282]\n"
     ]
    }
   ],
   "source": [
    "import caffe\n",
    "caffe.set_mode_cpu()\n",
    "\n",
    "image = caffe.io.load_image(path)\n",
    "model_def = ZEnv.g_project_root + '/Caffe/pb/deploy.prototxt' \n",
    "model_weights = ZEnv.g_project_root + '/Caffe/gen/stock_judge_train_iter_6000.caffemodel'\n",
    "model_mean_file = ZEnv.g_project_root + '/Caffe/gen/mean.binaryproto'\n",
    "\n",
    "net = caffe.Net(model_def, model_weights, caffe.TEST)  \n",
    "\n",
    "mean_blob = caffe.proto.caffe_pb2.BlobProto()\n",
    "mean_blob.ParseFromString(open(model_mean_file, 'rb').read())\n",
    "mean_npy = caffe.io.blobproto_to_array(mean_blob)\n",
    "mu = mean_npy.mean(2).mean(2)[0]\n",
    "print('mu = {}'.format(mu))\n",
    "transformer = caffe.io.Transformer({'data': net.blobs['data'].data.shape})\n",
    "transformer.set_transpose('data', (2,0,1))  \n",
    "transformer.set_mean('data', mu)           \n",
    "transformer.set_raw_scale('data', 255)     \n",
    "transformer.set_channel_swap('data', (2,1,0))  "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 154,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAQUAAAEBCAYAAABxB7CHAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nh6MNJwoOh6MNJwoOh6ON/w/4YWMo+KJzKQAAAABJRU5ErkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x12d64c510>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "transformed_image = transformer.preprocess('data', image)\n",
    "plt.imshow(image)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 155,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "demo classes is 1\n",
      "('predicted class is:', 7)\n",
      "('Top3 is:', array([7, 3, 1]))\n"
     ]
    }
   ],
   "source": [
    "net.blobs['data'].data[...] = transformed_image\n",
    "output = net.forward()\n",
    "output_prob = output['prob'][0]\n",
    "print('demo classes is {}'.format(classes))\n",
    "print('predicted class is:', output_prob.argmax())\n",
    "print('Top3 is:', output_prob.argsort()[::-1][:3])"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "如下， 测试数据集的头100个作为测试\n",
    "计算top3与top5的准确率\n",
    "top3: 54%\n",
    "top5: 63%"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 157,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(54, 63)"
      ]
     },
     "execution_count": 157,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "top3 = 0\n",
    "top5 = 0\n",
    "for ind in np.arange(0, 100):\n",
    "    path = img_root_test_folder + snap_pd_test.ix[ind].snap\n",
    "    classes = snap_pd_test.ix[ind]['class']\n",
    "    \n",
    "    image = caffe.io.load_image(path)\n",
    "    net.blobs['data'].data[...] = transformed_image\n",
    "    output = net.forward()\n",
    "    output_prob = output['prob'][0]\n",
    "    if classes in output_prob.argsort()[::-1][:3]:\n",
    "        top3 += 1\n",
    "        top5 += 1\n",
    "    elif classes in output_prob.argsort()[::-1][:5]:\n",
    "        top5 += 1\n",
    "top3, top5"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "\n",
    "\n",
    "### 使用深度学习确实对整个模式识别相比其它机器学习方法有了很大的飞跃，有了这个特征识别能力，就可以总结出新的交易拦截能力，进一步提升胜率，假想一下你的整个模型能力能有60的胜率而且每次赢的还比输的多，那它是不是就可以称为印钞机呢，当然了还有很多其它因素影响了印钞机的具体工作，这里无法一一阐述，有兴趣的可以加我的微信一起讨论！"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## 感谢🙏您能有耐心看到这里\n",
    "## 如果有什么问题可以加阿布的微信 \n",
    "## 微信号：aaaabbbuu"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![image](./mmexport1475383814280.jpg)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "anaconda-cloud": {},
  "kernelspec": {
   "display_name": "Python [Root]",
   "language": "python",
   "name": "Python [Root]"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.12"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}