LSTM_model.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import matplotlib.pyplot as plt\n",
    "# coding=utf-8\n",
    "# get_ipython().run_line_magic('matplotlib', 'inline')\n",
    "import pandas as pd\n",
    "import numpy as np\n",
    "from keras.layers.core import Flatten\n",
    "\n",
    "import tensorflow as tf\n",
    "from tensorflow.keras.preprocessing.sequence import pad_sequences\n",
    "from tensorflow.keras.models import Sequential,Model\n",
    "from tensorflow.keras.layers import Dense, Embedding, SpatialDropout1D\n",
    "from sklearn.model_selection import train_test_split\n",
    "from tensorflow.keras.callbacks import EarlyStopping\n",
    "from tensorflow.keras.callbacks import ModelCheckpoint\n",
    "from tensorflow.keras.layers import Dropout,LSTM,BatchNormalization,Conv1D,MaxPool1D,Reshape,Input,Conv2D\n",
    "from tensorflow.keras.optimizers import SGD\n",
    "from sklearn.utils import shuffle\n",
    "from sklearn.metrics import roc_auc_score\n",
    "import re\n",
    "import torch\n",
    "from keras.utils.vis_utils import plot_model\n",
    "import math \n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/html": [
       "<div>\n",
       "<style scoped>\n",
       "    .dataframe tbody tr th:only-of-type {\n",
       "        vertical-align: middle;\n",
       "    }\n",
       "\n",
       "    .dataframe tbody tr th {\n",
       "        vertical-align: top;\n",
       "    }\n",
       "\n",
       "    .dataframe thead th {\n",
       "        text-align: right;\n",
       "    }\n",
       "</style>\n",
       "<table border=\"1\" class=\"dataframe\">\n",
       "  <thead>\n",
       "    <tr style=\"text-align: right;\">\n",
       "      <th></th>\n",
       "      <th>Area</th>\n",
       "      <th>Perimeter</th>\n",
       "      <th>Major_Axis</th>\n",
       "      <th>Minor_Axis</th>\n",
       "      <th>Eccentricity</th>\n",
       "      <th>Eqdiasq</th>\n",
       "      <th>Solidity</th>\n",
       "      <th>Convex_Area</th>\n",
       "      <th>Extent</th>\n",
       "      <th>Aspect_Ratio</th>\n",
       "      <th>...</th>\n",
       "      <th>StdDev_RR</th>\n",
       "      <th>StdDev_RG</th>\n",
       "      <th>StdDev_RB</th>\n",
       "      <th>Skew_RR</th>\n",
       "      <th>Skew_RG</th>\n",
       "      <th>Skew_RB</th>\n",
       "      <th>Kurtosis_RR</th>\n",
       "      <th>Kurtosis_RG</th>\n",
       "      <th>Kurtosis_RB</th>\n",
       "      <th>Class</th>\n",
       "    </tr>\n",
       "  </thead>\n",
       "  <tbody>\n",
       "    <tr>\n",
       "      <th>766</th>\n",
       "      <td>84561</td>\n",
       "      <td>2020.3459</td>\n",
       "      <td>500.4059</td>\n",
       "      <td>242.4451</td>\n",
       "      <td>0.8748</td>\n",
       "      <td>328.1256</td>\n",
       "      <td>0.8579</td>\n",
       "      <td>98567</td>\n",
       "      <td>0.6699</td>\n",
       "      <td>2.0640</td>\n",
       "      <td>...</td>\n",
       "      <td>22.5937</td>\n",
       "      <td>23.7850</td>\n",
       "      <td>23.3334</td>\n",
       "      <td>-0.8506</td>\n",
       "      <td>-0.7833</td>\n",
       "      <td>-0.6849</td>\n",
       "      <td>2.8545</td>\n",
       "      <td>2.7554</td>\n",
       "      <td>2.7177</td>\n",
       "      <td>Kirmizi_Pistachio</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1879</th>\n",
       "      <td>93715</td>\n",
       "      <td>1297.0000</td>\n",
       "      <td>495.9520</td>\n",
       "      <td>242.0692</td>\n",
       "      <td>0.8728</td>\n",
       "      <td>345.4297</td>\n",
       "      <td>0.9788</td>\n",
       "      <td>95748</td>\n",
       "      <td>0.8001</td>\n",
       "      <td>2.0488</td>\n",
       "      <td>...</td>\n",
       "      <td>18.9635</td>\n",
       "      <td>19.1980</td>\n",
       "      <td>19.6101</td>\n",
       "      <td>-0.9200</td>\n",
       "      <td>-0.8737</td>\n",
       "      <td>-0.6758</td>\n",
       "      <td>3.3452</td>\n",
       "      <td>3.2793</td>\n",
       "      <td>3.4488</td>\n",
       "      <td>Siirt_Pistachio</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>377</th>\n",
       "      <td>68290</td>\n",
       "      <td>1236.1790</td>\n",
       "      <td>439.6265</td>\n",
       "      <td>200.1081</td>\n",
       "      <td>0.8904</td>\n",
       "      <td>294.8720</td>\n",
       "      <td>0.9673</td>\n",
       "      <td>70597</td>\n",
       "      <td>0.7652</td>\n",
       "      <td>2.1969</td>\n",
       "      <td>...</td>\n",
       "      <td>25.8404</td>\n",
       "      <td>27.2885</td>\n",
       "      <td>29.3617</td>\n",
       "      <td>-0.9461</td>\n",
       "      <td>-0.8359</td>\n",
       "      <td>-0.7615</td>\n",
       "      <td>2.8244</td>\n",
       "      <td>2.6158</td>\n",
       "      <td>2.5660</td>\n",
       "      <td>Kirmizi_Pistachio</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>584</th>\n",
       "      <td>73496</td>\n",
       "      <td>1118.4030</td>\n",
       "      <td>414.9055</td>\n",
       "      <td>227.9002</td>\n",
       "      <td>0.8356</td>\n",
       "      <td>305.9052</td>\n",
       "      <td>0.9738</td>\n",
       "      <td>75474</td>\n",
       "      <td>0.7397</td>\n",
       "      <td>1.8206</td>\n",
       "      <td>...</td>\n",
       "      <td>21.6237</td>\n",
       "      <td>22.9490</td>\n",
       "      <td>21.5302</td>\n",
       "      <td>-0.6926</td>\n",
       "      <td>-0.6324</td>\n",
       "      <td>-0.4304</td>\n",
       "      <td>2.4142</td>\n",
       "      <td>2.2298</td>\n",
       "      <td>2.5626</td>\n",
       "      <td>Kirmizi_Pistachio</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1322</th>\n",
       "      <td>95744</td>\n",
       "      <td>1344.7400</td>\n",
       "      <td>486.1907</td>\n",
       "      <td>273.9415</td>\n",
       "      <td>0.8262</td>\n",
       "      <td>349.1490</td>\n",
       "      <td>0.8889</td>\n",
       "      <td>107705</td>\n",
       "      <td>0.7280</td>\n",
       "      <td>1.7748</td>\n",
       "      <td>...</td>\n",
       "      <td>21.7916</td>\n",
       "      <td>25.7564</td>\n",
       "      <td>23.4475</td>\n",
       "      <td>-0.7043</td>\n",
       "      <td>-0.6322</td>\n",
       "      <td>-0.4200</td>\n",
       "      <td>2.4659</td>\n",
       "      <td>2.3822</td>\n",
       "      <td>2.2293</td>\n",
       "      <td>Siirt_Pistachio</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1300</th>\n",
       "      <td>104952</td>\n",
       "      <td>1309.2030</td>\n",
       "      <td>470.0492</td>\n",
       "      <td>295.9375</td>\n",
       "      <td>0.7769</td>\n",
       "      <td>365.5531</td>\n",
       "      <td>0.9348</td>\n",
       "      <td>112267</td>\n",
       "      <td>0.7297</td>\n",
       "      <td>1.5883</td>\n",
       "      <td>...</td>\n",
       "      <td>19.5857</td>\n",
       "      <td>23.3758</td>\n",
       "      <td>24.6418</td>\n",
       "      <td>-1.1376</td>\n",
       "      <td>-1.0231</td>\n",
       "      <td>-0.9553</td>\n",
       "      <td>4.1631</td>\n",
       "      <td>3.5007</td>\n",
       "      <td>3.6535</td>\n",
       "      <td>Siirt_Pistachio</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>556</th>\n",
       "      <td>84143</td>\n",
       "      <td>1170.3860</td>\n",
       "      <td>474.2799</td>\n",
       "      <td>227.7225</td>\n",
       "      <td>0.8772</td>\n",
       "      <td>327.3136</td>\n",
       "      <td>0.9864</td>\n",
       "      <td>85307</td>\n",
       "      <td>0.7556</td>\n",
       "      <td>2.0827</td>\n",
       "      <td>...</td>\n",
       "      <td>22.9112</td>\n",
       "      <td>24.6619</td>\n",
       "      <td>24.8875</td>\n",
       "      <td>-0.9349</td>\n",
       "      <td>-0.7006</td>\n",
       "      <td>-0.4097</td>\n",
       "      <td>3.3906</td>\n",
       "      <td>2.8394</td>\n",
       "      <td>2.5640</td>\n",
       "      <td>Kirmizi_Pistachio</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>49</th>\n",
       "      <td>60803</td>\n",
       "      <td>1317.9060</td>\n",
       "      <td>413.8860</td>\n",
       "      <td>195.5904</td>\n",
       "      <td>0.8813</td>\n",
       "      <td>278.2387</td>\n",
       "      <td>0.9106</td>\n",
       "      <td>66769</td>\n",
       "      <td>0.6834</td>\n",
       "      <td>2.1161</td>\n",
       "      <td>...</td>\n",
       "      <td>22.7933</td>\n",
       "      <td>23.7798</td>\n",
       "      <td>22.8103</td>\n",
       "      <td>-1.0110</td>\n",
       "      <td>-0.9310</td>\n",
       "      <td>-0.6495</td>\n",
       "      <td>3.2229</td>\n",
       "      <td>3.1221</td>\n",
       "      <td>3.0700</td>\n",
       "      <td>Kirmizi_Pistachio</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>785</th>\n",
       "      <td>73607</td>\n",
       "      <td>1070.1479</td>\n",
       "      <td>428.6051</td>\n",
       "      <td>219.3727</td>\n",
       "      <td>0.8591</td>\n",
       "      <td>306.1361</td>\n",
       "      <td>0.9909</td>\n",
       "      <td>74281</td>\n",
       "      <td>0.7982</td>\n",
       "      <td>1.9538</td>\n",
       "      <td>...</td>\n",
       "      <td>18.6741</td>\n",
       "      <td>18.9521</td>\n",
       "      <td>21.7716</td>\n",
       "      <td>-1.0000</td>\n",
       "      <td>-0.5618</td>\n",
       "      <td>0.0594</td>\n",
       "      <td>4.0868</td>\n",
       "      <td>3.5127</td>\n",
       "      <td>3.3700</td>\n",
       "      <td>Kirmizi_Pistachio</td>\n",
       "    </tr>\n",
       "    <tr>\n",
       "      <th>1263</th>\n",
       "      <td>92868</td>\n",
       "      <td>1163.1219</td>\n",
       "      <td>435.7098</td>\n",
       "      <td>273.1259</td>\n",
       "      <td>0.7791</td>\n",
       "      <td>343.8651</td>\n",
       "      <td>0.9920</td>\n",
       "      <td>93616</td>\n",
       "      <td>0.7304</td>\n",
       "      <td>1.5953</td>\n",
       "      <td>...</td>\n",
       "      <td>21.1671</td>\n",
       "      <td>21.9591</td>\n",
       "      <td>22.2506</td>\n",
       "      <td>-0.2361</td>\n",
       "      <td>-0.1603</td>\n",
       "      <td>-0.1153</td>\n",
       "      <td>2.3393</td>\n",
       "      <td>2.2724</td>\n",
       "      <td>2.1015</td>\n",
       "      <td>Siirt_Pistachio</td>\n",
       "    </tr>\n",
       "  </tbody>\n",
       "</table>\n",
       "<p>10 rows × 29 columns</p>\n",
       "</div>"
      ],
      "text/plain": [
       "        Area  Perimeter  Major_Axis  Minor_Axis  Eccentricity   Eqdiasq  \\\n",
       "766    84561  2020.3459    500.4059    242.4451        0.8748  328.1256   \n",
       "1879   93715  1297.0000    495.9520    242.0692        0.8728  345.4297   \n",
       "377    68290  1236.1790    439.6265    200.1081        0.8904  294.8720   \n",
       "584    73496  1118.4030    414.9055    227.9002        0.8356  305.9052   \n",
       "1322   95744  1344.7400    486.1907    273.9415        0.8262  349.1490   \n",
       "1300  104952  1309.2030    470.0492    295.9375        0.7769  365.5531   \n",
       "556    84143  1170.3860    474.2799    227.7225        0.8772  327.3136   \n",
       "49     60803  1317.9060    413.8860    195.5904        0.8813  278.2387   \n",
       "785    73607  1070.1479    428.6051    219.3727        0.8591  306.1361   \n",
       "1263   92868  1163.1219    435.7098    273.1259        0.7791  343.8651   \n",
       "\n",
       "      Solidity  Convex_Area  Extent  Aspect_Ratio  ...  StdDev_RR  StdDev_RG  \\\n",
       "766     0.8579        98567  0.6699        2.0640  ...    22.5937    23.7850   \n",
       "1879    0.9788        95748  0.8001        2.0488  ...    18.9635    19.1980   \n",
       "377     0.9673        70597  0.7652        2.1969  ...    25.8404    27.2885   \n",
       "584     0.9738        75474  0.7397        1.8206  ...    21.6237    22.9490   \n",
       "1322    0.8889       107705  0.7280        1.7748  ...    21.7916    25.7564   \n",
       "1300    0.9348       112267  0.7297        1.5883  ...    19.5857    23.3758   \n",
       "556     0.9864        85307  0.7556        2.0827  ...    22.9112    24.6619   \n",
       "49      0.9106        66769  0.6834        2.1161  ...    22.7933    23.7798   \n",
       "785     0.9909        74281  0.7982        1.9538  ...    18.6741    18.9521   \n",
       "1263    0.9920        93616  0.7304        1.5953  ...    21.1671    21.9591   \n",
       "\n",
       "      StdDev_RB  Skew_RR  Skew_RG  Skew_RB  Kurtosis_RR  Kurtosis_RG  \\\n",
       "766     23.3334  -0.8506  -0.7833  -0.6849       2.8545       2.7554   \n",
       "1879    19.6101  -0.9200  -0.8737  -0.6758       3.3452       3.2793   \n",
       "377     29.3617  -0.9461  -0.8359  -0.7615       2.8244       2.6158   \n",
       "584     21.5302  -0.6926  -0.6324  -0.4304       2.4142       2.2298   \n",
       "1322    23.4475  -0.7043  -0.6322  -0.4200       2.4659       2.3822   \n",
       "1300    24.6418  -1.1376  -1.0231  -0.9553       4.1631       3.5007   \n",
       "556     24.8875  -0.9349  -0.7006  -0.4097       3.3906       2.8394   \n",
       "49      22.8103  -1.0110  -0.9310  -0.6495       3.2229       3.1221   \n",
       "785     21.7716  -1.0000  -0.5618   0.0594       4.0868       3.5127   \n",
       "1263    22.2506  -0.2361  -0.1603  -0.1153       2.3393       2.2724   \n",
       "\n",
       "      Kurtosis_RB              Class  \n",
       "766        2.7177  Kirmizi_Pistachio  \n",
       "1879       3.4488    Siirt_Pistachio  \n",
       "377        2.5660  Kirmizi_Pistachio  \n",
       "584        2.5626  Kirmizi_Pistachio  \n",
       "1322       2.2293    Siirt_Pistachio  \n",
       "1300       3.6535    Siirt_Pistachio  \n",
       "556        2.5640  Kirmizi_Pistachio  \n",
       "49         3.0700  Kirmizi_Pistachio  \n",
       "785        3.3700  Kirmizi_Pistachio  \n",
       "1263       2.1015    Siirt_Pistachio  \n",
       "\n",
       "[10 rows x 29 columns]"
      ]
     },
     "execution_count": 2,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "dataset = pd.read_csv('./Pistachio_Dataset/Pistachio_28_Features_Dataset/Pistachio_28_Features_Dataset.csv') \n",
    "cols = ['Class','Area', 'Perimeter', 'Major_Axis', 'Minor_Axis', 'Eccentricity',\n",
    "        'Eqdiasq', 'Solidity', 'Convex_Area', 'Extent', 'Aspect_Ratio',\n",
    "        'Roundness', 'Compactness', 'Shapefactor_1', 'Shapefactor_2',\n",
    "        'Shapefactor_3', 'Shapefactor_4', 'Mean_RR', 'Mean_RG', 'Mean_RB',\n",
    "        'StdDev_RR', 'StdDev_RG', 'StdDev_RB', 'Skew_RR', 'Skew_RG', 'Skew_RB',\n",
    "        'Kurtosis_RR', 'Kurtosis_RG', 'Kurtosis_RB']\n",
    "# dataset = dataset.loc[:,cols]\n",
    "# d = {'Class': dataset['Class'].value_counts().index, 'count': dataset['Class'].value_counts()}\n",
    "# df_Class = pd.DataFrame(data=d).reset_index(drop=True)\n",
    "# df_Class\n",
    "\n",
    "dataset.sample(10)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "dataset = shuffle(dataset)\n",
    "Lisclass = dataset['Class'].values\n",
    "LisClassNp = []\n",
    "for i in range(len(Lisclass)):\n",
    "    if (Lisclass[i] == 'Kirmizi_Pistachio'):\n",
    "        LisClassNp.append([0,1])\n",
    "    else:\n",
    "        LisClassNp.append([1,0])\n",
    "Y = np.array(LisClassNp)\n",
    "dataset = dataset.drop(columns='Class')\n",
    "X = dataset.values\n",
    "from sklearn.preprocessing import MinMaxScaler\n",
    "scaler = MinMaxScaler(feature_range=(0, 1))\n",
    "X = scaler.fit_transform(X)\n",
    "#拆分训练集和测试集\n",
    "X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.10, random_state=40)\n",
    "X_train = X_train.reshape(X_train.shape[0],1,X_train.shape[1])\n",
    "X_test = X_test.reshape(X_test.shape[0],1,X_test.shape[1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# d = {'Class': dataset['Class'].value_counts().index, 'count': dataset['Class'].value_counts()}\n",
    "# df_Class = pd.DataFrame(data=d).reset_index(drop=True)\n",
    "# df_Class \n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [],
   "source": [
    "#X.dtype = 'int64'\n",
    "# Y.dtype = 'int64'\n",
    "# X = X.astype(int)\n",
    "\n",
    "# X = X.astype(int)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(2114, 28)\n",
      "(2114, 2)\n"
     ]
    }
   ],
   "source": [
    "print(X.shape)\n",
    "print(Y.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [],
   "source": [
    "import os,sys\n",
    "import shutil\n",
    "\n",
    "shutil.rmtree('models')  \n",
    "os.mkdir('models')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 8,
   "metadata": {},
   "outputs": [],
   "source": [
    "# import tensorflow as tf\n",
    "# def metric_precision(y_true,y_pred):\n",
    "#     TP=tf.reduce_sum(y_true/tf.round(y_pred))\n",
    "#     TN=tf.reduce_sum((1-y_true)(1-tf.round(y_pred)))\n",
    "#     FP=tf.reduce_sum((1-y_true)tf.round(y_pred))\n",
    "#     FN=tf.reduce_sum(y_true(1-tf.round(y_pred)))\n",
    "#     precision=TP/(TP+FP)\n",
    "#     return precision"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "def accsorce(Y_test,yhat):\n",
    "    sorce = 0\n",
    "    for i in range(len(Y_test)):\n",
    "        sorce += roc_auc_score(Y_test[i], yhat[i])\n",
    "\n",
    "    return sorce/len(Y_test)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 10,
   "metadata": {},
   "outputs": [],
   "source": [
    "X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.1, random_state=40)\n",
    "X_train = X_train.reshape(X_train.shape[0],1,X_train.shape[1])\n",
    "X_test = X_test.reshape(X_test.shape[0],1,X_test.shape[1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "<class 'numpy.ndarray'>\n"
     ]
    }
   ],
   "source": [
    "print(type(X_train))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "# X_train = torch.from_numpy(X_train)\n",
    "# Y_train = tuple(map(tuple, Y_train))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [],
   "source": [
    "# def Input():\n",
    "#     X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.1, random_state=40)\n",
    "#     X_train = X_train.reshape(X_train.shape[0],1,X_train.shape[1])\n",
    "#     X_test = X_test.reshape(X_test.shape[0],1,X_test.shape[1])\n",
    "#     X = np.array(X)\n",
    "#     return torch.from_numpy(X)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [],
   "source": [
    "# def sigmoid(x):\n",
    "#     return  1/(1+np.exp(-x))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {},
   "outputs": [],
   "source": [
    "# 参考LSTM结构图\n",
    "# https://blog.csdn.net/weixin_42175217/article/details/106183682\n",
    "class ConvLSTM(Model):\n",
    "    def __init__(self, inp_dim, oup_dim, kernel, dilation):\n",
    "        super().__init__()\n",
    "        pad_x = int(dilation * (kernel - 1) / 2)\n",
    "        self.conv_xf = Conv2D(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)\n",
    "        self.conv_xi = Conv2D(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)\n",
    "        self.conv_xo = Conv2D(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)\n",
    "        self.conv_xj = Conv2D(inp_dim, oup_dim, kernel, padding=pad_x, dilation=dilation)\n",
    "\n",
    "        pad_h = int((kernel - 1) / 2)\n",
    "        self.conv_hf = Conv2D(oup_dim, oup_dim, kernel, padding=pad_h)\n",
    "        self.conv_hi = Conv2D(oup_dim, oup_dim, kernel, padding=pad_h)\n",
    "        self.conv_ho = Conv2D(oup_dim, oup_dim, kernel, padding=pad_h)\n",
    "        self.conv_hj = Conv2D(oup_dim, oup_dim, kernel, padding=pad_h)\n",
    "\n",
    "        self.relu = np.LeakyReLU(0.2)\n",
    "\n",
    "    def forward(self, x, pair=None):\n",
    "        if pair is None:\n",
    "            i = np.sigmoid(self.conv_xi(x))\n",
    "            o = np.sigmoid(self.conv_xo(x))\n",
    "            j = np.tanh(self.conv_xj(x))\n",
    "            c = i * j\n",
    "            h = o * c\n",
    "        else:\n",
    "            h, c = pair\n",
    "            f = np.sigmoid(self.conv_xf(x) + self.conv_hf(h))\n",
    "            i = np.sigmoid(self.conv_xi(x) + self.conv_hi(h))\n",
    "            o = np.sigmoid(self.conv_xo(x) + self.conv_ho(h))\n",
    "            j = np.tanh(self.conv_xj(x) + self.conv_hj(h))\n",
    "            c = f * c + i * j\n",
    "            h = o * np.tanh(c)\n",
    "\n",
    "        h = self.relu(h)\n",
    "        return h, [h, c]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "inputdata KerasTensor(type_spec=TensorSpec(shape=(None, 1, 28), dtype=tf.float32, name='input_1'), name='input_1', description=\"created by layer 'input_1'\")\n",
      "Model: \"MyModelMLP\"\n",
      "__________________________________________________________________________________________________\n",
      "Layer (type)                    Output Shape         Param #     Connected to                     \n",
      "==================================================================================================\n",
      "input_1 (InputLayer)            [(None, 1, 28)]      0                                            \n",
      "__________________________________________________________________________________________________\n",
      "conv1d (Conv1D)                 (None, 1, 128)       3712        input_1[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "max_pooling1d (MaxPooling1D)    (None, 1, 128)       0           conv1d[0][0]                     \n",
      "__________________________________________________________________________________________________\n",
      "dropout (Dropout)               multiple             0           max_pooling1d[0][0]              \n",
      "                                                                 lstm_1[0][0]                     \n",
      "                                                                 lstm_1[1][0]                     \n",
      "                                                                 dense[0][0]                      \n",
      "__________________________________________________________________________________________________\n",
      "lstm_1 (LSTM)                   (None, 1, 128)       131584      dropout[0][0]                    \n",
      "                                                                 batch_normalization[0][0]        \n",
      "__________________________________________________________________________________________________\n",
      "batch_normalization (BatchNorma (None, 1, 128)       512         dropout[1][0]                    \n",
      "                                                                 dropout[2][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "dense (Dense)                   (None, 1, 8)         1032        batch_normalization[1][0]        \n",
      "__________________________________________________________________________________________________\n",
      "flatten (Flatten)               (None, 8)            0           dropout[3][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "dense_1 (Dense)                 (None, 2)            18          flatten[0][0]                    \n",
      "==================================================================================================\n",
      "Total params: 136,858\n",
      "Trainable params: 136,602\n",
      "Non-trainable params: 256\n",
      "__________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "class my_model(Model):\n",
    "    def __init__(self, input_shape, **kwargs):\n",
    "        super().__init__(**kwargs) \n",
    "        \n",
    "        the_units = 128\n",
    "        self.CNN   = Conv1D(the_units,1,padding=\"valid\",activation=\"relu\")\n",
    "        self.POLL  = MaxPool1D(1)\n",
    "        self.FLATT = Flatten()\n",
    "        self.LSTM0 = LSTM(the_units,return_sequences=True,\n",
    "                        kernel_initializer=tf.keras.initializers.glorot_normal(seed=2),bias_initializer=tf.keras.initializers.Zeros())\n",
    "        self.LSTM1 = LSTM(the_units,return_sequences=True)\n",
    "        self.LSTM2 = LSTM(the_units)\n",
    "        self.BAT   = BatchNormalization()\n",
    "        self.DROP  = Dropout(rate=0.3)\n",
    "        self.DENS  = Dense(8, activation='relu')\n",
    "        self.OUT   = Dense(2, activation='softmax')\n",
    "        self.resha = Reshape((1, 28))\n",
    "        \n",
    "        # 模型的参数\n",
    "        self.input_layer = tf.keras.layers.Input(input_shape)\n",
    "        self.out = self.call(self.input_layer)\n",
    "        \n",
    "        super().__init__( inputs=self.input_layer,outputs=self.out, **kwargs)\n",
    "    \n",
    "    # 初始化模型的参数\n",
    "    def build(self):\n",
    "        self._is_graph_network = True\n",
    "        self._init_graph_network(inputs=self.input_layer,outputs=self.out)\n",
    "    \n",
    "    def call(self, inputdata, **kwargs):\n",
    "                \n",
    "        #x1 = self.LSTM0(inputdata)\n",
    "        print('inputdata',inputdata)\n",
    "        x1 = self.CNN(inputdata)\n",
    "        #x1 = self.LSTM0(inputdata)\n",
    "        x1 = self.POLL(x1)\n",
    "        c = self.DROP(x1)\n",
    "        #c = self.BAT(x1)\n",
    "\n",
    "        c = self.LSTM1(c)\n",
    "        c = self.DROP(c)\n",
    "        c = self.BAT(c)\n",
    "        \n",
    "        c = self.LSTM1(c)\n",
    "        c = self.DROP(c)\n",
    "        c = self.BAT(c)\n",
    "        \n",
    "        # # d = self.LSTM2(c)\n",
    "        # # d = self.BAT(d)\n",
    "        # c = self.LSTM1(c)\n",
    "        # c = self.DROP(c)\n",
    "        # c = self.BAT(c)\n",
    "                        \n",
    "        d = self.DENS(c)\n",
    "        d = self.DROP(d)\n",
    "        \n",
    "        d = self.FLATT(d)\n",
    "        out = self.OUT(d)\n",
    "        return out\n",
    "    # AFAIK: The most convenient method to print model.summary() \n",
    "    # similar to the sequential or functional API like.\n",
    "    def build_graph(self):\n",
    "        #x = Input(shape=(dim))\n",
    "        x = Input()\n",
    "        return Model(inputs=[x], outputs=self.call(x))\n",
    "\n",
    "mymodel2 = my_model((1,28), name='MyModelMLP')\n",
    "# mymodel2(X_train)\n",
    "mymodel2.summary()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<IPython.core.display.Image object>"
      ]
     },
     "execution_count": 17,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "plot_model(mymodel2, to_file='mymodel2.png', show_shapes=True, \n",
    "                        show_layer_names=True, rankdir='TB', dpi=100, expand_nested=True)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "Epoch = 100  # 模型迭代的次数\n",
    "Batch_Size = 64  # 批量训练的样本的个数\n",
    "Out_Class = 2  # 输出的类别的个数,0-9共10类\n",
    "# 模型编译\n",
    "#mYMODEL = my_model((1,28),name=\"mYMODEL\")\n",
    "#mYMODEL(X_train)\n",
    "mymodel2.compile(loss='binary_crossentropy',#categorical_crossentropy', binary_crossentropy\n",
    "                optimizer='adam', metrics=['categorical_accuracy'])  \n",
    "\n",
    "# train_label_cate = tf.keras.utils.to_categorical(train_labels, 10)\n",
    "# testlabels = tf.keras.utils.to_categorical(test_labels, 10)\n",
    "checkpoint_path = \"./models/cp-{categorical_accuracy:.5f}.ckpt\"\n",
    "checkpoint_dir = os.path.dirname(checkpoint_path)\n",
    "\n",
    "# 创建一个回调，保证验证数据集准确率最大\n",
    "cp_callback = tf.keras.callbacks.ModelCheckpoint(filepath=checkpoint_path,\n",
    "                                                save_weights_only=True,\n",
    "                                                monitor='categorical_accuracy',\n",
    "                                                mode='max',\n",
    "                                                verbose=2,\n",
    "                                                save_best_only=True)\n",
    "#                                              filepath=filepath,\n",
    "#                                             save_weights_only=False,\n",
    "#                                             monitor='categorical_accuracy',\n",
    "#                                             mode='max',\n",
    "#                                             save_best_only=True)\n",
    "\n",
    "# 动态更改学习率：在模型的回调中使用\n",
    "def scheduler(epoch):  # 根据epoch动态更改学习率的参数\n",
    "    if epoch < 10:\n",
    "        return 0.13\n",
    "    else:\n",
    "        return 0.13 * tf.math.exp(0.1 * (10 - epoch))\n",
    "    \n",
    "lr_back = tf.keras.callbacks.LearningRateScheduler(scheduler)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 1/100\n",
      "inputdata Tensor(\"IteratorGetNext:0\", shape=(None, 1, 28), dtype=float32)\n",
      "inputdata Tensor(\"IteratorGetNext:0\", shape=(None, 1, 28), dtype=float32)\n",
      "inputdata Tensor(\"IteratorGetNext:0\", shape=(None, 1, 28), dtype=float32)\n",
      "27/27 - 8s - loss: 0.7650 - categorical_accuracy: 0.6137 - val_loss: 0.6877 - val_categorical_accuracy: 0.5602\n",
      "\n",
      "Epoch 00001: categorical_accuracy improved from -inf to 0.61368, saving model to ./models\\cp-0.61368.ckpt\n",
      "Epoch 2/100\n",
      "27/27 - 0s - loss: 0.5787 - categorical_accuracy: 0.7481 - val_loss: 0.6753 - val_categorical_accuracy: 0.5602\n",
      "\n",
      "Epoch 00002: categorical_accuracy improved from 0.61368 to 0.74810, saving model to ./models\\cp-0.74810.ckpt\n",
      "Epoch 3/100\n",
      "27/27 - 0s - loss: 0.5064 - categorical_accuracy: 0.8153 - val_loss: 0.6580 - val_categorical_accuracy: 0.5759\n",
      "\n",
      "Epoch 00003: categorical_accuracy improved from 0.74810 to 0.81531, saving model to ./models\\cp-0.81531.ckpt\n",
      "Epoch 4/100\n",
      "27/27 - 0s - loss: 0.4616 - categorical_accuracy: 0.8469 - val_loss: 0.6355 - val_categorical_accuracy: 0.6545\n",
      "\n",
      "Epoch 00004: categorical_accuracy improved from 0.81531 to 0.84687, saving model to ./models\\cp-0.84687.ckpt\n",
      "Epoch 5/100\n",
      "27/27 - 0s - loss: 0.4055 - categorical_accuracy: 0.8755 - val_loss: 0.5985 - val_categorical_accuracy: 0.7906\n",
      "\n",
      "Epoch 00005: categorical_accuracy improved from 0.84687 to 0.87551, saving model to ./models\\cp-0.87551.ckpt\n",
      "Epoch 6/100\n",
      "27/27 - 0s - loss: 0.3940 - categorical_accuracy: 0.8685 - val_loss: 0.5679 - val_categorical_accuracy: 0.8272\n",
      "\n",
      "Epoch 00006: categorical_accuracy did not improve from 0.87551\n",
      "Epoch 7/100\n",
      "27/27 - 0s - loss: 0.3797 - categorical_accuracy: 0.8767 - val_loss: 0.5238 - val_categorical_accuracy: 0.8796\n",
      "\n",
      "Epoch 00007: categorical_accuracy improved from 0.87551 to 0.87668, saving model to ./models\\cp-0.87668.ckpt\n",
      "Epoch 8/100\n",
      "27/27 - 0s - loss: 0.3659 - categorical_accuracy: 0.8819 - val_loss: 0.4627 - val_categorical_accuracy: 0.8429\n",
      "\n",
      "Epoch 00008: categorical_accuracy improved from 0.87668 to 0.88194, saving model to ./models\\cp-0.88194.ckpt\n",
      "Epoch 9/100\n",
      "27/27 - 0s - loss: 0.3564 - categorical_accuracy: 0.8790 - val_loss: 0.4051 - val_categorical_accuracy: 0.8796\n",
      "\n",
      "Epoch 00009: categorical_accuracy did not improve from 0.88194\n",
      "Epoch 10/100\n",
      "27/27 - 0s - loss: 0.3502 - categorical_accuracy: 0.8872 - val_loss: 0.3539 - val_categorical_accuracy: 0.8953\n",
      "\n",
      "Epoch 00010: categorical_accuracy improved from 0.88194 to 0.88720, saving model to ./models\\cp-0.88720.ckpt\n",
      "Epoch 11/100\n",
      "27/27 - 0s - loss: 0.3488 - categorical_accuracy: 0.8814 - val_loss: 0.3248 - val_categorical_accuracy: 0.9110\n",
      "\n",
      "Epoch 00011: categorical_accuracy did not improve from 0.88720\n",
      "Epoch 12/100\n",
      "27/27 - 0s - loss: 0.3292 - categorical_accuracy: 0.8802 - val_loss: 0.2559 - val_categorical_accuracy: 0.9110\n",
      "\n",
      "Epoch 00012: categorical_accuracy did not improve from 0.88720\n",
      "Epoch 13/100\n",
      "27/27 - 0s - loss: 0.3261 - categorical_accuracy: 0.8936 - val_loss: 0.2659 - val_categorical_accuracy: 0.9110\n",
      "\n",
      "Epoch 00013: categorical_accuracy improved from 0.88720 to 0.89363, saving model to ./models\\cp-0.89363.ckpt\n",
      "Epoch 14/100\n",
      "27/27 - 0s - loss: 0.3205 - categorical_accuracy: 0.8773 - val_loss: 0.2537 - val_categorical_accuracy: 0.9319\n",
      "\n",
      "Epoch 00014: categorical_accuracy did not improve from 0.89363\n",
      "Epoch 15/100\n",
      "27/27 - 0s - loss: 0.3210 - categorical_accuracy: 0.8901 - val_loss: 0.2313 - val_categorical_accuracy: 0.9215\n",
      "\n",
      "Epoch 00015: categorical_accuracy did not improve from 0.89363\n",
      "Epoch 16/100\n",
      "27/27 - 0s - loss: 0.3087 - categorical_accuracy: 0.8907 - val_loss: 0.2500 - val_categorical_accuracy: 0.9215\n",
      "\n",
      "Epoch 00016: categorical_accuracy did not improve from 0.89363\n",
      "Epoch 17/100\n",
      "27/27 - 0s - loss: 0.3061 - categorical_accuracy: 0.8936 - val_loss: 0.2399 - val_categorical_accuracy: 0.9372\n",
      "\n",
      "Epoch 00017: categorical_accuracy did not improve from 0.89363\n",
      "Epoch 18/100\n",
      "27/27 - 0s - loss: 0.2831 - categorical_accuracy: 0.9047 - val_loss: 0.2501 - val_categorical_accuracy: 0.9162\n",
      "\n",
      "Epoch 00018: categorical_accuracy improved from 0.89363 to 0.90473, saving model to ./models\\cp-0.90473.ckpt\n",
      "Epoch 19/100\n",
      "27/27 - 0s - loss: 0.3071 - categorical_accuracy: 0.8825 - val_loss: 0.2333 - val_categorical_accuracy: 0.9267\n",
      "\n",
      "Epoch 00019: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 20/100\n",
      "27/27 - 0s - loss: 0.3074 - categorical_accuracy: 0.8808 - val_loss: 0.2409 - val_categorical_accuracy: 0.9058\n",
      "\n",
      "Epoch 00020: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 21/100\n",
      "27/27 - 0s - loss: 0.2903 - categorical_accuracy: 0.8854 - val_loss: 0.2189 - val_categorical_accuracy: 0.9319\n",
      "\n",
      "Epoch 00021: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 22/100\n",
      "27/27 - 0s - loss: 0.2839 - categorical_accuracy: 0.8948 - val_loss: 0.2004 - val_categorical_accuracy: 0.9267\n",
      "\n",
      "Epoch 00022: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 23/100\n",
      "27/27 - 0s - loss: 0.2687 - categorical_accuracy: 0.9030 - val_loss: 0.2195 - val_categorical_accuracy: 0.9267\n",
      "\n",
      "Epoch 00023: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 24/100\n",
      "27/27 - 0s - loss: 0.2959 - categorical_accuracy: 0.8878 - val_loss: 0.2135 - val_categorical_accuracy: 0.9162\n",
      "\n",
      "Epoch 00024: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 25/100\n",
      "27/27 - 0s - loss: 0.2879 - categorical_accuracy: 0.8884 - val_loss: 0.2294 - val_categorical_accuracy: 0.9267\n",
      "\n",
      "Epoch 00025: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 26/100\n",
      "27/27 - 0s - loss: 0.2732 - categorical_accuracy: 0.8925 - val_loss: 0.2013 - val_categorical_accuracy: 0.9110\n",
      "\n",
      "Epoch 00026: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 27/100\n",
      "27/27 - 0s - loss: 0.2685 - categorical_accuracy: 0.8872 - val_loss: 0.2137 - val_categorical_accuracy: 0.9110\n",
      "\n",
      "Epoch 00027: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 28/100\n",
      "27/27 - 0s - loss: 0.2852 - categorical_accuracy: 0.8971 - val_loss: 0.2503 - val_categorical_accuracy: 0.9110\n",
      "\n",
      "Epoch 00028: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 29/100\n",
      "27/27 - 0s - loss: 0.2868 - categorical_accuracy: 0.8854 - val_loss: 0.2187 - val_categorical_accuracy: 0.9110\n",
      "\n",
      "Epoch 00029: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 30/100\n",
      "27/27 - 0s - loss: 0.2684 - categorical_accuracy: 0.8983 - val_loss: 0.2423 - val_categorical_accuracy: 0.9267\n",
      "\n",
      "Epoch 00030: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 31/100\n",
      "27/27 - 0s - loss: 0.2644 - categorical_accuracy: 0.8936 - val_loss: 0.2534 - val_categorical_accuracy: 0.9058\n",
      "\n",
      "Epoch 00031: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 32/100\n",
      "27/27 - 0s - loss: 0.2768 - categorical_accuracy: 0.8989 - val_loss: 0.2608 - val_categorical_accuracy: 0.9005\n",
      "\n",
      "Epoch 00032: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 33/100\n",
      "27/27 - 0s - loss: 0.2765 - categorical_accuracy: 0.8960 - val_loss: 0.2216 - val_categorical_accuracy: 0.9215\n",
      "\n",
      "Epoch 00033: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 34/100\n",
      "27/27 - 0s - loss: 0.2619 - categorical_accuracy: 0.8942 - val_loss: 0.2182 - val_categorical_accuracy: 0.9215\n",
      "\n",
      "Epoch 00034: categorical_accuracy did not improve from 0.90473\n",
      "Epoch 35/100\n",
      "27/27 - 0s - loss: 0.2555 - categorical_accuracy: 0.9082 - val_loss: 0.2277 - val_categorical_accuracy: 0.9215\n",
      "\n",
      "Epoch 00035: categorical_accuracy improved from 0.90473 to 0.90824, saving model to ./models\\cp-0.90824.ckpt\n",
      "Epoch 36/100\n",
      "27/27 - 0s - loss: 0.2542 - categorical_accuracy: 0.9018 - val_loss: 0.2377 - val_categorical_accuracy: 0.9005\n",
      "\n",
      "Epoch 00036: categorical_accuracy did not improve from 0.90824\n",
      "Epoch 37/100\n",
      "27/27 - 0s - loss: 0.2526 - categorical_accuracy: 0.9106 - val_loss: 0.2117 - val_categorical_accuracy: 0.9319\n",
      "\n",
      "Epoch 00037: categorical_accuracy improved from 0.90824 to 0.91058, saving model to ./models\\cp-0.91058.ckpt\n",
      "Epoch 38/100\n",
      "27/27 - 0s - loss: 0.2684 - categorical_accuracy: 0.8878 - val_loss: 0.2814 - val_categorical_accuracy: 0.8953\n",
      "\n",
      "Epoch 00038: categorical_accuracy did not improve from 0.91058\n",
      "Epoch 39/100\n",
      "27/27 - 0s - loss: 0.2783 - categorical_accuracy: 0.8936 - val_loss: 0.2834 - val_categorical_accuracy: 0.8953\n",
      "\n",
      "Epoch 00039: categorical_accuracy did not improve from 0.91058\n",
      "Epoch 40/100\n",
      "27/27 - 0s - loss: 0.2708 - categorical_accuracy: 0.8930 - val_loss: 0.2717 - val_categorical_accuracy: 0.8848\n",
      "\n",
      "Epoch 00040: categorical_accuracy did not improve from 0.91058\n",
      "Epoch 41/100\n",
      "27/27 - 0s - loss: 0.2523 - categorical_accuracy: 0.8977 - val_loss: 0.2796 - val_categorical_accuracy: 0.8848\n",
      "\n",
      "Epoch 00041: categorical_accuracy did not improve from 0.91058\n",
      "Epoch 42/100\n",
      "27/27 - 1s - loss: 0.2694 - categorical_accuracy: 0.9082 - val_loss: 0.2684 - val_categorical_accuracy: 0.8953\n",
      "\n",
      "Epoch 00042: categorical_accuracy did not improve from 0.91058\n",
      "Epoch 43/100\n",
      "27/27 - 0s - loss: 0.2430 - categorical_accuracy: 0.9001 - val_loss: 0.2874 - val_categorical_accuracy: 0.8743\n",
      "\n",
      "Epoch 00043: categorical_accuracy did not improve from 0.91058\n",
      "Epoch 44/100\n",
      "27/27 - 0s - loss: 0.2572 - categorical_accuracy: 0.9065 - val_loss: 0.3305 - val_categorical_accuracy: 0.8482\n",
      "\n",
      "Epoch 00044: categorical_accuracy did not improve from 0.91058\n",
      "Epoch 45/100\n",
      "27/27 - 0s - loss: 0.2561 - categorical_accuracy: 0.8977 - val_loss: 0.3036 - val_categorical_accuracy: 0.8691\n",
      "\n",
      "Epoch 00045: categorical_accuracy did not improve from 0.91058\n",
      "Epoch 46/100\n",
      "27/27 - 0s - loss: 0.2521 - categorical_accuracy: 0.9041 - val_loss: 0.2564 - val_categorical_accuracy: 0.8953\n",
      "\n",
      "Epoch 00046: categorical_accuracy did not improve from 0.91058\n",
      "Epoch 47/100\n",
      "27/27 - 0s - loss: 0.2391 - categorical_accuracy: 0.9123 - val_loss: 0.3723 - val_categorical_accuracy: 0.8115\n",
      "\n",
      "Epoch 00047: categorical_accuracy improved from 0.91058 to 0.91233, saving model to ./models\\cp-0.91233.ckpt\n",
      "Epoch 48/100\n",
      "27/27 - 0s - loss: 0.2519 - categorical_accuracy: 0.8989 - val_loss: 0.3043 - val_categorical_accuracy: 0.8586\n",
      "\n",
      "Epoch 00048: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 49/100\n",
      "27/27 - 0s - loss: 0.2375 - categorical_accuracy: 0.9053 - val_loss: 0.3149 - val_categorical_accuracy: 0.8429\n",
      "\n",
      "Epoch 00049: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 50/100\n",
      "27/27 - 0s - loss: 0.2369 - categorical_accuracy: 0.9041 - val_loss: 0.3053 - val_categorical_accuracy: 0.8429\n",
      "\n",
      "Epoch 00050: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 51/100\n",
      "27/27 - 0s - loss: 0.2399 - categorical_accuracy: 0.9041 - val_loss: 0.3764 - val_categorical_accuracy: 0.8063\n",
      "\n",
      "Epoch 00051: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 52/100\n",
      "27/27 - 0s - loss: 0.2394 - categorical_accuracy: 0.8989 - val_loss: 0.3520 - val_categorical_accuracy: 0.8168\n",
      "\n",
      "Epoch 00052: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 53/100\n",
      "27/27 - 0s - loss: 0.2383 - categorical_accuracy: 0.9100 - val_loss: 0.3778 - val_categorical_accuracy: 0.8010\n",
      "\n",
      "Epoch 00053: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 54/100\n",
      "27/27 - 0s - loss: 0.2502 - categorical_accuracy: 0.9047 - val_loss: 0.2767 - val_categorical_accuracy: 0.8953\n",
      "\n",
      "Epoch 00054: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 55/100\n",
      "27/27 - 0s - loss: 0.2487 - categorical_accuracy: 0.9030 - val_loss: 0.2643 - val_categorical_accuracy: 0.9162\n",
      "\n",
      "Epoch 00055: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 56/100\n",
      "27/27 - 0s - loss: 0.2405 - categorical_accuracy: 0.9053 - val_loss: 0.3788 - val_categorical_accuracy: 0.7958\n",
      "\n",
      "Epoch 00056: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 57/100\n",
      "27/27 - 0s - loss: 0.2358 - categorical_accuracy: 0.9082 - val_loss: 0.4322 - val_categorical_accuracy: 0.7225\n",
      "\n",
      "Epoch 00057: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 58/100\n",
      "27/27 - 0s - loss: 0.2599 - categorical_accuracy: 0.9094 - val_loss: 0.2477 - val_categorical_accuracy: 0.9110\n",
      "\n",
      "Epoch 00058: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 59/100\n",
      "27/27 - 0s - loss: 0.2519 - categorical_accuracy: 0.9041 - val_loss: 0.4160 - val_categorical_accuracy: 0.7749\n",
      "\n",
      "Epoch 00059: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 60/100\n",
      "27/27 - 0s - loss: 0.2406 - categorical_accuracy: 0.8966 - val_loss: 0.3353 - val_categorical_accuracy: 0.8377\n",
      "\n",
      "Epoch 00060: categorical_accuracy did not improve from 0.91233\n",
      "Epoch 61/100\n",
      "27/27 - 0s - loss: 0.2353 - categorical_accuracy: 0.9153 - val_loss: 0.4167 - val_categorical_accuracy: 0.7435\n",
      "\n",
      "Epoch 00061: categorical_accuracy improved from 0.91233 to 0.91525, saving model to ./models\\cp-0.91525.ckpt\n",
      "Epoch 62/100\n",
      "27/27 - 0s - loss: 0.2380 - categorical_accuracy: 0.9100 - val_loss: 0.3129 - val_categorical_accuracy: 0.8534\n",
      "\n",
      "Epoch 00062: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 63/100\n",
      "27/27 - 0s - loss: 0.2448 - categorical_accuracy: 0.9100 - val_loss: 0.3482 - val_categorical_accuracy: 0.7958\n",
      "\n",
      "Epoch 00063: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 64/100\n",
      "27/27 - 0s - loss: 0.2587 - categorical_accuracy: 0.9036 - val_loss: 0.3203 - val_categorical_accuracy: 0.8429\n",
      "\n",
      "Epoch 00064: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 65/100\n",
      "27/27 - 0s - loss: 0.2245 - categorical_accuracy: 0.9106 - val_loss: 0.3109 - val_categorical_accuracy: 0.8429\n",
      "\n",
      "Epoch 00065: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 66/100\n",
      "27/27 - 0s - loss: 0.2448 - categorical_accuracy: 0.9024 - val_loss: 0.2636 - val_categorical_accuracy: 0.8901\n",
      "\n",
      "Epoch 00066: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 67/100\n",
      "27/27 - 0s - loss: 0.2508 - categorical_accuracy: 0.9036 - val_loss: 0.3033 - val_categorical_accuracy: 0.8325\n",
      "\n",
      "Epoch 00067: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 68/100\n",
      "27/27 - 0s - loss: 0.2387 - categorical_accuracy: 0.9059 - val_loss: 0.3019 - val_categorical_accuracy: 0.8377\n",
      "\n",
      "Epoch 00068: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 69/100\n",
      "27/27 - 0s - loss: 0.2488 - categorical_accuracy: 0.9047 - val_loss: 0.2582 - val_categorical_accuracy: 0.9215\n",
      "\n",
      "Epoch 00069: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 70/100\n",
      "27/27 - 0s - loss: 0.2328 - categorical_accuracy: 0.9094 - val_loss: 0.2494 - val_categorical_accuracy: 0.8901\n",
      "\n",
      "Epoch 00070: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 71/100\n",
      "27/27 - 0s - loss: 0.2451 - categorical_accuracy: 0.9030 - val_loss: 0.2072 - val_categorical_accuracy: 0.9267\n",
      "\n",
      "Epoch 00071: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 72/100\n",
      "27/27 - 0s - loss: 0.2339 - categorical_accuracy: 0.9117 - val_loss: 0.2504 - val_categorical_accuracy: 0.9319\n",
      "\n",
      "Epoch 00072: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 73/100\n",
      "27/27 - 0s - loss: 0.2559 - categorical_accuracy: 0.9088 - val_loss: 0.2497 - val_categorical_accuracy: 0.9267\n",
      "\n",
      "Epoch 00073: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 74/100\n",
      "27/27 - 0s - loss: 0.2499 - categorical_accuracy: 0.9030 - val_loss: 0.2615 - val_categorical_accuracy: 0.9372\n",
      "\n",
      "Epoch 00074: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 75/100\n",
      "27/27 - 0s - loss: 0.2343 - categorical_accuracy: 0.9059 - val_loss: 0.2204 - val_categorical_accuracy: 0.9372\n",
      "\n",
      "Epoch 00075: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 76/100\n",
      "27/27 - 0s - loss: 0.2498 - categorical_accuracy: 0.9106 - val_loss: 0.2055 - val_categorical_accuracy: 0.9424\n",
      "\n",
      "Epoch 00076: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 77/100\n",
      "27/27 - 0s - loss: 0.2294 - categorical_accuracy: 0.9071 - val_loss: 0.2480 - val_categorical_accuracy: 0.9058\n",
      "\n",
      "Epoch 00077: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 78/100\n",
      "27/27 - 0s - loss: 0.2371 - categorical_accuracy: 0.9012 - val_loss: 0.2141 - val_categorical_accuracy: 0.9215\n",
      "\n",
      "Epoch 00078: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 79/100\n",
      "27/27 - 0s - loss: 0.2358 - categorical_accuracy: 0.9135 - val_loss: 0.2221 - val_categorical_accuracy: 0.9267\n",
      "\n",
      "Epoch 00079: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 80/100\n",
      "27/27 - 0s - loss: 0.2317 - categorical_accuracy: 0.9147 - val_loss: 0.2225 - val_categorical_accuracy: 0.9162\n",
      "\n",
      "Epoch 00080: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 81/100\n",
      "27/27 - 0s - loss: 0.2274 - categorical_accuracy: 0.9100 - val_loss: 0.2447 - val_categorical_accuracy: 0.9267\n",
      "\n",
      "Epoch 00081: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 82/100\n",
      "27/27 - 0s - loss: 0.2298 - categorical_accuracy: 0.9129 - val_loss: 0.2346 - val_categorical_accuracy: 0.9267\n",
      "\n",
      "Epoch 00082: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 83/100\n",
      "27/27 - 0s - loss: 0.2248 - categorical_accuracy: 0.9117 - val_loss: 0.2327 - val_categorical_accuracy: 0.9372\n",
      "\n",
      "Epoch 00083: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 84/100\n",
      "27/27 - 0s - loss: 0.2255 - categorical_accuracy: 0.9135 - val_loss: 0.2587 - val_categorical_accuracy: 0.9058\n",
      "\n",
      "Epoch 00084: categorical_accuracy did not improve from 0.91525\n",
      "Epoch 85/100\n",
      "27/27 - 0s - loss: 0.2230 - categorical_accuracy: 0.9205 - val_loss: 0.2495 - val_categorical_accuracy: 0.9319\n",
      "\n",
      "Epoch 00085: categorical_accuracy improved from 0.91525 to 0.92051, saving model to ./models\\cp-0.92051.ckpt\n",
      "Epoch 86/100\n",
      "27/27 - 0s - loss: 0.2187 - categorical_accuracy: 0.9199 - val_loss: 0.2664 - val_categorical_accuracy: 0.9372\n",
      "\n",
      "Epoch 00086: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 87/100\n",
      "27/27 - 0s - loss: 0.2175 - categorical_accuracy: 0.9164 - val_loss: 0.2491 - val_categorical_accuracy: 0.9267\n",
      "\n",
      "Epoch 00087: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 88/100\n",
      "27/27 - 0s - loss: 0.2176 - categorical_accuracy: 0.9170 - val_loss: 0.2463 - val_categorical_accuracy: 0.9215\n",
      "\n",
      "Epoch 00088: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 89/100\n",
      "27/27 - 0s - loss: 0.2245 - categorical_accuracy: 0.9047 - val_loss: 0.2500 - val_categorical_accuracy: 0.9110\n",
      "\n",
      "Epoch 00089: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 90/100\n",
      "27/27 - 0s - loss: 0.2396 - categorical_accuracy: 0.9088 - val_loss: 0.2465 - val_categorical_accuracy: 0.9162\n",
      "\n",
      "Epoch 00090: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 91/100\n",
      "27/27 - 0s - loss: 0.2333 - categorical_accuracy: 0.9059 - val_loss: 0.2502 - val_categorical_accuracy: 0.9110\n",
      "\n",
      "Epoch 00091: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 92/100\n",
      "27/27 - 0s - loss: 0.2147 - categorical_accuracy: 0.9193 - val_loss: 0.2934 - val_categorical_accuracy: 0.9162\n",
      "\n",
      "Epoch 00092: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 93/100\n",
      "27/27 - 0s - loss: 0.2270 - categorical_accuracy: 0.9117 - val_loss: 0.2708 - val_categorical_accuracy: 0.9058\n",
      "\n",
      "Epoch 00093: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 94/100\n",
      "27/27 - 0s - loss: 0.2238 - categorical_accuracy: 0.9153 - val_loss: 0.2086 - val_categorical_accuracy: 0.9372\n",
      "\n",
      "Epoch 00094: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 95/100\n",
      "27/27 - 0s - loss: 0.2490 - categorical_accuracy: 0.9053 - val_loss: 0.2913 - val_categorical_accuracy: 0.9005\n",
      "\n",
      "Epoch 00095: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 96/100\n",
      "27/27 - 0s - loss: 0.2515 - categorical_accuracy: 0.9100 - val_loss: 0.2533 - val_categorical_accuracy: 0.9372\n",
      "\n",
      "Epoch 00096: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 97/100\n",
      "27/27 - 0s - loss: 0.2152 - categorical_accuracy: 0.9147 - val_loss: 0.2200 - val_categorical_accuracy: 0.9424\n",
      "\n",
      "Epoch 00097: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 98/100\n",
      "27/27 - 0s - loss: 0.2244 - categorical_accuracy: 0.9077 - val_loss: 0.2511 - val_categorical_accuracy: 0.9162\n",
      "\n",
      "Epoch 00098: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 99/100\n",
      "27/27 - 0s - loss: 0.2253 - categorical_accuracy: 0.9094 - val_loss: 0.2387 - val_categorical_accuracy: 0.9319\n",
      "\n",
      "Epoch 00099: categorical_accuracy did not improve from 0.92051\n",
      "Epoch 100/100\n",
      "27/27 - 0s - loss: 0.2244 - categorical_accuracy: 0.9117 - val_loss: 0.2441 - val_categorical_accuracy: 0.9110\n",
      "\n",
      "Epoch 00100: categorical_accuracy did not improve from 0.92051\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<keras.callbacks.History at 0x23a200e46d0>"
      ]
     },
     "execution_count": 19,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "mymodel2.fit(X_train,Y_train, batch_size=Batch_Size, \n",
    "            epochs=Epoch, verbose=2, validation_split=0.1, \n",
    "            callbacks=cp_callback)\n",
    "            #callbacks=[cp_callback, lr_back])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 20,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "7/7 [==============================] - 0s 2ms/step - loss: 0.3302 - categorical_accuracy: 0.8821\n",
      "测试数据集成本：0.33020797,准确率88.20754886%%\n"
     ]
    }
   ],
   "source": [
    "# 评估模型,按照模型最后的参数计算\n",
    "test_loss, test_acc = mymodel2.evaluate(X_test, Y_test)\n",
    "\n",
    "print('测试数据集成本：{:.8f},准确率{:.8f}%%'. format(test_loss, 100*test_acc))\n",
    "\n",
    "\n",
    "best_para = tf.train.latest_checkpoint(checkpoint_dir)\n",
    "print('最优的参数文件：', best_para)\n",
    "#best_para = './models\\cp-0.92987.ckpt'\n",
    "mymodel2.load_weights(best_para)\n",
    "\n",
    "predict_loss, predict_acc = mymodel2.evaluate(X_test, Y_test)\n",
    "print('使用训练后的参数','成本:', predict_loss, '准确率', predict_acc)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "最优的参数文件： ./models\\cp-0.92051.ckpt\n",
      "7/7 [==============================] - 0s 2ms/step - loss: 0.3158 - categorical_accuracy: 0.9104\n",
      "使用训练后的参数 成本: 0.3158160150051117 准确率 0.9103773832321167\n"
     ]
    }
   ],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "# del mymodel2\n",
    "# # 参数加载\n",
    "# # 新的模型结构保持一致。\n",
    "# model_new = my_model((1,28))\n",
    "# #mymodel2 = my_model((1,28), name='MyModelMLP')\n",
    "# # 需要经过编译，参数也要和原来的一致\n",
    "# model_new.compile(loss='binary_crossentropy',#categorical_crossentropy', binary_crossentropy\n",
    "#                 optimizer='adam', metrics=['categorical_accuracy'])  \n",
    "# # checkpoint_path = \"./models/cp-{categorical_accuracy:.5f}.ckpt\"\n",
    "# # checkpoint_dir = os.path.dirname(checkpoint_path)\n",
    "# # 加载已经训练好的参数\n",
    "# best_para = tf.train.latest_checkpoint(checkpoint_dir)\n",
    "# print('最优的参数文件：', best_para)\n",
    "# model_new.load_weights(best_para)\n",
    "# predict_loss, predict_acc = model_new.evaluate(X_test, Y_test)\n",
    "# print('使用训练后的参数','成本:', predict_loss, '准确率', predict_acc)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {},
   "outputs": [],
   "source": [
    "# class my_model(Model):\n",
    "#     def __init__(self, dim):\n",
    "#         super(my_model, self).__init__()\n",
    "#         the_units = 128\n",
    "#         self.CNN   = Conv1D(16,5,padding=\"same\",activation=\"relu\",input_shape=(dim))\n",
    "#         self.POLL  = MaxPool1D(2)\n",
    "#         self.FLATT = Flatten()\n",
    "#         self.LSTM1 = LSTM(the_units,return_sequences=True)\n",
    "#         self.LSTM2 = LSTM(the_units)\n",
    "#         self.BAT   = BatchNormalization()\n",
    "#         self.DROP  = Dropout(rate=0.1)\n",
    "#         self.DENS  = Dense(the_units, activation='relu')\n",
    "#         self.OUT   = Dense(2, activation='sigmoid')\n",
    "#         self.resha = Reshape((None,1, 128))\n",
    "    \n",
    "#     def call(self, inputs):\n",
    "#         x = self.CNN(inputs)\n",
    "#         x = self.POLL(x)\n",
    "#         x = self.FLATT(x)\n",
    "#         x = self.DENS(x)\n",
    "        \n",
    "#         a = self.resha(x)\n",
    "        \n",
    "#         b = self.LSTM1(a)\n",
    "#         b = self.DROP(a)\n",
    "#         b = self.BAT(b)\n",
    "        \n",
    "#         c = self.LSTM2(b)\n",
    "#         c = self.DROP(c)\n",
    "#         c = self.BAT(c)\n",
    "        \n",
    "#         d = self.DENS(c)\n",
    "#         d = self.DROP(d)\n",
    "#         return self.OUT(d)\n",
    "    \n",
    "#     # AFAIK: The most convenient method to print model.summary() \n",
    "#     # similar to the sequential or functional API like.\n",
    "#     # def build_graph(self):\n",
    "#     #     x = Input(shape=(dim))\n",
    "#     #     return Model(inputs=[x], outputs=self.call(x))\n",
    "\n",
    "# # dim = (124,124,3)\n",
    "# # model = my_model((dim))\n",
    "# # model.build((None, *dim))\n",
    "# # model.build_graph().summary()\n",
    "# dim = (None,28,1)\n",
    "# model = my_model(dim)\n",
    "# model.build(input_shape=dim)\n",
    "# model.summary()\n",
    "\n",
    "# model = tf.keras.utils.plot_model(model=model,\n",
    "#         show_shapes=True, to_file='model.png')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {},
   "outputs": [],
   "source": [
    "\n",
    "# #定义模型\n",
    "# import random\n",
    "# from string import printable\n",
    "# from tokenize import Double\n",
    "# sgd = SGD(learning_rate=0.1, decay=1e-6, momentum=0.9, nesterov=True)\n",
    "\n",
    "# # lstm_layers = [1,2,3,4,5]\n",
    "# # dense_layers = [1,2,3,4,5,6]\n",
    "# # units = [16,32,64,128]\n",
    "# # dropout = [0.05,0.1,0.15,0.25]\n",
    "# # Batch_size = [32,64,128]\n",
    "# # optimizer = ['adam',sgd]\n",
    "# lstm_layers = [2]\n",
    "# dense_layers = [4]\n",
    "# units = [256]\n",
    "# dropout = [0.05]\n",
    "# Batch_size = [64]\n",
    "# optimizer = ['adam']\n",
    "# for the_batch_size in Batch_size:\n",
    "#     for the_dropout in dropout:\n",
    "#         for the_optimizer in optimizer:\n",
    "#             for the_dense_layers in dense_layers:\n",
    "#                 for the_lstm_layers in lstm_layers:\n",
    "#                     for the_units in units:\n",
    "#                         sgd = SGD(learning_rate=0.1, decay=1e-6, momentum=0.9, nesterov=True)\n",
    "#                         X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.1, random_state=40)\n",
    "#                         X_train = X_train.reshape(X_train.shape[0],1,X_train.shape[1])\n",
    "#                         X_test = X_test.reshape(X_test.shape[0],1,X_test.shape[1])\n",
    "#                         model = Sequential()\n",
    "#                         # model.build(input_shape=(277,277,2))\n",
    "#                         #print(model.summary())\n",
    "#                         #model.add(SpatialDropout1D(0.2))\n",
    "#                         print(X_train.shape[0],X_train.shape[1],X_train.shape[2])\n",
    "#                         print('*************')\n",
    "#                         model.add(Conv1D(16,5,padding=\"same\",activation=\"relu\",input_shape=(X_train.shape[2],1)))\n",
    "#                         model.add(MaxPool1D(2))\n",
    "#                         model.add(Flatten())\n",
    "#                         model.add(Dense(the_dense_layers,activation='relu'))\n",
    "#                         #model.add(LSTM(the_units ,input_shape=(X_train.shape[1],X_train.shape[2]),return_sequences = True))\n",
    "#                         model.add(LSTM(the_units,return_sequences=True))\n",
    "#                         model.add(Dropout(the_dropout))\n",
    "#                         model.add(BatchNormalization())\n",
    "#                         # #第二层\n",
    "#                         for i in range(the_lstm_layers):\n",
    "#                             model.add(LSTM(the_units,return_sequences=True))\n",
    "#                             model.add(Dropout(the_dropout))\n",
    "#                             model.add(BatchNormalization())\n",
    "\n",
    "#                         model.add(LSTM(the_units))\n",
    "#                         model.add(Dropout(the_dropout))\n",
    "#                         model.add(BatchNormalization())\n",
    "#                         #全连接层\n",
    "#                         for i in range(the_dense_layers):\n",
    "#                             model.add(Dense(the_dense_layers,activation='relu'))\n",
    "#                             model.add(Dropout(the_dropout))\n",
    "                            \n",
    "#                         # model.add(Flatten()) \n",
    "                        \n",
    "                        \n",
    "#                         model.add(Dense(2, activation='softmax'))\n",
    "                        \n",
    "#                         #sgd = SGD(learning_rate=0.01, momentum=0.9 , decay=0.1, nesterov=False)\n",
    "                        \n",
    "#                         # learning_rate = 0.1\n",
    "#                         # decay = 0.001\n",
    "#                         # epochs = 50\n",
    "#                         # batch_size = 64\n",
    "                        \n",
    "                        \n",
    "#                         model.compile(  loss='binary_crossentropy',#categorical_crossentropy', binary_crossentropy\n",
    "#                                         optimizer=the_optimizer, metrics=['categorical_accuracy'])\n",
    "#                         print(model.summary())\n",
    "\n",
    "#                         epochs = 100\n",
    "#                         batch_size = the_batch_size\n",
    "#                         if(the_optimizer == sgd):\n",
    "#                             the_optimizer = 'sgd'\n",
    "#                         filepath = './models/{categorical_accuracy:.4f}_{epoch:02d}_'+f'dropout_{the_dropout}_batch_size_{the_batch_size}_optimizer_{the_optimizer}_dense_layers_{the_dense_layers}_lstm_layers_{the_lstm_layers}_unit_{the_units}.h5'\n",
    "#                         checkpoint = ModelCheckpoint(\n",
    "#                                             filepath=filepath,\n",
    "#                                             save_weights_only=False,\n",
    "#                                             monitor='categorical_accuracy',\n",
    "#                                             mode='max',\n",
    "#                                             save_best_only=True)\n",
    "#                         history = model.fit(X_train, Y_train, epochs=epochs, batch_size=batch_size, validation_split=0.1,\n",
    "#                                             #callbacks=[EarlyStopping(monitor='val_loss', patience=3, min_delta=0.0001)])\n",
    "#                                             callbacks=checkpoint)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {},
   "outputs": [],
   "source": [
    "#model.save('mymodel.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 26,
   "metadata": {},
   "outputs": [],
   "source": [
    "# from tensorflow.keras.models import load_model\n",
    "# best_model = load_model('./models/0.9557_78_dropout_0.05pbatch_size_32optimizer_adamdense_layers_1_lstm_layers_2_unit_128.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 27,
   "metadata": {},
   "outputs": [],
   "source": [
    "# dataset = pd.read_csv('test.csv') \n",
    "# cols = ['Class','Area', 'Perimeter', 'Major_Axis', 'Minor_Axis', 'Eccentricity',\n",
    "#         'Eqdiasq', 'Solidity', 'Convex_Area', 'Extent', 'Aspect_Ratio',\n",
    "#         'Roundness', 'Compactness', 'Shapefactor_1', 'Shapefactor_2',\n",
    "#         'Shapefactor_3', 'Shapefactor_4', 'Mean_RR', 'Mean_RG', 'Mean_RB',\n",
    "#         'StdDev_RR', 'StdDev_RG', 'StdDev_RB', 'Skew_RR', 'Skew_RG', 'Skew_RB',\n",
    "#         'Kurtosis_RR', 'Kurtosis_RG', 'Kurtosis_RB']\n",
    "# # dataset = dataset.loc[:,cols]\n",
    "# dataset.sample(10)\n",
    "# dataset = shuffle(dataset)\n",
    "# Lisclass = dataset['Class'].values\n",
    "# LisClassNp = []\n",
    "# for i in range(len(Lisclass)):\n",
    "#     if (Lisclass[i] == 'Kirmizi_Pistachio'):\n",
    "#         LisClassNp.append([0,1])\n",
    "#     else:\n",
    "#         LisClassNp.append([1,0])\n",
    "# Y = np.array(LisClassNp)\n",
    "# dataset = dataset.drop(columns='Class')\n",
    "# X = dataset.values\n",
    "# # from sklearn.preprocessing import MinMaxScaler\n",
    "# # scaler = MinMaxScaler(feature_range=(0, 1))\n",
    "# # X = scaler.fit_transform(X)\n",
    "# #进行预测 make a prediction\n",
    "# X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.90, random_state=random.randint(10,100))\n",
    "# X_train = X_train.reshape(X_train.shape[0],1,X_train.shape[1])\n",
    "# X_test = X_test.reshape(X_test.shape[0],1,X_test.shape[1])\n",
    "# # print ('train_x.shape, train_y.shape, test_x.shape, test_y.shape')\n",
    "# # print(train_X.shape, train_y.shape, test_X.shape, test_y.shape)\n",
    "# yhat = best_model.predict(X_test)\n",
    "# # print(yhat.shape)\n",
    "# print(yhat)\n",
    "# print(Y_test)\n",
    "# sorce = 0\n",
    "# for i in range(len(Y_test)):\n",
    "#     sorce += roc_auc_score(Y_test[i], yhat[i])\n",
    "\n",
    "# print('ACC:',sorce/len(Y_test))\n",
    "# #拆分训练集和测试集\n",
    "# # X_train, X_test, Y_train, Y_test = train_test_split(X, Y, test_size=0.10, random_state=random.randint(10,100))\n",
    "# # X_train = X_train.reshape(X_train.shape[0],1,X_train.shape[1])\n",
    "# # X_test = X_test.reshape(X_test.shape[0],1,X_test.shape[1])"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 28,
   "metadata": {},
   "outputs": [],
   "source": [
    "# print(Y_test.shape)\n",
    "# print(yhat.shape)\n",
    "# # Y_test = Y_test.reshape(430)\n",
    "# # yhat = yhat.reshape(430)\n",
    "# # Y_test = list(Y_test)\n",
    "# # yhat = list(yhat)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {},
   "outputs": [],
   "source": [
    "# print(yhat)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "\n",
    "\n",
    "\n",
    "\n",
    "\n"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3.8.1 64-bit",
   "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": "3.8.1"
  },
  "orig_nbformat": 4,
  "vscode": {
   "interpreter": {
    "hash": "33d73dfb9d58d1e4c0ed15f266f7c1fd1b5e79268076ebae9c660cb33abbd60e"
   }
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}