EEGnet_CNN_Dense_Binary_Classifier_VERT.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 85,
   "metadata": {},
   "outputs": [],
   "source": [
    "import scipy.io as sio\n",
    "\n",
    "# load data from the .mat file:\n",
    "# mat_contents = sio.loadmat('EEG_big.mat') # (EEG_data.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/EEG_big4CNN.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/EEG_big4CNN_1subj.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/KOS_100Hz_noICA->0.5-38Hz,Env=1.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/Merged123->0.5-38Hz,Env=1.mat')\n",
    "\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/KOS_100Hz_ICA->0.5-38Hz,Env=1.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/Merged123->1-48Hz,Env=1.mat')\n",
    "\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/BCI_TD.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/BCI_FD.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/BCI_0101T_TD.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/BCI_0202T_TD.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/BCI_0403T_TD.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/BCI_0503T_TD.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/BCI_0801T_TD.mat')\n",
    "\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/ks.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/ks_FD.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/ks_2TD.mat')\n",
    "mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/ksenia2_long_chunked.mat') \n",
    "\n",
    "\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/BCI_FD_fullspectrum.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/BCI_FD_lowspectrum.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/KOS_100Hz_ICA->1-38Hz,Env=1_FD.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/Merged123->1-100Hz,Env=1_FD')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/Merged123->1-47Hz,Env=1_FD.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/Merged123->15-47Hz,Env=1_TD.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/Merged123->15-47Hz,Env=1_TD5_64_.mat')\n",
    "# mat_contents = sio.loadmat('/home/amplifier/home/DATASETS/Merged123->2-30Hz,Env=1_TD5_64_.mat')\n",
    "\n",
    "\n",
    "BATCH_SIZE = 20\n",
    "SHUFFLE = True\n",
    "TEST_TRAIN = 0.2\n",
    "\n",
    "KERNEL_SIZE = ((100), (20,1), (2,2))\n",
    "\n",
    "N_KERNELS = (100,200,400)\n",
    "POOL_SIZE = ((5),(5,5),(2,1))\n",
    "\n",
    "DROPOUT = (0.2, 0.3, 0.6, 0.7)\n",
    "DENSE_LEN = 10"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 86,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Original data shape: (134, 24, 500)\n",
      "Original labels shape: (134, 1)\n",
      "Original data type: float64\n",
      "Normalized data type: float16\n",
      "test input shape (27, 24, 500, 1) Nomralized MEAN: -0.0017519 min -65.188 max 164.0\n",
      "train input shape (107, 24, 500, 1) Nomralized MEAN: 0.00096178 min -73.625 max 161.62\n",
      "test labels shape (27, 2) Nomralized MEAN: 0.5 min 0.0 max 1.0\n",
      "train labels shape (107, 2) Nomralized MEAN: 0.5 min 0.0 max 1.0\n"
     ]
    }
   ],
   "source": [
    "# get the Dataset:\n",
    "\n",
    "import numpy as np\n",
    "from sklearn.model_selection import train_test_split\n",
    "from sklearn import preprocessing\n",
    "\n",
    "X = mat_contents['X']\n",
    "Y = mat_contents['Z']\n",
    "    \n",
    "\n",
    "if Y.shape[1] > Y.shape[0]:\n",
    "    Y = Y.T\n",
    "    \n",
    "print('Original data shape:', X.shape)\n",
    "print('Original labels shape:', Y.shape)\n",
    "\n",
    "# verify that the model REALLY finds a mapping between the input and the labels. If we get\n",
    "# our accuracy by chance, then we should get the same accuracy on a permuted dataset:\n",
    "# Y = np.random.permutation(Y)\n",
    "\n",
    "\n",
    "# winsize = mat_contents['winsize']\n",
    "# stepsize = mat_contents['stepsize']\n",
    "# trial_len = mat_contents['trial_len']\n",
    "# low_cutoff = mat_contents['low_cutoff']\n",
    "# high_cutoff = mat_contents['high_cutoff']\n",
    "# source = mat_contents['filein']\n",
    "\n",
    "x_train, x_test, y_train, y_test = train_test_split(X, Y, test_size=TEST_TRAIN, shuffle=SHUFFLE)\n",
    "print('Original data type:', x_train.dtype)\n",
    "\n",
    "# convert to float64 for numerical stability:\n",
    "x_train = x_train.astype('float64')\n",
    "y_train = y_train.astype('float64')\n",
    "x_test = x_test.astype('float64')\n",
    "y_test = y_test.astype('float64')\n",
    "\n",
    "# normalize to unit variance and zero mean:\n",
    "# for i in range(x_train.shape[0]):\n",
    "#     for j in range(x_train.shape[1]):\n",
    "#         x_train[i,j,:] = preprocessing.scale(x_train[i,j,:])\n",
    "# #         x_train[i,j,:] = preprocessing.MaxAbsScaler(x_train[i,j,:])\n",
    "\n",
    "# for i in range(x_test.shape[0]):\n",
    "#     for j in range(x_test.shape[1]):\n",
    "#         x_test[i,j,:] = preprocessing.scale(x_test[i,j,:])\n",
    "# #         x_test[i,j,:] = preprocessing.minmax_scale(x_test[i,j,:])\n",
    "\n",
    "# ensure the tensors are sized right:\n",
    "if len(X.shape)==3:\n",
    "    x_train = np.expand_dims(x_train,3)\n",
    "    x_test = np.expand_dims(x_test,3)\n",
    "    \n",
    "\n",
    "# one hot encode the labels:\n",
    "onehot_encoder = preprocessing.OneHotEncoder(sparse=False)\n",
    "y_train = onehot_encoder.fit_transform(y_train)\n",
    "y_test = onehot_encoder.fit_transform(y_test)\n",
    "\n",
    "# convert to float16 to save space:\n",
    "x_train = x_train.astype('float16')\n",
    "y_train = y_train.astype('float16')\n",
    "x_test = x_test.astype('float16')\n",
    "y_test = y_test.astype('float16')\n",
    "print('Normalized data type:', x_train.dtype)\n",
    "\n",
    "chan = X.shape[1] # if you work in the FD, this is the height of the sample time-frequency image, othewise EEG channels\n",
    "leng = X.shape[2] # if you work in the FD, this is the width of the sample time-frequency image, othewise time samples of EEG signal\n",
    "\n",
    "if len(X.shape)==3:\n",
    "    streams = 1 # this is EEG channels if you work with frequency domain, in the TD streams = 1\n",
    "if len(X.shape)==4:\n",
    "    streams = X.shape[3] # this is EEG channels if you work with frequency domain, in the TD streams = 1\n",
    "\n",
    "\n",
    "print('test input shape', x_test.shape, \"Nomralized MEAN:\", np.mean(x_test), \"min\", np.min(x_test),\"max\", np.max(x_test))\n",
    "print('train input shape', x_train.shape, \"Nomralized MEAN:\", np.mean(x_train), \"min\", np.min(x_train),\"max\", np.max(x_train))\n",
    "\n",
    "print('test labels shape', y_test.shape, \"Nomralized MEAN:\", np.mean(y_test), \"min\", np.min(y_test),\"max\", np.max(y_test))\n",
    "print('train labels shape', y_train.shape, \"Nomralized MEAN:\", np.mean(y_train), \"min\", np.min(y_train),\"max\", np.max(y_train))\n",
    "\n",
    "# print('Window length', winsize)\n",
    "# print('Step size:', stepsize)\n",
    "# print('Trial length:', trial_len)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 87,
   "metadata": {},
   "outputs": [],
   "source": [
    "# we first define the autoencoder's architecture:\n",
    "\n",
    "# make sure you don't hog all the video memory\n",
    "import tensorflow as tf\n",
    "config = tf.ConfigProto()\n",
    "config.gpu_options.allow_growth = True\n",
    "sess = tf.Session(config=config)\n",
    "from keras import backend as K\n",
    "K.set_session(sess)\n",
    "###################################\n",
    "\n",
    "from keras.layers import Input, Dense, Conv2D, Conv1D, ELU, Permute, Lambda, Flatten\n",
    "from keras.layers import MaxPooling2D, MaxPooling2D, MaxPooling1D, Dropout, BatchNormalization\n",
    "from keras.models import Model\n",
    "from keras.models import load_model\n",
    "from keras.callbacks import TensorBoard, EarlyStopping, ModelCheckpoint\n",
    "from keras import optimizers\n",
    "from keras.initializers import Orthogonal as orth\n",
    "\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "\n",
    "def Squeeze(input_img):\n",
    "    return K.squeeze(input_img, axis=3)\n",
    "\n",
    "def Exp_dims(input_img):\n",
    "    return K.expand_dims(input_img, axis=3)\n",
    "\n",
    "def conv_net(leng):\n",
    "    print(leng)\n",
    "    input_img = Input(shape=(chan, leng, streams))  # adapt this if using `channels_first` image\n",
    "    squeezed = Lambda(Squeeze)(input_img)\n",
    "    perm_img = Permute((2,1))(squeezed)\n",
    "    input_norm = BatchNormalization()(perm_img)\n",
    "    e1 = Conv1D(N_KERNELS[0], KERNEL_SIZE[0], padding='same',\n",
    "                                    kernel_initializer = orth(gain=1.0, seed=None))(input_norm)\n",
    "    e1act = ELU(alpha=1.0)(e1)\n",
    "    e1p = MaxPooling1D(POOL_SIZE[0], padding='same')(e1act)\n",
    "    e1e = Lambda(Exp_dims)(e1p)\n",
    "    \n",
    "\n",
    "    e2b = BatchNormalization()(e1e)\n",
    "    e2a = Dropout(DROPOUT[0])(e2b)\n",
    "    e2c = Conv2D(N_KERNELS[1], KERNEL_SIZE[1], padding='same',\n",
    "                                    kernel_initializer = orth(gain=1.0, seed=None))(e2a)\n",
    "    e2act = ELU(alpha=1.0)(e2c)\n",
    "    e2p = MaxPooling2D(POOL_SIZE[1], padding='same')(e2act)\n",
    "    \n",
    "    \n",
    "\n",
    "    e3b = BatchNormalization()(e2p)\n",
    "    e3a = Dropout(DROPOUT[1])(e3b)\n",
    "    e3c = Conv2D(N_KERNELS[2], KERNEL_SIZE[2], padding='same',\n",
    "                                    kernel_initializer = orth(gain=1.0, seed=None))(e3a)\n",
    "    e3act = ELU(alpha=1.0)(e3c)\n",
    "    e3p = MaxPooling2D(POOL_SIZE[2], padding='same')(e3act)\n",
    "      \n",
    "    e10 = Flatten()(e3p)\n",
    "    e10b = BatchNormalization()(e10)\n",
    "    e10d = Dropout(DROPOUT[2])(e10b)\n",
    "    e11 = Dense(DENSE_LEN, activation='sigmoid',\n",
    "                                     kernel_initializer = orth(gain=1.0, seed=None))(e10d)\n",
    "    e12 = Dropout(DROPOUT[3])(e11)\n",
    "    e12b = BatchNormalization()(e12)\n",
    "    e13 = Dense(2, activation='softmax')(e12b)\n",
    "\n",
    "    return Model(input_img, e13)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 88,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "500\n",
      "_________________________________________________________________\n",
      "Layer (type)                 Output Shape              Param #   \n",
      "=================================================================\n",
      "input_26 (InputLayer)        (None, 24, 500, 1)        0         \n",
      "_________________________________________________________________\n",
      "lambda_25 (Lambda)           (None, 24, 500)           0         \n",
      "_________________________________________________________________\n",
      "permute_13 (Permute)         (None, 500, 24)           0         \n",
      "_________________________________________________________________\n",
      "batch_normalization_61 (Batc (None, 500, 24)           96        \n",
      "_________________________________________________________________\n",
      "conv1d_13 (Conv1D)           (None, 500, 100)          240100    \n",
      "_________________________________________________________________\n",
      "elu_25 (ELU)                 (None, 500, 100)          0         \n",
      "_________________________________________________________________\n",
      "max_pooling1d_13 (MaxPooling (None, 100, 100)          0         \n",
      "_________________________________________________________________\n",
      "lambda_26 (Lambda)           (None, 100, 100, 1)       0         \n",
      "_________________________________________________________________\n",
      "batch_normalization_62 (Batc (None, 100, 100, 1)       4         \n",
      "_________________________________________________________________\n",
      "dropout_49 (Dropout)         (None, 100, 100, 1)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_25 (Conv2D)           (None, 100, 100, 200)     4200      \n",
      "_________________________________________________________________\n",
      "elu_26 (ELU)                 (None, 100, 100, 200)     0         \n",
      "_________________________________________________________________\n",
      "max_pooling2d_25 (MaxPooling (None, 20, 20, 200)       0         \n",
      "_________________________________________________________________\n",
      "batch_normalization_63 (Batc (None, 20, 20, 200)       800       \n",
      "_________________________________________________________________\n",
      "dropout_50 (Dropout)         (None, 20, 20, 200)       0         \n",
      "_________________________________________________________________\n",
      "conv2d_26 (Conv2D)           (None, 20, 20, 400)       320400    \n",
      "_________________________________________________________________\n",
      "elu_27 (ELU)                 (None, 20, 20, 400)       0         \n",
      "_________________________________________________________________\n",
      "max_pooling2d_26 (MaxPooling (None, 10, 20, 400)       0         \n",
      "_________________________________________________________________\n",
      "flatten_13 (Flatten)         (None, 80000)             0         \n",
      "_________________________________________________________________\n",
      "batch_normalization_64 (Batc (None, 80000)             320000    \n",
      "_________________________________________________________________\n",
      "dropout_51 (Dropout)         (None, 80000)             0         \n",
      "_________________________________________________________________\n",
      "dense_25 (Dense)             (None, 10)                800010    \n",
      "_________________________________________________________________\n",
      "dropout_52 (Dropout)         (None, 10)                0         \n",
      "_________________________________________________________________\n",
      "batch_normalization_65 (Batc (None, 10)                40        \n",
      "_________________________________________________________________\n",
      "dense_26 (Dense)             (None, 2)                 22        \n",
      "=================================================================\n",
      "Total params: 1,685,672\n",
      "Trainable params: 1,525,202\n",
      "Non-trainable params: 160,470\n",
      "_________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "# define input to the model:\n",
    "x = Input(shape=(chan, leng, streams))\n",
    "\n",
    "# make the model:\n",
    "model = conv_net(leng)\n",
    "\n",
    "# compile the model:\n",
    "sgd = optimizers.SGD(lr=0.001, decay=1e-6, momentum=0.95, nesterov=True)\n",
    "# model.compile(loss='binary_crossentropy', optimizer=sgd, metrics=['accuracy'])\n",
    "# model.compile(optimizer='rmsprop', loss=correlation_coefficient_loss)\n",
    "model.compile(loss='binary_crossentropy', optimizer='adadelta', metrics=['accuracy'])\n",
    "\n",
    "\n",
    "# SEE INSIDE THE MODEL:\n",
    "# conv_net().summary() #or\n",
    "model.summary()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 89,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 107 samples, validate on 27 samples\n",
      "Epoch 1/1000\n",
      "\n",
      "Epoch 00001: val_acc improved from -inf to 0.55556, saving model to /home/amplifier/home/NEW_DL/weights/EEGnet_weights_VERT.h5\n",
      " - 2s - loss: 0.7522 - acc: 0.6168 - val_loss: 0.6750 - val_acc: 0.5556\n",
      "Epoch 2/1000\n",
      "\n",
      "Epoch 00002: val_acc did not improve\n",
      " - 0s - loss: 0.6509 - acc: 0.6542 - val_loss: 0.6334 - val_acc: 0.5556\n",
      "Epoch 3/1000\n",
      "\n",
      "Epoch 00003: val_acc improved from 0.55556 to 0.59259, saving model to /home/amplifier/home/NEW_DL/weights/EEGnet_weights_VERT.h5\n",
      " - 0s - loss: 0.5532 - acc: 0.7009 - val_loss: 0.6720 - val_acc: 0.5926\n",
      "Epoch 4/1000\n",
      "\n",
      "Epoch 00004: val_acc did not improve\n",
      " - 0s - loss: 0.5503 - acc: 0.7009 - val_loss: 0.7279 - val_acc: 0.5926\n",
      "Epoch 5/1000\n",
      "\n",
      "Epoch 00005: val_acc did not improve\n",
      " - 0s - loss: 0.6340 - acc: 0.7009 - val_loss: 0.7154 - val_acc: 0.5926\n",
      "Epoch 6/1000\n",
      "\n",
      "Epoch 00006: val_acc did not improve\n",
      " - 0s - loss: 0.5040 - acc: 0.7570 - val_loss: 0.7485 - val_acc: 0.5926\n",
      "Epoch 7/1000\n",
      "\n",
      "Epoch 00007: val_acc did not improve\n",
      " - 0s - loss: 0.4658 - acc: 0.7757 - val_loss: 0.7306 - val_acc: 0.5556\n",
      "Epoch 8/1000\n",
      "\n",
      "Epoch 00008: val_acc did not improve\n",
      " - 0s - loss: 0.5386 - acc: 0.7477 - val_loss: 0.7507 - val_acc: 0.5926\n",
      "Epoch 9/1000\n",
      "\n",
      "Epoch 00009: val_acc did not improve\n",
      " - 0s - loss: 0.4787 - acc: 0.8318 - val_loss: 0.7079 - val_acc: 0.5926\n",
      "Epoch 10/1000\n",
      "\n",
      "Epoch 00010: val_acc did not improve\n",
      " - 0s - loss: 0.4804 - acc: 0.7664 - val_loss: 0.7231 - val_acc: 0.5926\n",
      "Epoch 11/1000\n",
      "\n",
      "Epoch 00011: val_acc did not improve\n",
      " - 0s - loss: 0.4755 - acc: 0.7757 - val_loss: 0.7116 - val_acc: 0.5926\n",
      "Epoch 12/1000\n",
      "\n",
      "Epoch 00012: val_acc did not improve\n",
      " - 0s - loss: 0.4159 - acc: 0.7944 - val_loss: 0.7275 - val_acc: 0.5926\n",
      "Epoch 13/1000\n",
      "\n",
      "Epoch 00013: val_acc did not improve\n",
      " - 0s - loss: 0.4312 - acc: 0.8224 - val_loss: 0.7336 - val_acc: 0.5926\n",
      "Epoch 14/1000\n",
      "\n",
      "Epoch 00014: val_acc did not improve\n",
      " - 0s - loss: 0.4363 - acc: 0.7944 - val_loss: 0.7994 - val_acc: 0.5926\n",
      "Epoch 15/1000\n",
      "\n",
      "Epoch 00015: val_acc did not improve\n",
      " - 0s - loss: 0.4387 - acc: 0.8411 - val_loss: 0.7941 - val_acc: 0.5926\n",
      "Epoch 16/1000\n",
      "\n",
      "Epoch 00016: val_acc did not improve\n",
      " - 0s - loss: 0.3659 - acc: 0.8785 - val_loss: 0.8069 - val_acc: 0.5926\n",
      "Epoch 17/1000\n",
      "\n",
      "Epoch 00017: val_acc did not improve\n",
      " - 0s - loss: 0.3904 - acc: 0.8692 - val_loss: 0.8115 - val_acc: 0.5926\n",
      "Epoch 18/1000\n",
      "\n",
      "Epoch 00018: val_acc did not improve\n",
      " - 0s - loss: 0.4076 - acc: 0.8505 - val_loss: 0.8085 - val_acc: 0.5926\n",
      "Epoch 19/1000\n",
      "\n",
      "Epoch 00019: val_acc did not improve\n",
      " - 0s - loss: 0.3709 - acc: 0.8692 - val_loss: 0.8134 - val_acc: 0.5926\n",
      "Epoch 20/1000\n",
      "\n",
      "Epoch 00020: val_acc did not improve\n",
      " - 0s - loss: 0.3494 - acc: 0.8598 - val_loss: 0.8006 - val_acc: 0.5556\n",
      "Epoch 21/1000\n",
      "\n",
      "Epoch 00021: val_acc did not improve\n",
      " - 0s - loss: 0.3323 - acc: 0.8598 - val_loss: 0.7546 - val_acc: 0.5926\n",
      "Epoch 22/1000\n",
      "\n",
      "Epoch 00022: val_acc did not improve\n",
      " - 0s - loss: 0.4013 - acc: 0.8598 - val_loss: 0.8158 - val_acc: 0.5556\n",
      "Epoch 23/1000\n",
      "\n",
      "Epoch 00023: val_acc did not improve\n",
      " - 0s - loss: 0.3990 - acc: 0.8411 - val_loss: 0.8096 - val_acc: 0.5556\n",
      "Epoch 24/1000\n",
      "\n",
      "Epoch 00024: val_acc improved from 0.59259 to 0.66667, saving model to /home/amplifier/home/NEW_DL/weights/EEGnet_weights_VERT.h5\n",
      " - 0s - loss: 0.4282 - acc: 0.8224 - val_loss: 0.6768 - val_acc: 0.6667\n",
      "Epoch 25/1000\n",
      "\n",
      "Epoch 00025: val_acc did not improve\n",
      " - 0s - loss: 0.3395 - acc: 0.8972 - val_loss: 0.7380 - val_acc: 0.6296\n",
      "Epoch 26/1000\n",
      "\n",
      "Epoch 00026: val_acc did not improve\n",
      " - 0s - loss: 0.3960 - acc: 0.8131 - val_loss: 0.6679 - val_acc: 0.6296\n",
      "Epoch 27/1000\n",
      "\n",
      "Epoch 00027: val_acc did not improve\n",
      " - 0s - loss: 0.3281 - acc: 0.8785 - val_loss: 0.6297 - val_acc: 0.6296\n",
      "Epoch 28/1000\n",
      "\n",
      "Epoch 00028: val_acc improved from 0.66667 to 0.74074, saving model to /home/amplifier/home/NEW_DL/weights/EEGnet_weights_VERT.h5\n",
      " - 0s - loss: 0.3600 - acc: 0.8785 - val_loss: 0.5967 - val_acc: 0.7407\n",
      "Epoch 29/1000\n",
      "\n",
      "Epoch 00029: val_acc did not improve\n",
      " - 0s - loss: 0.2825 - acc: 0.9252 - val_loss: 0.6583 - val_acc: 0.6667\n",
      "Epoch 30/1000\n",
      "\n",
      "Epoch 00030: val_acc did not improve\n",
      " - 0s - loss: 0.2869 - acc: 0.9065 - val_loss: 0.7140 - val_acc: 0.6296\n",
      "Epoch 31/1000\n",
      "\n",
      "Epoch 00031: val_acc did not improve\n",
      " - 0s - loss: 0.2930 - acc: 0.8972 - val_loss: 0.7539 - val_acc: 0.5926\n",
      "Epoch 32/1000\n",
      "\n",
      "Epoch 00032: val_acc did not improve\n",
      " - 0s - loss: 0.2908 - acc: 0.9159 - val_loss: 0.8110 - val_acc: 0.5926\n",
      "Epoch 33/1000\n",
      "\n",
      "Epoch 00033: val_acc did not improve\n",
      " - 0s - loss: 0.2905 - acc: 0.9065 - val_loss: 0.8303 - val_acc: 0.6296\n",
      "Epoch 34/1000\n",
      "\n",
      "Epoch 00034: val_acc did not improve\n",
      " - 0s - loss: 0.3021 - acc: 0.8879 - val_loss: 0.7628 - val_acc: 0.6667\n",
      "Epoch 35/1000\n",
      "\n",
      "Epoch 00035: val_acc did not improve\n",
      " - 0s - loss: 0.3139 - acc: 0.8598 - val_loss: 0.7630 - val_acc: 0.6667\n",
      "Epoch 36/1000\n",
      "\n",
      "Epoch 00036: val_acc did not improve\n",
      " - 0s - loss: 0.2821 - acc: 0.9159 - val_loss: 0.7688 - val_acc: 0.6296\n",
      "Epoch 37/1000\n",
      "\n",
      "Epoch 00037: val_acc did not improve\n",
      " - 0s - loss: 0.3039 - acc: 0.8972 - val_loss: 0.7908 - val_acc: 0.6296\n",
      "Epoch 38/1000\n",
      "\n",
      "Epoch 00038: val_acc did not improve\n",
      " - 0s - loss: 0.2987 - acc: 0.8598 - val_loss: 0.7178 - val_acc: 0.6667\n",
      "Epoch 39/1000\n",
      "\n",
      "Epoch 00039: val_acc did not improve\n",
      " - 0s - loss: 0.2329 - acc: 0.9533 - val_loss: 0.7774 - val_acc: 0.6296\n",
      "Epoch 40/1000\n",
      "\n",
      "Epoch 00040: val_acc did not improve\n",
      " - 0s - loss: 0.2701 - acc: 0.9159 - val_loss: 0.7809 - val_acc: 0.6296\n",
      "Epoch 41/1000\n",
      "\n",
      "Epoch 00041: val_acc did not improve\n",
      " - 0s - loss: 0.2049 - acc: 0.9626 - val_loss: 0.8158 - val_acc: 0.6296\n",
      "Epoch 42/1000\n",
      "\n",
      "Epoch 00042: val_acc did not improve\n",
      " - 0s - loss: 0.2235 - acc: 0.9439 - val_loss: 0.8353 - val_acc: 0.6296\n",
      "Epoch 43/1000\n",
      "\n",
      "Epoch 00043: val_acc did not improve\n",
      " - 0s - loss: 0.2040 - acc: 0.9626 - val_loss: 0.8324 - val_acc: 0.6296\n",
      "Epoch 44/1000\n",
      "\n",
      "Epoch 00044: val_acc did not improve\n",
      " - 0s - loss: 0.2470 - acc: 0.9159 - val_loss: 0.7797 - val_acc: 0.6296\n",
      "Epoch 45/1000\n",
      "\n",
      "Epoch 00045: val_acc did not improve\n",
      " - 0s - loss: 0.2567 - acc: 0.9346 - val_loss: 0.8149 - val_acc: 0.6296\n",
      "Epoch 46/1000\n",
      "\n",
      "Epoch 00046: val_acc did not improve\n",
      " - 0s - loss: 0.3005 - acc: 0.9159 - val_loss: 0.8275 - val_acc: 0.6296\n",
      "Epoch 47/1000\n",
      "\n",
      "Epoch 00047: val_acc did not improve\n",
      " - 0s - loss: 0.2379 - acc: 0.9346 - val_loss: 0.8534 - val_acc: 0.5926\n",
      "Epoch 48/1000\n",
      "\n",
      "Epoch 00048: val_acc did not improve\n",
      " - 0s - loss: 0.3206 - acc: 0.9065 - val_loss: 0.8781 - val_acc: 0.5926\n",
      "Epoch 49/1000\n",
      "\n",
      "Epoch 00049: val_acc did not improve\n",
      " - 0s - loss: 0.2517 - acc: 0.9159 - val_loss: 0.9005 - val_acc: 0.5926\n",
      "Epoch 50/1000\n",
      "\n",
      "Epoch 00050: val_acc did not improve\n",
      " - 0s - loss: 0.2084 - acc: 0.9533 - val_loss: 0.9353 - val_acc: 0.5926\n",
      "Epoch 51/1000\n",
      "\n",
      "Epoch 00051: val_acc did not improve\n",
      " - 0s - loss: 0.2328 - acc: 0.9346 - val_loss: 0.9130 - val_acc: 0.5926\n",
      "Epoch 52/1000\n",
      "\n",
      "Epoch 00052: val_acc did not improve\n",
      " - 0s - loss: 0.2270 - acc: 0.9159 - val_loss: 0.9183 - val_acc: 0.5926\n",
      "Epoch 53/1000\n",
      "\n",
      "Epoch 00053: val_acc did not improve\n",
      " - 0s - loss: 0.2311 - acc: 0.9346 - val_loss: 0.9341 - val_acc: 0.5556\n",
      "Epoch 54/1000\n",
      "\n",
      "Epoch 00054: val_acc did not improve\n",
      " - 0s - loss: 0.2267 - acc: 0.9439 - val_loss: 0.9470 - val_acc: 0.5556\n",
      "Epoch 55/1000\n",
      "\n",
      "Epoch 00055: val_acc did not improve\n",
      " - 0s - loss: 0.2169 - acc: 0.9252 - val_loss: 0.9503 - val_acc: 0.5926\n",
      "Epoch 56/1000\n",
      "\n",
      "Epoch 00056: val_acc did not improve\n",
      " - 0s - loss: 0.2161 - acc: 0.9346 - val_loss: 0.9947 - val_acc: 0.5556\n",
      "Epoch 57/1000\n",
      "\n",
      "Epoch 00057: val_acc did not improve\n",
      " - 0s - loss: 0.2642 - acc: 0.8972 - val_loss: 1.0083 - val_acc: 0.5926\n",
      "Epoch 58/1000\n",
      "\n",
      "Epoch 00058: val_acc did not improve\n",
      " - 0s - loss: 0.2178 - acc: 0.9346 - val_loss: 1.0022 - val_acc: 0.5556\n",
      "Epoch 59/1000\n",
      "\n",
      "Epoch 00059: val_acc did not improve\n",
      " - 0s - loss: 0.2372 - acc: 0.9533 - val_loss: 1.0071 - val_acc: 0.5556\n",
      "Epoch 60/1000\n",
      "\n",
      "Epoch 00060: val_acc did not improve\n",
      " - 0s - loss: 0.1979 - acc: 0.9533 - val_loss: 1.0625 - val_acc: 0.5556\n",
      "Epoch 61/1000\n",
      "\n",
      "Epoch 00061: val_acc did not improve\n",
      " - 0s - loss: 0.2002 - acc: 0.9439 - val_loss: 1.0818 - val_acc: 0.5556\n",
      "Epoch 62/1000\n",
      "\n",
      "Epoch 00062: val_acc did not improve\n",
      " - 0s - loss: 0.1949 - acc: 0.9533 - val_loss: 1.1350 - val_acc: 0.5556\n",
      "Epoch 63/1000\n",
      "\n",
      "Epoch 00063: val_acc did not improve\n",
      " - 0s - loss: 0.2385 - acc: 0.8879 - val_loss: 1.0936 - val_acc: 0.5926\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 64/1000\n",
      "\n",
      "Epoch 00064: val_acc did not improve\n",
      " - 0s - loss: 0.2650 - acc: 0.9065 - val_loss: 1.0651 - val_acc: 0.5926\n",
      "Epoch 65/1000\n",
      "\n",
      "Epoch 00065: val_acc did not improve\n",
      " - 0s - loss: 0.2450 - acc: 0.8785 - val_loss: 1.0583 - val_acc: 0.5926\n",
      "Epoch 66/1000\n",
      "\n",
      "Epoch 00066: val_acc did not improve\n",
      " - 0s - loss: 0.2186 - acc: 0.9626 - val_loss: 1.0887 - val_acc: 0.5926\n",
      "Epoch 67/1000\n",
      "\n",
      "Epoch 00067: val_acc did not improve\n",
      " - 0s - loss: 0.2296 - acc: 0.9439 - val_loss: 1.0742 - val_acc: 0.5926\n",
      "Epoch 68/1000\n",
      "\n",
      "Epoch 00068: val_acc did not improve\n",
      " - 0s - loss: 0.2234 - acc: 0.9346 - val_loss: 1.1084 - val_acc: 0.5926\n",
      "Epoch 69/1000\n",
      "\n",
      "Epoch 00069: val_acc did not improve\n",
      " - 0s - loss: 0.1712 - acc: 0.9720 - val_loss: 1.1046 - val_acc: 0.5926\n",
      "Epoch 70/1000\n",
      "\n",
      "Epoch 00070: val_acc did not improve\n",
      " - 0s - loss: 0.1692 - acc: 0.9439 - val_loss: 1.1357 - val_acc: 0.5556\n",
      "Epoch 71/1000\n",
      "\n",
      "Epoch 00071: val_acc did not improve\n",
      " - 0s - loss: 0.1909 - acc: 0.9533 - val_loss: 1.1164 - val_acc: 0.5556\n",
      "Epoch 72/1000\n",
      "\n",
      "Epoch 00072: val_acc did not improve\n",
      " - 0s - loss: 0.1788 - acc: 0.9533 - val_loss: 1.1156 - val_acc: 0.5926\n",
      "Epoch 73/1000\n",
      "\n",
      "Epoch 00073: val_acc did not improve\n",
      " - 0s - loss: 0.2396 - acc: 0.9065 - val_loss: 1.1767 - val_acc: 0.5926\n",
      "Epoch 74/1000\n",
      "\n",
      "Epoch 00074: val_acc did not improve\n",
      " - 0s - loss: 0.2691 - acc: 0.8879 - val_loss: 1.1918 - val_acc: 0.5556\n",
      "Epoch 75/1000\n",
      "\n",
      "Epoch 00075: val_acc did not improve\n",
      " - 0s - loss: 0.1887 - acc: 0.9533 - val_loss: 1.1874 - val_acc: 0.5556\n",
      "Epoch 76/1000\n",
      "\n",
      "Epoch 00076: val_acc did not improve\n",
      " - 0s - loss: 0.1934 - acc: 0.9533 - val_loss: 1.1890 - val_acc: 0.5556\n",
      "Epoch 77/1000\n",
      "\n",
      "Epoch 00077: val_acc did not improve\n",
      " - 0s - loss: 0.1969 - acc: 0.9346 - val_loss: 1.2070 - val_acc: 0.5556\n",
      "Epoch 78/1000\n",
      "\n",
      "Epoch 00078: val_acc did not improve\n",
      " - 0s - loss: 0.1782 - acc: 0.9533 - val_loss: 1.2210 - val_acc: 0.5556\n",
      "Epoch 79/1000\n",
      "\n",
      "Epoch 00079: val_acc did not improve\n",
      " - 0s - loss: 0.2213 - acc: 0.9159 - val_loss: 1.2223 - val_acc: 0.5926\n",
      "Epoch 80/1000\n",
      "\n",
      "Epoch 00080: val_acc did not improve\n",
      " - 0s - loss: 0.1961 - acc: 0.9346 - val_loss: 1.2562 - val_acc: 0.5556\n",
      "Epoch 81/1000\n",
      "\n",
      "Epoch 00081: val_acc did not improve\n",
      " - 0s - loss: 0.1829 - acc: 0.9626 - val_loss: 1.1691 - val_acc: 0.5926\n",
      "Epoch 82/1000\n",
      "\n",
      "Epoch 00082: val_acc did not improve\n",
      " - 0s - loss: 0.1834 - acc: 0.9252 - val_loss: 1.2386 - val_acc: 0.5556\n",
      "Epoch 83/1000\n",
      "\n",
      "Epoch 00083: val_acc did not improve\n",
      " - 0s - loss: 0.2875 - acc: 0.9065 - val_loss: 1.2264 - val_acc: 0.5926\n",
      "Epoch 84/1000\n",
      "\n",
      "Epoch 00084: val_acc did not improve\n",
      " - 0s - loss: 0.2040 - acc: 0.9252 - val_loss: 1.1965 - val_acc: 0.5185\n",
      "Epoch 85/1000\n",
      "\n",
      "Epoch 00085: val_acc did not improve\n",
      " - 0s - loss: 0.1555 - acc: 0.9720 - val_loss: 1.2295 - val_acc: 0.5556\n",
      "Epoch 86/1000\n",
      "\n",
      "Epoch 00086: val_acc did not improve\n",
      " - 0s - loss: 0.1649 - acc: 0.9626 - val_loss: 1.2242 - val_acc: 0.5556\n",
      "Epoch 87/1000\n",
      "\n",
      "Epoch 00087: val_acc did not improve\n",
      " - 0s - loss: 0.1904 - acc: 0.9533 - val_loss: 1.2493 - val_acc: 0.5926\n",
      "Epoch 88/1000\n",
      "\n",
      "Epoch 00088: val_acc did not improve\n",
      " - 0s - loss: 0.2206 - acc: 0.8972 - val_loss: 1.2198 - val_acc: 0.5926\n",
      "Epoch 89/1000\n",
      "\n",
      "Epoch 00089: val_acc did not improve\n",
      " - 0s - loss: 0.1414 - acc: 0.9813 - val_loss: 1.2318 - val_acc: 0.5926\n",
      "Epoch 90/1000\n",
      "\n",
      "Epoch 00090: val_acc did not improve\n",
      " - 0s - loss: 0.3392 - acc: 0.8785 - val_loss: 1.1861 - val_acc: 0.5926\n",
      "Epoch 91/1000\n",
      "\n",
      "Epoch 00091: val_acc did not improve\n",
      " - 0s - loss: 0.1634 - acc: 0.9439 - val_loss: 1.2062 - val_acc: 0.5926\n",
      "Epoch 92/1000\n",
      "\n",
      "Epoch 00092: val_acc did not improve\n",
      " - 0s - loss: 0.1585 - acc: 0.9720 - val_loss: 1.2070 - val_acc: 0.5926\n",
      "Epoch 93/1000\n",
      "\n",
      "Epoch 00093: val_acc did not improve\n",
      " - 0s - loss: 0.2180 - acc: 0.9065 - val_loss: 1.2352 - val_acc: 0.5926\n",
      "Epoch 94/1000\n",
      "\n",
      "Epoch 00094: val_acc did not improve\n",
      " - 0s - loss: 0.2034 - acc: 0.9346 - val_loss: 1.2407 - val_acc: 0.5926\n",
      "Epoch 95/1000\n",
      "\n",
      "Epoch 00095: val_acc did not improve\n",
      " - 0s - loss: 0.2320 - acc: 0.9346 - val_loss: 1.2221 - val_acc: 0.5556\n",
      "Epoch 96/1000\n",
      "\n",
      "Epoch 00096: val_acc did not improve\n",
      " - 0s - loss: 0.2408 - acc: 0.8598 - val_loss: 1.2270 - val_acc: 0.5556\n",
      "Epoch 97/1000\n",
      "\n",
      "Epoch 00097: val_acc did not improve\n",
      " - 0s - loss: 0.2484 - acc: 0.9065 - val_loss: 1.2005 - val_acc: 0.5556\n",
      "Epoch 98/1000\n",
      "\n",
      "Epoch 00098: val_acc did not improve\n",
      " - 0s - loss: 0.1731 - acc: 0.9626 - val_loss: 1.2134 - val_acc: 0.5556\n",
      "Epoch 99/1000\n",
      "\n",
      "Epoch 00099: val_acc did not improve\n",
      " - 0s - loss: 0.2787 - acc: 0.9159 - val_loss: 1.3166 - val_acc: 0.5185\n",
      "Epoch 100/1000\n",
      "\n",
      "Epoch 00100: val_acc did not improve\n",
      " - 0s - loss: 0.3184 - acc: 0.8879 - val_loss: 1.3663 - val_acc: 0.5185\n",
      "Epoch 101/1000\n",
      "\n",
      "Epoch 00101: val_acc did not improve\n",
      " - 0s - loss: 0.2109 - acc: 0.9533 - val_loss: 1.3496 - val_acc: 0.5185\n",
      "Epoch 102/1000\n",
      "\n",
      "Epoch 00102: val_acc did not improve\n",
      " - 0s - loss: 0.1990 - acc: 0.9159 - val_loss: 1.3686 - val_acc: 0.5185\n",
      "Epoch 103/1000\n",
      "\n",
      "Epoch 00103: val_acc did not improve\n",
      " - 0s - loss: 0.3999 - acc: 0.8598 - val_loss: 1.2638 - val_acc: 0.5185\n",
      "Epoch 104/1000\n",
      "\n",
      "Epoch 00104: val_acc did not improve\n",
      " - 0s - loss: 0.1470 - acc: 0.9626 - val_loss: 1.3256 - val_acc: 0.5185\n",
      "Epoch 105/1000\n",
      "\n",
      "Epoch 00105: val_acc did not improve\n",
      " - 0s - loss: 0.2518 - acc: 0.9252 - val_loss: 1.3432 - val_acc: 0.5185\n",
      "Epoch 106/1000\n",
      "\n",
      "Epoch 00106: val_acc did not improve\n",
      " - 0s - loss: 0.2224 - acc: 0.9159 - val_loss: 1.2945 - val_acc: 0.5185\n",
      "Epoch 107/1000\n",
      "\n",
      "Epoch 00107: val_acc did not improve\n",
      " - 0s - loss: 0.1668 - acc: 0.9439 - val_loss: 1.3109 - val_acc: 0.5185\n",
      "Epoch 108/1000\n",
      "\n",
      "Epoch 00108: val_acc did not improve\n",
      " - 0s - loss: 0.2018 - acc: 0.9533 - val_loss: 1.1395 - val_acc: 0.6296\n",
      "Epoch 109/1000\n",
      "\n",
      "Epoch 00109: val_acc did not improve\n",
      " - 0s - loss: 0.1591 - acc: 0.9626 - val_loss: 1.1653 - val_acc: 0.6296\n",
      "Epoch 110/1000\n",
      "\n",
      "Epoch 00110: val_acc did not improve\n",
      " - 0s - loss: 0.2111 - acc: 0.9252 - val_loss: 1.1765 - val_acc: 0.6296\n",
      "Epoch 111/1000\n",
      "\n",
      "Epoch 00111: val_acc did not improve\n",
      " - 0s - loss: 0.1798 - acc: 0.9533 - val_loss: 1.1614 - val_acc: 0.6296\n",
      "Epoch 112/1000\n",
      "\n",
      "Epoch 00112: val_acc did not improve\n",
      " - 0s - loss: 0.1501 - acc: 0.9720 - val_loss: 1.1878 - val_acc: 0.6296\n",
      "Epoch 113/1000\n",
      "\n",
      "Epoch 00113: val_acc did not improve\n",
      " - 0s - loss: 0.3438 - acc: 0.8505 - val_loss: 1.1929 - val_acc: 0.5926\n",
      "Epoch 114/1000\n",
      "\n",
      "Epoch 00114: val_acc did not improve\n",
      " - 0s - loss: 0.1882 - acc: 0.9626 - val_loss: 1.2740 - val_acc: 0.5926\n",
      "Epoch 115/1000\n",
      "\n",
      "Epoch 00115: val_acc did not improve\n",
      " - 0s - loss: 0.1719 - acc: 0.9626 - val_loss: 1.2758 - val_acc: 0.5926\n",
      "Epoch 116/1000\n",
      "\n",
      "Epoch 00116: val_acc did not improve\n",
      " - 0s - loss: 0.1504 - acc: 0.9533 - val_loss: 1.3150 - val_acc: 0.5926\n",
      "Epoch 117/1000\n",
      "\n",
      "Epoch 00117: val_acc did not improve\n",
      " - 0s - loss: 0.1887 - acc: 0.9065 - val_loss: 1.2892 - val_acc: 0.5926\n",
      "Epoch 118/1000\n",
      "\n",
      "Epoch 00118: val_acc did not improve\n",
      " - 0s - loss: 0.1982 - acc: 0.9159 - val_loss: 1.2725 - val_acc: 0.5926\n",
      "Epoch 119/1000\n",
      "\n",
      "Epoch 00119: val_acc did not improve\n",
      " - 0s - loss: 0.2351 - acc: 0.9346 - val_loss: 1.3023 - val_acc: 0.5556\n",
      "Epoch 120/1000\n",
      "\n",
      "Epoch 00120: val_acc did not improve\n",
      " - 0s - loss: 0.1818 - acc: 0.9346 - val_loss: 1.2869 - val_acc: 0.5926\n",
      "Epoch 121/1000\n",
      "\n",
      "Epoch 00121: val_acc did not improve\n",
      " - 0s - loss: 0.1852 - acc: 0.9720 - val_loss: 1.3881 - val_acc: 0.5556\n",
      "Epoch 122/1000\n",
      "\n",
      "Epoch 00122: val_acc did not improve\n",
      " - 0s - loss: 0.1526 - acc: 0.9720 - val_loss: 1.3757 - val_acc: 0.5926\n",
      "Epoch 123/1000\n",
      "\n",
      "Epoch 00123: val_acc did not improve\n",
      " - 0s - loss: 0.2754 - acc: 0.8598 - val_loss: 1.3759 - val_acc: 0.5185\n",
      "Epoch 124/1000\n",
      "\n",
      "Epoch 00124: val_acc did not improve\n",
      " - 0s - loss: 0.1496 - acc: 0.9533 - val_loss: 1.4067 - val_acc: 0.5556\n",
      "Epoch 125/1000\n",
      "\n",
      "Epoch 00125: val_acc did not improve\n",
      " - 0s - loss: 0.2397 - acc: 0.9626 - val_loss: 1.3790 - val_acc: 0.5556\n",
      "Epoch 126/1000\n",
      "\n",
      "Epoch 00126: val_acc did not improve\n",
      " - 0s - loss: 0.3096 - acc: 0.8879 - val_loss: 1.3719 - val_acc: 0.5556\n",
      "Epoch 127/1000\n",
      "\n",
      "Epoch 00127: val_acc did not improve\n",
      " - 0s - loss: 0.2075 - acc: 0.9346 - val_loss: 1.3519 - val_acc: 0.5926\n",
      "Epoch 128/1000\n",
      "\n",
      "Epoch 00128: val_acc did not improve\n",
      " - 0s - loss: 0.1623 - acc: 0.9533 - val_loss: 1.3676 - val_acc: 0.5926\n"
     ]
    }
   ],
   "source": [
    "# Training time!\n",
    "early_stopping = EarlyStopping(monitor='val_loss', patience=100, mode='min')\n",
    "checkpointer = ModelCheckpoint(filepath='/home/amplifier/home/NEW_DL/weights/EEGnet_weights_VERT.h5',\n",
    "                               verbose=1,\n",
    "                               monitor='val_acc',\n",
    "                               save_best_only=True)\n",
    "TB = TensorBoard(log_dir='/home/amplifier/home/CNN_stim')\n",
    "\n",
    "train_history = model.fit(x_train, y_train,\n",
    "                epochs=1000,\n",
    "                batch_size=BATCH_SIZE,\n",
    "                verbose=2,\n",
    "                shuffle=True,\n",
    "                validation_data=(x_test, y_test),\n",
    "                callbacks=[TB, checkpointer, early_stopping])\n",
    "\n",
    "# model.save('/home/amplifier/home/NEW_DL/models/EEGnet.h5')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 90,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(107, 2)\n",
      "(27, 2)\n"
     ]
    }
   ],
   "source": [
    "# load the best model:\n",
    "model.load_weights('/home/amplifier/home/NEW_DL/weights/EEGnet_weights_VERT.h5')\n",
    "\n",
    "# let's predict a TRAINING sammple:\n",
    "decoded_train = model.predict(x_train)\n",
    "decoded_test = model.predict(x_test)\n",
    "print(decoded_train.shape)\n",
    "print(decoded_test.shape)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 91,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[[13  2]\n",
      " [ 5  7]]\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f9875760048>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "W = []\n",
    "GT = []\n",
    "\n",
    "for i in range(len(decoded_test)):\n",
    "    if decoded_test[i][0]<decoded_test[i][1]:\n",
    "        W.append('right')\n",
    "    if decoded_test[i][0]>decoded_test[i][1]:\n",
    "        W.append('left')\n",
    "\n",
    "    if y_test[i][0]<y_test[i][1]:\n",
    "        GT.append('right')\n",
    "    if y_test[i][0]>y_test[i][1]:\n",
    "        GT.append('left')\n",
    "        \n",
    "from sklearn.metrics import confusion_matrix\n",
    "labels = ['left', 'right'] # 1 for right, 0 for left\n",
    "cm = confusion_matrix(GT, W, labels)\n",
    "fig = plt.figure()\n",
    "ax = fig.add_subplot(111)\n",
    "\n",
    "ax.matshow(cm)\n",
    "plt.xlabel('Predicted')\n",
    "plt.ylabel('True')\n",
    "plt.title('Confusion Matrix')\n",
    "ax.set_xticklabels([''] + labels)\n",
    "ax.set_yticklabels([''] + labels)\n",
    "print(cm)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 92,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f986ec25d30>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "#####################################################\n",
    "\n",
    "# print(np.array(history.losses))\n",
    "loss = train_history.history['loss']\n",
    "val_loss = train_history.history['val_loss']\n",
    "plt.plot(loss)\n",
    "plt.plot(val_loss)\n",
    "plt.legend(['loss', 'val_loss'])\n",
    "plt.title('AdaDelta')\n",
    "plt.ylabel('loss')\n",
    "plt.xlabel('epoch')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 93,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Accuracy on the test data: 0.740740740741\n",
      "Accuracy on the training data: 0.943925233645\n"
     ]
    }
   ],
   "source": [
    "##########################################################\n",
    "\n",
    "# now test sample by sample\n",
    "cor_test = []\n",
    "L = x_test.shape\n",
    "for sample_no in range(x_test.shape[0]):\n",
    "    pred = model.predict(np.expand_dims(x_test[sample_no],0))\n",
    "    if ((pred[0][0]<pred[0][1]) and (y_test[sample_no,0] < y_test[sample_no,1]) or\n",
    "       ((pred[0][0]>pred[0][1]) & (y_test[sample_no,0] > y_test[sample_no,1]))):\n",
    "        cor_test.append(1)\n",
    "    else:\n",
    "        cor_test.append(0)\n",
    "\n",
    "cor_train = []\n",
    "L = x_train.shape\n",
    "for sample_no in range(x_train.shape[0]):\n",
    "    pred = model.predict(np.expand_dims(x_train[sample_no],0))\n",
    "    if ((pred[0][0]<pred[0][1]) and (y_train[sample_no,0] < y_train[sample_no,1]) or\n",
    "       ((pred[0][0]>pred[0][1]) & (y_train[sample_no,0] > y_train[sample_no,1]))):\n",
    "        cor_train.append(1)\n",
    "    else:\n",
    "        cor_train.append(0)\n",
    "\n",
    "# Report accuracies\n",
    "print('Accuracy on the test data:', np.mean(cor_test))\n",
    "print('Accuracy on the training data:', np.mean(cor_train))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Visualize a few kernels from the first convolutional layer:\n",
    "# enc.layers[2].get_config()\n",
    "# dir(enc.layers[2])\n",
    "x = model.layers[2].get_weights()\n",
    "print ('Layer number:', 2, 'Layer name:', model.layers[2].name)\n",
    "sh = x[0].shape\n",
    "print(sh) # kernel length x input channels x output channels\n",
    "\n",
    "c = 0\n",
    "fig, axs = plt.subplots(sh[2],sh[3])\n",
    "fig.set_size_inches(18.5, 10.5)\n",
    "axs = axs.ravel()\n",
    "for i in range(sh[2]):\n",
    "    for j in range(sh[3]):\n",
    "        axs[c].imshow(x[0][:,:,i,j])\n",
    "        axs[c].get_xaxis().set_visible(False)\n",
    "        axs[c].get_yaxis().set_visible(False)        \n",
    "        axs[c].set_title(str(i)+str(j))\n",
    "        c = c + 1\n",
    "plt.suptitle(('Convolution kernels in layer ' + model.layers[2].name), fontsize=16)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "from IPython.display import SVG\n",
    "from keras.utils.vis_utils import model_to_dot\n",
    "\n",
    "SVG(model_to_dot(model, show_shapes=True).create(prog='dot', format='svg'))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}