tutorial/tutorial_transfer_learning.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<img src=\"figure/clairvoyance_logo.png\">\n",
    "\n",
    "# Clairvoyance: Transfer learning on time-series prediction\n",
    "\n",
    "## ML-AIM (http://vanderschaar-lab.com/)\n",
    "\n",
    "This notebook describes the user-guide of a transfer learning on time-series predictions application using Clairvoyance framework. The separate tutorial for the time-series prediction can be found in tutorial_prediction.ipynb. Transfer learning is defined as following: transfering the knowledge of the trained model (trained on different dataset) to improve the predictive model to the interested dataset. For instance, transfer the knowledge of the trained time-series prediction model on Hospital A to improve the predictive model for Hospital B.\n",
    " \n",
    "<img src=\"figure/transfer-learning-definition.png\">\n",
    "\n",
    "To run this tutorial, you need:\n",
    "### Temporal and static datasets for training and testing\n",
    "\n",
    "If users come with their own temporal and static datasets for training and testing, the users should save those files as 'data_name_temporal_train_data_eav.csv.gz', 'data_name_static_train_data.csv.gz', 'data_name_temporal_test_data_eav.csv.gz', 'data_name_static_test_data.csv.gz' in '../datasets/data/data_name/' directory.\n",
    "\n",
    "\n",
    "### Prerequisite\n",
    "Clone https://github.com/jsyoon0823/time-series-automl.git to the current directory."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Transfer learning on time-series prediction pipeline summary\n",
    "\n",
    "<img src=\"figure/transfer-learning-block-diagram.png\">\n",
    "\n",
    "### Step 1: Load dataset\n",
    "  - Extract csv files from the original raw datasets in ../datasets/data/data_name/ directory.  \n",
    "  \n",
    "### Step 2: Preprocess dataset\n",
    "  - Preprocessing the raw data using various filters such as (1) replacing negative values to NaN, (2) do one-hot encidng for certain features, (3) do normalization.  \n",
    "  \n",
    "### Step 3: Define problem\n",
    "  - Set the time-series prediction problem that we want to solve. Set the problem (whether it is one-shot or online prediction), set the label, set the maximum sequence length, and set the treatment features. We also define the metric for evaluation and the task itself (whether classification or regression).\n",
    "\n",
    "### Step 4: Impute dataset\n",
    "  - Impute missing values in the preprocessed static and temporal datasets and return complete datasets.\n",
    "  \n",
    "### Step 5: Feature selection\n",
    "  - Select the relevant static and temporal features to the labels. You can skip the feature selection (set feature selection method = None).\n",
    "  \n",
    "### Step 6: Original time-series model fit and predict\n",
    "  - After finishing the data preparation, we define the original predictive models and train the model using the training dataset. After training, we use the trained model to predict the labels of the testing dataset.\n",
    "  \n",
    "### Step 7: Transfer learning time-series model fit and predict\n",
    "  - After finishing the data preparation and original predictive model training, we define the transfer learning model and train the model on top of trained original predictive model using the transfer dataset. After training, we use the trained transfer model to predict the labels of the testing dataset.\n",
    "\n",
    "### Step 8: Visualize results\n",
    "  - Visualize the various results such as performance and predictions."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 0: Import necessary packages\n",
    "\n",
    "Import necessary packages for the entire tutorials."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Necessary packages\n",
    "from __future__ import absolute_import\n",
    "from __future__ import division\n",
    "from __future__ import print_function\n",
    "\n",
    "import numpy as np\n",
    "import warnings\n",
    "warnings.filterwarnings('ignore')\n",
    "import sys\n",
    "sys.path.append('../')\n",
    "\n",
    "from utils import PipelineComposer"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 1: Load dataset\n",
    "\n",
    "Extract temporal and static datasets from 'data_name_temporal_train_data_eav.csv.gz', 'data_name_static_train_data.csv.gz', 'data_name_temporal_test_data_eav.csv.gz', 'data_name_static_test_data.csv.gz' in '../datasets/data/data_name/' directory.\n",
    "\n",
    "- CSVLoader: Load csv files from the original raw datasets in ../datasets/data/data_name/ directory.\n",
    "- file_names: mimic in this tutorial."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Finish data loading.\n"
     ]
    }
   ],
   "source": [
    "from datasets import CSVLoader\n",
    "\n",
    "# Define data name\n",
    "data_name = 'mimic'\n",
    "# Define data dictionary\n",
    "data_directory = '../datasets/data/'+data_name + '/' + data_name + '_'\n",
    "\n",
    "# Load train and test datasets\n",
    "data_loader_training = CSVLoader(static_file=data_directory + 'static_train_data.csv.gz',\n",
    "                                 temporal_file=data_directory + 'temporal_train_data_eav.csv.gz')\n",
    "\n",
    "data_loader_testing = CSVLoader(static_file=data_directory + 'static_test_data.csv.gz',\n",
    "                                temporal_file=data_directory + 'temporal_test_data_eav.csv.gz')\n",
    "\n",
    "dataset_training = data_loader_training.load()\n",
    "dataset_testing = data_loader_testing.load()\n",
    "\n",
    "print('Finish data loading.')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 2: Preprocess dataset\n",
    "\n",
    "Preprocess the raw data using multiple filters. In this tutorial, we replace all the negative values to NaN (using NegativeFilter), do one-hot encoding on 'admission_type' feature (using OneHotEncoder), and do MinMax Normalization (using Normalization). Preprocessing is done for both training and testing datasets. \n",
    "  - NegativeFilter: Replace negative values to NaN\n",
    "  - OneHotEncoder: One hot encoding certain features\n",
    "    - one_hot_encoding: input features that need to be one-hot encoded\n",
    "  - Normalization (3 options): MinMax, Standard, None"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Finish preprocessing.\n"
     ]
    }
   ],
   "source": [
    "from preprocessing import FilterNegative, OneHotEncoder, Normalizer\n",
    "\n",
    "# (1) filter out negative values\n",
    "negative_filter = FilterNegative()\n",
    "# (2) one-hot encode categorical features\n",
    "one_hot_encoding = 'admission_type'\n",
    "onehot_encoder = OneHotEncoder(one_hot_encoding_features=[one_hot_encoding])\n",
    "# (3) Normalize features: 3 options (minmax, standard, none)\n",
    "normalization = 'minmax'\n",
    "normalizer = Normalizer(normalization)\n",
    "\n",
    "# Data preprocessing\n",
    "filter_pipeline = PipelineComposer(negative_filter, onehot_encoder, normalizer)\n",
    "\n",
    "dataset_training = filter_pipeline.fit_transform(dataset_training)\n",
    "dataset_testing = filter_pipeline.transform(dataset_testing)\n",
    "\n",
    "print('Finish preprocessing.')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 3: Define problem   \n",
    "\n",
    "Set the time-series prediction problem that we want to solve. Set the problem (whether it is one-shot or online prediction), set the label, set the maximum sequence length, and set the treatment features. We also define the metric for evaluation and the task itself (whether classification or regression). In this tutorial, we predict whether the patients will die at the end of the hospital stay (one-shot setting).\n",
    "  - problem: 'one-shot'(one time prediction) or 'online'(rolling window prediction)\n",
    "    - 'one-shot': one time prediction at the end of the time-series \n",
    "    - 'online': preditcion at every time stamps of the time-series\n",
    "  - max_seq_len: maximum sequence length of time-series sequence\n",
    "  - label_name: the column name for the label(s)\n",
    "  - treatment: the column name for treatments\n",
    "  - window: x-hour ahead prediction (for online prediction).\n",
    "  \n",
    "  - other parameters:\n",
    "    - metric_name: auc, apr, mse, mae\n",
    "    - task: classification or regression"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "100%|██████████| 18490/18490 [00:48<00:00, 378.11it/s]\n",
      "100%|██████████| 18490/18490 [00:51<00:00, 355.70it/s]\n",
      "100%|██████████| 4610/4610 [00:07<00:00, 590.43it/s]\n",
      "100%|██████████| 4610/4610 [00:08<00:00, 531.60it/s]\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Finish defining problem.\n"
     ]
    }
   ],
   "source": [
    "from preprocessing import ProblemMaker\n",
    "\n",
    "# Define parameters\n",
    "problem = 'one-shot'\n",
    "max_seq_len = 24\n",
    "label_name = 'death'\n",
    "treatment = None\n",
    "window = 0\n",
    "\n",
    "# Define problem \n",
    "problem_maker = ProblemMaker(problem=problem, label=[label_name],\n",
    "                             max_seq_len=max_seq_len, treatment=treatment, window = window)\n",
    "\n",
    "dataset_training = problem_maker.fit_transform(dataset_training)\n",
    "dataset_testing = problem_maker.fit_transform(dataset_testing)\n",
    "\n",
    "# Set other parameters\n",
    "metric_name = 'auc'\n",
    "task = 'classification'\n",
    "\n",
    "metric_sets = [metric_name]\n",
    "metric_parameters =  {'problem': problem, 'label_name': [label_name]}\n",
    "\n",
    "print('Finish defining problem.')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 4: Impute dataset\n",
    "\n",
    "Impute missing values in the preprocessed static and temporal datasets and return complete datasets.\n",
    "  - Static imputation (6 options): mean, median, mice, missforest, knn, gain\n",
    "  - Temporal imputation (8 options): mean, median, linear, quadratic, cubic, spline, mrnn, tgain"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Finish imputation.\n"
     ]
    }
   ],
   "source": [
    "from imputation import Imputation\n",
    "\n",
    "# Set imputation models\n",
    "static_imputation_model = 'median'\n",
    "temporal_imputation_model = 'median'\n",
    "\n",
    "# Impute the missing data\n",
    "static_imputation = Imputation(imputation_model_name = static_imputation_model, data_type = 'static')\n",
    "temporal_imputation = Imputation(imputation_model_name = temporal_imputation_model, data_type = 'temporal')\n",
    "\n",
    "imputation_pipeline = PipelineComposer(static_imputation, temporal_imputation)\n",
    "\n",
    "dataset_training = imputation_pipeline.fit_transform(dataset_training)\n",
    "dataset_testing = imputation_pipeline.transform(dataset_testing)\n",
    "\n",
    "print('Finish imputation.')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 5: Feature selection\n",
    "\n",
    "Select the relevant static and temporal features to the labels. If you do not want, we can skip the feature selection (set feature selection method = None).\n",
    "  - feature selection method: \n",
    "      - feature_selection_model: greedy-addition, greedy-deletion, recursive-addition, recursive-deletion, None\n",
    "      - feature_number: selected featuer number"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 6,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Finish feature selection.\n"
     ]
    }
   ],
   "source": [
    "from feature_selection import FeatureSelection\n",
    "\n",
    "# Set feature selection parameters\n",
    "static_feature_selection_model = None\n",
    "temporal_feature_selection_model = None\n",
    "static_feature_selection_number = None\n",
    "temporal_feature_selection_number = None\n",
    "\n",
    "# Select relevant features\n",
    "static_feature_selection = \\\n",
    "FeatureSelection(feature_selection_model_name = static_feature_selection_model,\n",
    "                 feature_type = 'static', feature_number = static_feature_selection_number,\n",
    "                 task = task, metric_name = metric_name,\n",
    "                 metric_parameters = metric_parameters)\n",
    "\n",
    "temporal_feature_selection = \\\n",
    "FeatureSelection(feature_selection_model_name = temporal_feature_selection_model,\n",
    "                 feature_type = 'temporal', feature_number = temporal_feature_selection_number,\n",
    "                 task = task, metric_name = metric_name,\n",
    "                 metric_parameters = metric_parameters)\n",
    "\n",
    "feature_selection_pipeline = PipelineComposer(static_feature_selection, temporal_feature_selection)\n",
    "\n",
    "dataset_training = feature_selection_pipeline.fit_transform(dataset_training)\n",
    "dataset_testing = feature_selection_pipeline.transform(dataset_testing)\n",
    "\n",
    "print('Finish feature selection.')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 6: Original time-series model fit and predict\n",
    "\n",
    "After finishing the data preparation, we define the original predictive models (6 options, RNN, GRU, LSTM, Attention, Temporal CNN, and Transformer), and train the original model using the training dataset. We set validation set as the 20% of the training set for early stopping and best model saving. After training, we use the original trained model to predict the labels of the testing dataset.\n",
    "\n",
    "- predictor_parameters:\n",
    "  - ori_model_name: rnn, gru, lstm\n",
    "  - ori_model_parameters: network parameters such as numer of layers\n",
    "    - h_dim: hidden dimensions\n",
    "    - n_layer: layer number\n",
    "    - n_head: head number (only for transformer model)\n",
    "    - batch_size: number of samples in mini-batch\n",
    "    - epochs: number of epochs\n",
    "    - learning_rate: learning rate\n",
    "  - static_mode: how to utilize static features (concatenate or None)\n",
    "  - time_mode: how to utilize time information (concatenate or None)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 7,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 14792 samples, validate on 3698 samples\n",
      "Epoch 1/20\n",
      "14792/14792 [==============================] - 5s 343us/sample - loss: 0.3233 - val_loss: 0.3153\n",
      "Epoch 2/20\n",
      "14792/14792 [==============================] - 4s 269us/sample - loss: 0.2931 - val_loss: 0.3022\n",
      "Epoch 3/20\n",
      "14792/14792 [==============================] - 4s 276us/sample - loss: 0.2825 - val_loss: 0.2986\n",
      "Epoch 4/20\n",
      "14792/14792 [==============================] - 4s 274us/sample - loss: 0.2766 - val_loss: 0.2901\n",
      "Epoch 5/20\n",
      "14792/14792 [==============================] - 4s 268us/sample - loss: 0.2740 - val_loss: 0.2836\n",
      "Epoch 6/20\n",
      "14792/14792 [==============================] - 4s 270us/sample - loss: 0.2737 - val_loss: 0.2953\n",
      "Epoch 7/20\n",
      "14792/14792 [==============================] - 4s 271us/sample - loss: 0.2673 - val_loss: 0.2789\n",
      "Epoch 8/20\n",
      "14792/14792 [==============================] - 4s 275us/sample - loss: 0.2651 - val_loss: 0.2860\n",
      "Epoch 9/20\n",
      "14792/14792 [==============================] - 4s 272us/sample - loss: 0.2602 - val_loss: 0.2758\n",
      "Epoch 10/20\n",
      "14792/14792 [==============================] - 4s 273us/sample - loss: 0.2572 - val_loss: 0.2717\n",
      "Epoch 11/20\n",
      "14792/14792 [==============================] - 4s 270us/sample - loss: 0.2557 - val_loss: 0.2769\n",
      "Epoch 12/20\n",
      "14792/14792 [==============================] - 4s 269us/sample - loss: 0.2544 - val_loss: 0.2851\n",
      "Epoch 13/20\n",
      "14792/14792 [==============================] - 4s 272us/sample - loss: 0.2521 - val_loss: 0.2685\n",
      "Epoch 14/20\n",
      "14792/14792 [==============================] - 4s 279us/sample - loss: 0.2494 - val_loss: 0.2707\n",
      "Epoch 15/20\n",
      "14792/14792 [==============================] - 4s 279us/sample - loss: 0.2473 - val_loss: 0.2972\n",
      "Epoch 16/20\n",
      "14792/14792 [==============================] - 4s 267us/sample - loss: 0.2474 - val_loss: 0.2810\n",
      "Epoch 17/20\n",
      "14792/14792 [==============================] - 4s 268us/sample - loss: 0.2442 - val_loss: 0.2668\n",
      "Epoch 18/20\n",
      "14792/14792 [==============================] - 4s 272us/sample - loss: 0.2419 - val_loss: 0.2714\n",
      "Epoch 19/20\n",
      "14792/14792 [==============================] - 4s 270us/sample - loss: 0.2444 - val_loss: 0.2997\n",
      "Epoch 20/20\n",
      "14792/14792 [==============================] - 4s 275us/sample - loss: 0.2399 - val_loss: 0.2745\n",
      "Finish predictor model training and testing.\n"
     ]
    }
   ],
   "source": [
    "from prediction import prediction\n",
    "\n",
    "# Set original predictive model\n",
    "ori_model_name = 'gru'\n",
    "\n",
    "# Set original model parameters\n",
    "ori_model_parameters = {'h_dim': 100,\n",
    "                        'n_layer': 2,\n",
    "                        'n_head': 2,\n",
    "                        'batch_size': 128,\n",
    "                        'epoch': 20,\n",
    "                        'model_type': ori_model_name,\n",
    "                        'learning_rate': 0.001,\n",
    "                        'static_mode': 'Concatenate',\n",
    "                        'time_mode': 'Concatenate',\n",
    "                        'verbose': True}\n",
    "\n",
    "# Set up validation for early stopping and best model saving\n",
    "dataset_training.train_val_test_split(prob_val=0.2, prob_test = 0.0)\n",
    "\n",
    "# Train the original predictive model\n",
    "ori_pred_class = prediction(ori_model_name, ori_model_parameters, task)\n",
    "ori_pred_class.fit(dataset_training)\n",
    "# Return the predictions on the testing set by the original predictive model\n",
    "ori_test_y_hat = ori_pred_class.predict(dataset_testing)\n",
    "\n",
    "print('Finish predictor model training and testing.')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 7: Transfer learning time-series model fit and predict\n",
    "\n",
    "After finishing the data preparation and original predictive model training, we define the transfer learning predictive models (3 options, RNN, GRU, LSTM), and train the model using the transfer dataset. After training, we use the trained transfer model to predict the labels of the testing dataset.\n",
    "\n",
    "- predictor_parameters:\n",
    "  - trans_model_name: rnn, gru, lstm\n",
    "  - trans_model_parameters: network parameters such as numer of layers\n",
    "    - h_dim: hidden dimensions\n",
    "    - n_layer: layer number\n",
    "    - n_head: head number (only for transformer model)\n",
    "    - batch_size: number of samples in mini-batch\n",
    "    - epochs: number of epochs\n",
    "    - learning_rate: learning rate\n",
    "  - static_mode: how to utilize static features (concatenate or None)\n",
    "  - time_mode: how to utilize time information (concatenate or None)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Train on 4610 samples, validate on 4610 samples\n",
      "Epoch 1/500\n",
      "4610/4610 [==============================] - 0s 107us/sample - loss: 0.3398 - val_loss: 0.2662\n",
      "Epoch 2/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2573 - val_loss: 0.2502\n",
      "Epoch 3/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2471 - val_loss: 0.2443\n",
      "Epoch 4/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2434 - val_loss: 0.2417\n",
      "Epoch 5/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2412 - val_loss: 0.2406\n",
      "Epoch 6/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2410 - val_loss: 0.2403\n",
      "Epoch 7/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2409 - val_loss: 0.2403\n",
      "Epoch 8/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2427 - val_loss: 0.2401\n",
      "Epoch 9/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2403 - val_loss: 0.2400\n",
      "Epoch 10/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2400 - val_loss: 0.2395\n",
      "Epoch 11/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2397 - val_loss: 0.2394\n",
      "Epoch 12/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2398 - val_loss: 0.2391\n",
      "Epoch 13/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2399 - val_loss: 0.2391\n",
      "Epoch 14/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2410 - val_loss: 0.2392\n",
      "Epoch 15/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2401 - val_loss: 0.2391\n",
      "Epoch 16/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2393 - val_loss: 0.2388\n",
      "Epoch 17/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2392 - val_loss: 0.2388\n",
      "Epoch 18/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2403 - val_loss: 0.2385\n",
      "Epoch 19/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2388 - val_loss: 0.2386\n",
      "Epoch 20/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2391 - val_loss: 0.2386\n",
      "Epoch 21/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2402 - val_loss: 0.2388\n",
      "Epoch 22/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2386 - val_loss: 0.2382\n",
      "Epoch 23/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2385 - val_loss: 0.2380\n",
      "Epoch 24/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2383 - val_loss: 0.2380\n",
      "Epoch 25/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2383 - val_loss: 0.2379\n",
      "Epoch 26/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2387 - val_loss: 0.2381\n",
      "Epoch 27/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2382 - val_loss: 0.2377\n",
      "Epoch 28/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2383 - val_loss: 0.2378\n",
      "Epoch 29/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2383 - val_loss: 0.2376\n",
      "Epoch 30/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2378 - val_loss: 0.2375\n",
      "Epoch 31/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2379 - val_loss: 0.2376\n",
      "Epoch 32/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2384 - val_loss: 0.2377\n",
      "Epoch 33/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2381 - val_loss: 0.2376\n",
      "Epoch 34/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2379 - val_loss: 0.2373\n",
      "Epoch 35/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2379 - val_loss: 0.2379\n",
      "Epoch 36/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2419 - val_loss: 0.2376\n",
      "Epoch 37/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2390 - val_loss: 0.2376\n",
      "Epoch 38/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2383 - val_loss: 0.2374\n",
      "Epoch 39/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2381 - val_loss: 0.2375\n",
      "Epoch 40/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2376 - val_loss: 0.2371\n",
      "Epoch 41/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2374 - val_loss: 0.2371\n",
      "Epoch 42/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2375 - val_loss: 0.2371\n",
      "Epoch 43/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2377 - val_loss: 0.2373\n",
      "Epoch 44/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2405 - val_loss: 0.2383\n",
      "Epoch 45/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2397 - val_loss: 0.2372\n",
      "Epoch 46/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2375 - val_loss: 0.2372\n",
      "Epoch 47/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2405 - val_loss: 0.2374\n",
      "Epoch 48/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2372 - val_loss: 0.2369\n",
      "Epoch 49/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2374 - val_loss: 0.2369\n",
      "Epoch 50/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2370 - val_loss: 0.2369\n",
      "Epoch 51/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2373 - val_loss: 0.2370\n",
      "Epoch 52/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2399 - val_loss: 0.2370\n",
      "Epoch 53/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2384 - val_loss: 0.2370\n",
      "Epoch 54/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2372 - val_loss: 0.2366\n",
      "Epoch 55/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2371 - val_loss: 0.2366\n",
      "Epoch 56/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2370 - val_loss: 0.2371\n",
      "Epoch 57/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2379 - val_loss: 0.2367\n",
      "Epoch 58/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2377 - val_loss: 0.2366\n",
      "Epoch 59/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2370 - val_loss: 0.2367\n",
      "Epoch 60/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2369 - val_loss: 0.2366\n",
      "Epoch 61/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2369 - val_loss: 0.2363\n",
      "Epoch 62/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2370 - val_loss: 0.2365\n",
      "Epoch 63/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2377 - val_loss: 0.2364\n",
      "Epoch 64/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2370 - val_loss: 0.2364\n",
      "Epoch 65/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2367 - val_loss: 0.2365\n",
      "Epoch 66/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2368 - val_loss: 0.2362\n",
      "Epoch 67/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2367 - val_loss: 0.2362\n",
      "Epoch 68/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2365 - val_loss: 0.2364\n",
      "Epoch 69/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2367 - val_loss: 0.2366\n",
      "Epoch 70/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2391 - val_loss: 0.2369\n",
      "Epoch 71/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2369 - val_loss: 0.2361\n",
      "Epoch 72/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2365 - val_loss: 0.2361\n",
      "Epoch 73/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2391 - val_loss: 0.2369\n",
      "Epoch 74/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2374 - val_loss: 0.2361\n",
      "Epoch 75/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2374 - val_loss: 0.2364\n",
      "Epoch 76/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2365 - val_loss: 0.2360\n",
      "Epoch 77/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2378 - val_loss: 0.2363\n",
      "Epoch 78/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2376 - val_loss: 0.2361\n",
      "Epoch 79/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2373 - val_loss: 0.2363\n",
      "Epoch 80/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2364 - val_loss: 0.2359\n",
      "Epoch 81/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2366 - val_loss: 0.2361\n",
      "Epoch 82/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2371 - val_loss: 0.2363\n",
      "Epoch 83/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2371 - val_loss: 0.2359\n",
      "Epoch 84/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2361 - val_loss: 0.2362\n",
      "Epoch 85/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2362 - val_loss: 0.2367\n",
      "Epoch 86/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2390 - val_loss: 0.2365\n",
      "Epoch 87/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2368 - val_loss: 0.2358\n",
      "Epoch 88/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2382 - val_loss: 0.2359\n",
      "Epoch 89/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2371 - val_loss: 0.2359\n",
      "Epoch 90/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2376 - val_loss: 0.2366\n",
      "Epoch 91/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2385 - val_loss: 0.2362\n",
      "Epoch 92/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2363 - val_loss: 0.2361\n",
      "Epoch 93/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2373 - val_loss: 0.2365\n",
      "Epoch 94/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2369 - val_loss: 0.2359\n",
      "Epoch 95/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2361 - val_loss: 0.2358\n",
      "Epoch 96/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2365 - val_loss: 0.2359\n",
      "Epoch 97/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2359 - val_loss: 0.2356\n",
      "Epoch 98/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2367 - val_loss: 0.2353\n",
      "Epoch 99/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2358 - val_loss: 0.2353\n",
      "Epoch 100/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2357 - val_loss: 0.2353\n",
      "Epoch 101/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2359 - val_loss: 0.2353\n",
      "Epoch 102/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2358 - val_loss: 0.2353\n",
      "Epoch 103/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2363 - val_loss: 0.2352\n",
      "Epoch 104/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2355 - val_loss: 0.2352\n",
      "Epoch 105/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2355 - val_loss: 0.2353\n",
      "Epoch 106/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2357 - val_loss: 0.2359\n",
      "Epoch 107/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2360 - val_loss: 0.2351\n",
      "Epoch 108/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2354 - val_loss: 0.2352\n",
      "Epoch 109/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2353 - val_loss: 0.2350\n",
      "Epoch 110/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2355 - val_loss: 0.2351\n",
      "Epoch 111/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2366 - val_loss: 0.2354\n",
      "Epoch 112/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2357 - val_loss: 0.2352\n",
      "Epoch 113/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2355 - val_loss: 0.2349\n",
      "Epoch 114/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2360 - val_loss: 0.2350\n",
      "Epoch 115/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2354 - val_loss: 0.2352\n",
      "Epoch 116/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2395 - val_loss: 0.2354\n",
      "Epoch 117/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2377 - val_loss: 0.2358\n",
      "Epoch 118/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2358 - val_loss: 0.2350\n",
      "Epoch 119/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2354 - val_loss: 0.2350\n",
      "Epoch 120/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2359 - val_loss: 0.2351\n",
      "Epoch 121/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2353 - val_loss: 0.2349\n",
      "Epoch 122/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2352 - val_loss: 0.2349\n",
      "Epoch 123/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2352 - val_loss: 0.2348\n",
      "Epoch 124/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2352 - val_loss: 0.2348\n",
      "Epoch 125/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2352 - val_loss: 0.2348\n",
      "Epoch 126/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2351 - val_loss: 0.2348\n",
      "Epoch 127/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2352 - val_loss: 0.2349\n",
      "Epoch 128/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2352 - val_loss: 0.2347\n",
      "Epoch 129/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2350 - val_loss: 0.2349\n",
      "Epoch 130/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2400 - val_loss: 0.2354\n",
      "Epoch 131/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2352 - val_loss: 0.2348\n",
      "Epoch 132/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2380 - val_loss: 0.2359\n",
      "Epoch 133/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2362 - val_loss: 0.2348\n",
      "Epoch 134/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2355 - val_loss: 0.2347\n",
      "Epoch 135/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2353 - val_loss: 0.2362\n",
      "Epoch 136/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2379 - val_loss: 0.2358\n",
      "Epoch 137/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2357 - val_loss: 0.2347\n",
      "Epoch 138/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2352 - val_loss: 0.2350\n",
      "Epoch 139/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2350 - val_loss: 0.2347\n",
      "Epoch 140/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2349 - val_loss: 0.2345\n",
      "Epoch 141/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2352 - val_loss: 0.2347\n",
      "Epoch 142/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2348 - val_loss: 0.2345\n",
      "Epoch 143/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2353 - val_loss: 0.2347\n",
      "Epoch 144/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2370 - val_loss: 0.2348\n",
      "Epoch 145/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2350 - val_loss: 0.2349\n",
      "Epoch 146/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2361 - val_loss: 0.2345\n",
      "Epoch 147/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2349 - val_loss: 0.2344\n",
      "Epoch 148/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2349 - val_loss: 0.2344\n",
      "Epoch 149/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2361 - val_loss: 0.2351\n",
      "Epoch 150/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2350 - val_loss: 0.2348\n",
      "Epoch 151/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2350 - val_loss: 0.2343\n",
      "Epoch 152/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2349 - val_loss: 0.2343\n",
      "Epoch 153/500\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2350 - val_loss: 0.2345\n",
      "Epoch 154/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2400 - val_loss: 0.2353\n",
      "Epoch 155/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2343 - val_loss: 0.2356\n",
      "Epoch 156/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2392 - val_loss: 0.2349\n",
      "Epoch 157/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2351 - val_loss: 0.2343\n",
      "Epoch 158/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2409 - val_loss: 0.2350\n",
      "Epoch 159/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2352 - val_loss: 0.2341\n",
      "Epoch 160/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2353 - val_loss: 0.2354\n",
      "Epoch 161/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2383 - val_loss: 0.2352\n",
      "Epoch 162/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2349 - val_loss: 0.2341\n",
      "Epoch 163/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2344 - val_loss: 0.2341\n",
      "Epoch 164/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2360 - val_loss: 0.2343\n",
      "Epoch 165/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2345 - val_loss: 0.2341\n",
      "Epoch 166/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2365 - val_loss: 0.2351\n",
      "Epoch 167/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2359 - val_loss: 0.2348\n",
      "Epoch 168/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2356 - val_loss: 0.2344\n",
      "Epoch 169/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2352 - val_loss: 0.2342\n",
      "Epoch 170/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2349 - val_loss: 0.2341\n",
      "Epoch 171/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2344 - val_loss: 0.2341\n",
      "Epoch 172/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2371 - val_loss: 0.2347\n",
      "Epoch 173/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2349 - val_loss: 0.2339\n",
      "Epoch 174/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2342 - val_loss: 0.2339\n",
      "Epoch 175/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2345 - val_loss: 0.2351\n",
      "Epoch 176/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2373 - val_loss: 0.2347\n",
      "Epoch 177/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2372 - val_loss: 0.2345\n",
      "Epoch 178/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2357 - val_loss: 0.2344\n",
      "Epoch 179/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2352 - val_loss: 0.2340\n",
      "Epoch 180/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2342 - val_loss: 0.2338\n",
      "Epoch 181/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2348 - val_loss: 0.2339\n",
      "Epoch 182/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2342 - val_loss: 0.2337\n",
      "Epoch 183/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2340 - val_loss: 0.2337\n",
      "Epoch 184/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2344 - val_loss: 0.2339\n",
      "Epoch 185/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2345 - val_loss: 0.2340\n",
      "Epoch 186/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2373 - val_loss: 0.2343\n",
      "Epoch 187/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2343 - val_loss: 0.2338\n",
      "Epoch 188/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2345 - val_loss: 0.2337\n",
      "Epoch 189/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2346 - val_loss: 0.2337\n",
      "Epoch 190/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2341 - val_loss: 0.2337\n",
      "Epoch 191/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2345 - val_loss: 0.2337\n",
      "Epoch 192/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2343 - val_loss: 0.2336\n",
      "Epoch 193/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2343 - val_loss: 0.2340\n",
      "Epoch 194/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2346 - val_loss: 0.2337\n",
      "Epoch 195/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2347 - val_loss: 0.2336\n",
      "Epoch 196/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2341 - val_loss: 0.2336\n",
      "Epoch 197/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2376 - val_loss: 0.2351\n",
      "Epoch 198/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2350 - val_loss: 0.2339\n",
      "Epoch 199/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2347 - val_loss: 0.2336\n",
      "Epoch 200/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2342 - val_loss: 0.2349\n",
      "Epoch 201/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2393 - val_loss: 0.2347\n",
      "Epoch 202/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2340 - val_loss: 0.2335\n",
      "Epoch 203/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2338 - val_loss: 0.2335\n",
      "Epoch 204/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2349 - val_loss: 0.2339\n",
      "Epoch 205/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2346 - val_loss: 0.2334\n",
      "Epoch 206/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2340 - val_loss: 0.2345\n",
      "Epoch 207/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2409 - val_loss: 0.2352\n",
      "Epoch 208/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2354 - val_loss: 0.2342\n",
      "Epoch 209/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2343 - val_loss: 0.2336\n",
      "Epoch 210/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2339 - val_loss: 0.2333\n",
      "Epoch 211/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2342 - val_loss: 0.2334\n",
      "Epoch 212/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2338 - val_loss: 0.2333\n",
      "Epoch 213/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2337 - val_loss: 0.2337\n",
      "Epoch 214/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2353 - val_loss: 0.2334\n",
      "Epoch 215/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2335 - val_loss: 0.2333\n",
      "Epoch 216/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2337 - val_loss: 0.2332\n",
      "Epoch 217/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2339 - val_loss: 0.2339\n",
      "Epoch 218/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2347 - val_loss: 0.2333\n",
      "Epoch 219/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2336 - val_loss: 0.2336\n",
      "Epoch 220/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2409 - val_loss: 0.2341\n",
      "Epoch 221/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2341 - val_loss: 0.2331\n",
      "Epoch 222/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2335 - val_loss: 0.2334\n",
      "Epoch 223/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2356 - val_loss: 0.2339\n",
      "Epoch 224/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2340 - val_loss: 0.2332\n",
      "Epoch 225/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2336 - val_loss: 0.2331\n",
      "Epoch 226/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2333 - val_loss: 0.2331\n",
      "Epoch 227/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2337 - val_loss: 0.2332\n",
      "Epoch 228/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2337 - val_loss: 0.2331\n",
      "Epoch 229/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2340 - val_loss: 0.2332\n",
      "Epoch 230/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2332 - val_loss: 0.2332\n",
      "Epoch 231/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2345 - val_loss: 0.2334\n",
      "Epoch 232/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2352 - val_loss: 0.2338\n",
      "Epoch 233/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2337 - val_loss: 0.2330\n",
      "Epoch 234/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2335 - val_loss: 0.2329\n",
      "Epoch 235/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2339 - val_loss: 0.2329\n",
      "Epoch 236/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2333 - val_loss: 0.2329\n",
      "Epoch 237/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2332 - val_loss: 0.2333\n",
      "Epoch 238/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2377 - val_loss: 0.2331\n",
      "Epoch 239/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2335 - val_loss: 0.2329\n",
      "Epoch 240/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2332 - val_loss: 0.2333\n",
      "Epoch 241/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2334 - val_loss: 0.2328\n",
      "Epoch 242/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2335 - val_loss: 0.2338\n",
      "Epoch 243/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2367 - val_loss: 0.2337\n",
      "Epoch 244/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2334 - val_loss: 0.2330\n",
      "Epoch 245/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2332 - val_loss: 0.2331\n",
      "Epoch 246/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2336 - val_loss: 0.2328\n",
      "Epoch 247/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2337 - val_loss: 0.2330\n",
      "Epoch 248/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2335 - val_loss: 0.2328\n",
      "Epoch 249/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2332 - val_loss: 0.2328\n",
      "Epoch 250/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2334 - val_loss: 0.2328\n",
      "Epoch 251/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2332 - val_loss: 0.2326\n",
      "Epoch 252/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2359 - val_loss: 0.2329\n",
      "Epoch 253/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2331 - val_loss: 0.2326\n",
      "Epoch 254/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2348 - val_loss: 0.2342\n",
      "Epoch 255/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2404 - val_loss: 0.2332\n",
      "Epoch 256/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2329 - val_loss: 0.2333\n",
      "Epoch 257/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2355 - val_loss: 0.2329\n",
      "Epoch 258/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2363 - val_loss: 0.2344\n",
      "Epoch 259/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2339 - val_loss: 0.2327\n",
      "Epoch 260/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2328 - val_loss: 0.2331\n",
      "Epoch 261/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2332 - val_loss: 0.2326\n",
      "Epoch 262/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2330 - val_loss: 0.2325\n",
      "Epoch 263/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2328 - val_loss: 0.2325\n",
      "Epoch 264/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2330 - val_loss: 0.2324\n",
      "Epoch 265/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2328 - val_loss: 0.2324\n",
      "Epoch 266/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2345 - val_loss: 0.2329\n",
      "Epoch 267/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2334 - val_loss: 0.2325\n",
      "Epoch 268/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2334 - val_loss: 0.2332\n",
      "Epoch 269/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2352 - val_loss: 0.2334\n",
      "Epoch 270/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2332 - val_loss: 0.2324\n",
      "Epoch 271/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2331 - val_loss: 0.2324\n",
      "Epoch 272/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2329 - val_loss: 0.2324\n",
      "Epoch 273/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2359 - val_loss: 0.2328\n",
      "Epoch 274/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2343 - val_loss: 0.2333\n",
      "Epoch 275/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2330 - val_loss: 0.2322\n",
      "Epoch 276/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2324 - val_loss: 0.2322\n",
      "Epoch 277/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2325 - val_loss: 0.2322\n",
      "Epoch 278/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2329 - val_loss: 0.2322\n",
      "Epoch 279/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2327 - val_loss: 0.2321\n",
      "Epoch 280/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2426 - val_loss: 0.2338\n",
      "Epoch 281/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2340 - val_loss: 0.2326\n",
      "Epoch 282/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2330 - val_loss: 0.2324\n",
      "Epoch 283/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2329 - val_loss: 0.2321\n",
      "Epoch 284/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2332 - val_loss: 0.2321\n",
      "Epoch 285/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2325 - val_loss: 0.2324\n",
      "Epoch 286/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2326 - val_loss: 0.2321\n",
      "Epoch 287/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2324 - val_loss: 0.2321\n",
      "Epoch 288/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2324 - val_loss: 0.2320\n",
      "Epoch 289/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2325 - val_loss: 0.2321\n",
      "Epoch 290/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2324 - val_loss: 0.2320\n",
      "Epoch 291/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2342 - val_loss: 0.2322\n",
      "Epoch 292/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2327 - val_loss: 0.2322\n",
      "Epoch 293/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2322 - val_loss: 0.2320\n",
      "Epoch 294/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2328 - val_loss: 0.2323\n",
      "Epoch 295/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2323 - val_loss: 0.2321\n",
      "Epoch 296/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2348 - val_loss: 0.2324\n",
      "Epoch 297/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2325 - val_loss: 0.2319\n",
      "Epoch 298/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2326 - val_loss: 0.2319\n",
      "Epoch 299/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2323 - val_loss: 0.2320\n",
      "Epoch 300/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2324 - val_loss: 0.2318\n",
      "Epoch 301/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2323 - val_loss: 0.2318\n",
      "Epoch 302/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2353 - val_loss: 0.2324\n",
      "Epoch 303/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2325 - val_loss: 0.2321\n",
      "Epoch 304/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2329 - val_loss: 0.2321\n",
      "Epoch 305/500\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2327 - val_loss: 0.2322\n",
      "Epoch 306/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2325 - val_loss: 0.2319\n",
      "Epoch 307/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2324 - val_loss: 0.2317\n",
      "Epoch 308/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2319 - val_loss: 0.2318\n",
      "Epoch 309/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2328 - val_loss: 0.2321\n",
      "Epoch 310/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2332 - val_loss: 0.2317\n",
      "Epoch 311/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2322 - val_loss: 0.2319\n",
      "Epoch 312/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2325 - val_loss: 0.2320\n",
      "Epoch 313/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2333 - val_loss: 0.2335\n",
      "Epoch 314/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2367 - val_loss: 0.2317\n",
      "Epoch 315/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2320 - val_loss: 0.2315\n",
      "Epoch 316/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2320 - val_loss: 0.2315\n",
      "Epoch 317/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2319 - val_loss: 0.2317\n",
      "Epoch 318/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2322 - val_loss: 0.2315\n",
      "Epoch 319/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2317 - val_loss: 0.2316\n",
      "Epoch 320/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2326 - val_loss: 0.2314\n",
      "Epoch 321/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2315 - val_loss: 0.2313\n",
      "Epoch 322/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2323 - val_loss: 0.2315\n",
      "Epoch 323/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2320 - val_loss: 0.2312\n",
      "Epoch 324/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2316 - val_loss: 0.2312\n",
      "Epoch 325/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2316 - val_loss: 0.2312\n",
      "Epoch 326/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2318 - val_loss: 0.2313\n",
      "Epoch 327/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2314 - val_loss: 0.2312\n",
      "Epoch 328/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2319 - val_loss: 0.2313\n",
      "Epoch 329/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2318 - val_loss: 0.2310\n",
      "Epoch 330/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2314 - val_loss: 0.2310\n",
      "Epoch 331/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2318 - val_loss: 0.2312\n",
      "Epoch 332/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2315 - val_loss: 0.2310\n",
      "Epoch 333/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2313 - val_loss: 0.2310\n",
      "Epoch 334/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2316 - val_loss: 0.2310\n",
      "Epoch 335/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2325 - val_loss: 0.2311\n",
      "Epoch 336/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2332 - val_loss: 0.2318\n",
      "Epoch 337/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2323 - val_loss: 0.2317\n",
      "Epoch 338/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2313 - val_loss: 0.2317\n",
      "Epoch 339/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2316 - val_loss: 0.2310\n",
      "Epoch 340/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2318 - val_loss: 0.2308\n",
      "Epoch 341/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2312 - val_loss: 0.2309\n",
      "Epoch 342/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2312 - val_loss: 0.2307\n",
      "Epoch 343/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2311 - val_loss: 0.2310\n",
      "Epoch 344/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2319 - val_loss: 0.2320\n",
      "Epoch 345/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2339 - val_loss: 0.2310\n",
      "Epoch 346/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2313 - val_loss: 0.2312\n",
      "Epoch 347/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2325 - val_loss: 0.2316\n",
      "Epoch 348/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2357 - val_loss: 0.2313\n",
      "Epoch 349/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2316 - val_loss: 0.2306\n",
      "Epoch 350/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2312 - val_loss: 0.2305\n",
      "Epoch 351/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2312 - val_loss: 0.2309\n",
      "Epoch 352/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2309 - val_loss: 0.2308\n",
      "Epoch 353/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2310 - val_loss: 0.2304\n",
      "Epoch 354/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2308 - val_loss: 0.2304\n",
      "Epoch 355/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2310 - val_loss: 0.2305\n",
      "Epoch 356/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2313 - val_loss: 0.2306\n",
      "Epoch 357/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2311 - val_loss: 0.2309\n",
      "Epoch 358/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2311 - val_loss: 0.2304\n",
      "Epoch 359/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2331 - val_loss: 0.2311\n",
      "Epoch 360/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2314 - val_loss: 0.2306\n",
      "Epoch 361/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2312 - val_loss: 0.2305\n",
      "Epoch 362/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2308 - val_loss: 0.2303\n",
      "Epoch 363/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2309 - val_loss: 0.2301\n",
      "Epoch 364/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2314 - val_loss: 0.2311\n",
      "Epoch 365/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2311 - val_loss: 0.2306\n",
      "Epoch 366/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2357 - val_loss: 0.2313\n",
      "Epoch 367/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2307 - val_loss: 0.2301\n",
      "Epoch 368/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2305 - val_loss: 0.2306\n",
      "Epoch 369/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2360 - val_loss: 0.2316\n",
      "Epoch 370/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2309 - val_loss: 0.2302\n",
      "Epoch 371/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2309 - val_loss: 0.2303\n",
      "Epoch 372/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2307 - val_loss: 0.2300\n",
      "Epoch 373/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2318 - val_loss: 0.2303\n",
      "Epoch 374/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2305 - val_loss: 0.2317\n",
      "Epoch 375/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2331 - val_loss: 0.2301\n",
      "Epoch 376/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2307 - val_loss: 0.2299\n",
      "Epoch 377/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2308 - val_loss: 0.2299\n",
      "Epoch 378/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2305 - val_loss: 0.2299\n",
      "Epoch 379/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2310 - val_loss: 0.2300\n",
      "Epoch 380/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2302 - val_loss: 0.2300\n",
      "Epoch 381/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2302 - val_loss: 0.2299\n",
      "Epoch 382/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2308 - val_loss: 0.2297\n",
      "Epoch 383/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2311 - val_loss: 0.2300\n",
      "Epoch 384/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2301 - val_loss: 0.2297\n",
      "Epoch 385/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2301 - val_loss: 0.2298\n",
      "Epoch 386/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2303 - val_loss: 0.2296\n",
      "Epoch 387/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2311 - val_loss: 0.2296\n",
      "Epoch 388/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2301 - val_loss: 0.2299\n",
      "Epoch 389/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2303 - val_loss: 0.2297\n",
      "Epoch 390/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2297 - val_loss: 0.2297\n",
      "Epoch 391/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2300 - val_loss: 0.2299\n",
      "Epoch 392/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2356 - val_loss: 0.2316\n",
      "Epoch 393/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2306 - val_loss: 0.2305\n",
      "Epoch 394/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2409 - val_loss: 0.2308\n",
      "Epoch 395/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2304 - val_loss: 0.2297\n",
      "Epoch 396/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2304 - val_loss: 0.2305\n",
      "Epoch 397/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2373 - val_loss: 0.2306\n",
      "Epoch 398/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2303 - val_loss: 0.2294\n",
      "Epoch 399/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2301 - val_loss: 0.2293\n",
      "Epoch 400/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2296 - val_loss: 0.2294\n",
      "Epoch 401/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2305 - val_loss: 0.2292\n",
      "Epoch 402/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2311 - val_loss: 0.2294\n",
      "Epoch 403/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2298 - val_loss: 0.2292\n",
      "Epoch 404/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2297 - val_loss: 0.2292\n",
      "Epoch 405/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2294 - val_loss: 0.2292\n",
      "Epoch 406/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2303 - val_loss: 0.2295\n",
      "Epoch 407/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2296 - val_loss: 0.2292\n",
      "Epoch 408/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2297 - val_loss: 0.2293\n",
      "Epoch 409/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2318 - val_loss: 0.2295\n",
      "Epoch 410/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2305 - val_loss: 0.2293\n",
      "Epoch 411/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2307 - val_loss: 0.2296\n",
      "Epoch 412/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2298 - val_loss: 0.2292\n",
      "Epoch 413/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2297 - val_loss: 0.2298\n",
      "Epoch 414/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2295 - val_loss: 0.2290\n",
      "Epoch 415/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2294 - val_loss: 0.2291\n",
      "Epoch 416/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2293 - val_loss: 0.2290\n",
      "Epoch 417/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2292 - val_loss: 0.2288\n",
      "Epoch 418/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2294 - val_loss: 0.2288\n",
      "Epoch 419/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2292 - val_loss: 0.2288\n",
      "Epoch 420/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2292 - val_loss: 0.2288\n",
      "Epoch 421/500\n",
      "4610/4610 [==============================] - 0s 12us/sample - loss: 0.2291 - val_loss: 0.2287\n",
      "Epoch 422/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2292 - val_loss: 0.2288\n",
      "Epoch 423/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2326 - val_loss: 0.2293\n",
      "Epoch 424/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2293 - val_loss: 0.2303\n",
      "Epoch 425/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2323 - val_loss: 0.2291\n",
      "Epoch 426/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2291 - val_loss: 0.2289\n",
      "Epoch 427/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2301 - val_loss: 0.2288\n",
      "Epoch 428/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2290 - val_loss: 0.2287\n",
      "Epoch 429/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2291 - val_loss: 0.2289\n",
      "Epoch 430/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2314 - val_loss: 0.2288\n",
      "Epoch 431/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2292 - val_loss: 0.2290\n",
      "Epoch 432/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2302 - val_loss: 0.2307\n",
      "Epoch 433/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2317 - val_loss: 0.2292\n",
      "Epoch 434/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2290 - val_loss: 0.2284\n",
      "Epoch 435/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2289 - val_loss: 0.2284\n",
      "Epoch 436/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2312 - val_loss: 0.2287\n",
      "Epoch 437/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2334 - val_loss: 0.2298\n",
      "Epoch 438/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2295 - val_loss: 0.2289\n",
      "Epoch 439/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2290 - val_loss: 0.2285\n",
      "Epoch 440/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2291 - val_loss: 0.2285\n",
      "Epoch 441/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2293 - val_loss: 0.2283\n",
      "Epoch 442/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2299 - val_loss: 0.2289\n",
      "Epoch 443/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2291 - val_loss: 0.2284\n",
      "Epoch 444/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2292 - val_loss: 0.2287\n",
      "Epoch 445/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2287 - val_loss: 0.2282\n",
      "Epoch 446/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2287 - val_loss: 0.2281\n",
      "Epoch 447/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2286 - val_loss: 0.2281\n",
      "Epoch 448/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2298 - val_loss: 0.2282\n",
      "Epoch 449/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2289 - val_loss: 0.2281\n",
      "Epoch 450/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2287 - val_loss: 0.2286\n",
      "Epoch 451/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2314 - val_loss: 0.2284\n",
      "Epoch 452/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2283 - val_loss: 0.2289\n",
      "Epoch 453/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2344 - val_loss: 0.2308\n",
      "Epoch 454/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2303 - val_loss: 0.2287\n",
      "Epoch 455/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2282 - val_loss: 0.2279\n",
      "Epoch 456/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2307 - val_loss: 0.2283\n",
      "Epoch 457/500\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2294 - val_loss: 0.2281\n",
      "Epoch 458/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2285 - val_loss: 0.2279\n",
      "Epoch 459/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2285 - val_loss: 0.2278\n",
      "Epoch 460/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2283 - val_loss: 0.2277\n",
      "Epoch 461/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2281 - val_loss: 0.2284\n",
      "Epoch 462/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2281 - val_loss: 0.2278\n",
      "Epoch 463/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2325 - val_loss: 0.2292\n",
      "Epoch 464/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2288 - val_loss: 0.2278\n",
      "Epoch 465/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2285 - val_loss: 0.2284\n",
      "Epoch 466/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2303 - val_loss: 0.2280\n",
      "Epoch 467/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2285 - val_loss: 0.2285\n",
      "Epoch 468/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2281 - val_loss: 0.2279\n",
      "Epoch 469/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2283 - val_loss: 0.2275\n",
      "Epoch 470/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2280 - val_loss: 0.2275\n",
      "Epoch 471/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2279 - val_loss: 0.2283\n",
      "Epoch 472/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2295 - val_loss: 0.2280\n",
      "Epoch 473/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2330 - val_loss: 0.2285\n",
      "Epoch 474/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2286 - val_loss: 0.2276\n",
      "Epoch 475/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2277 - val_loss: 0.2274\n",
      "Epoch 476/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2285 - val_loss: 0.2283\n",
      "Epoch 477/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2283 - val_loss: 0.2274\n",
      "Epoch 478/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2281 - val_loss: 0.2273\n",
      "Epoch 479/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2277 - val_loss: 0.2273\n",
      "Epoch 480/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2279 - val_loss: 0.2273\n",
      "Epoch 481/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2284 - val_loss: 0.2275\n",
      "Epoch 482/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2280 - val_loss: 0.2274\n",
      "Epoch 483/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2284 - val_loss: 0.2274\n",
      "Epoch 484/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2281 - val_loss: 0.2286\n",
      "Epoch 485/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2301 - val_loss: 0.2281\n",
      "Epoch 486/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2281 - val_loss: 0.2274\n",
      "Epoch 487/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2287 - val_loss: 0.2272\n",
      "Epoch 488/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2297 - val_loss: 0.2286\n",
      "Epoch 489/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2278 - val_loss: 0.2271\n",
      "Epoch 490/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2292 - val_loss: 0.2277\n",
      "Epoch 491/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2284 - val_loss: 0.2271\n",
      "Epoch 492/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2281 - val_loss: 0.2272\n",
      "Epoch 493/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2282 - val_loss: 0.2294\n",
      "Epoch 494/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2366 - val_loss: 0.2291\n",
      "Epoch 495/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2282 - val_loss: 0.2271\n",
      "Epoch 496/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2275 - val_loss: 0.2270\n",
      "Epoch 497/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2276 - val_loss: 0.2270\n",
      "Epoch 498/500\n",
      "4610/4610 [==============================] - 0s 9us/sample - loss: 0.2272 - val_loss: 0.2270\n",
      "Epoch 499/500\n",
      "4610/4610 [==============================] - 0s 10us/sample - loss: 0.2273 - val_loss: 0.2272\n",
      "Epoch 500/500\n",
      "4610/4610 [==============================] - 0s 11us/sample - loss: 0.2276 - val_loss: 0.2268\n",
      "Finish transfer learning model training and testing.\n"
     ]
    }
   ],
   "source": [
    "from prediction import TransferLearning\n",
    "\n",
    "# Set transfer learning model\n",
    "trans_model_name = 'gru'\n",
    "\n",
    "# Set model parameters\n",
    "trans_model_parameters = {'h_dim': 100,\n",
    "                        'n_layer': 2,\n",
    "                        'n_head': 2,\n",
    "                        'batch_size': 128,\n",
    "                        'epoch': 50,\n",
    "                        'model_type': trans_model_name,\n",
    "                        'learning_rate': 0.001,\n",
    "                        'static_mode': 'Concatenate',\n",
    "                        'time_mode': 'Concatenate',\n",
    "                        'verbose': True}\n",
    "\n",
    "trans_class = TransferLearning(ori_pred_class, trans_model_parameters, task)\n",
    "trans_class.fit(dataset_testing)\n",
    "trans_test_y_hat = trans_class.predict(dataset_testing) \n",
    "    \n",
    "print('Finish transfer learning model training and testing.') "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 8: Visualize results\n",
    "\n",
    "(1) Visualize the performance of the trained model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Finish original and transfer predictor model evaluation.\n",
      "Overall performance of the original model\n",
      "+--------------+-------+\n",
      "| metric/label | death |\n",
      "+==============+=======+\n",
      "|     auc      | 0.865 |\n",
      "+--------------+-------+\n",
      "Overall performance of the transfer model\n",
      "+--------------+-------+\n",
      "| metric/label | death |\n",
      "+==============+=======+\n",
      "|     auc      | 0.880 |\n",
      "+--------------+-------+\n"
     ]
    }
   ],
   "source": [
    "from evaluation import Metrics\n",
    "from evaluation import print_performance\n",
    "\n",
    "# Evaluate predictor model\n",
    "_, _, test_y, _, _ = dataset_testing.get_fold(0, 'test')\n",
    "ori_result = Metrics(metric_sets, metric_parameters).evaluate(test_y, ori_test_y_hat)\n",
    "trans_result = Metrics(metric_sets, metric_parameters).evaluate(test_y, trans_test_y_hat)\n",
    "print('Finish original and transfer predictor model evaluation.')\n",
    "\n",
    "print('Overall performance of the original model')\n",
    "print_performance(ori_result, metric_sets, metric_parameters)\n",
    "print('Overall performance of the transfer model')\n",
    "print_performance(trans_result, metric_sets, metric_parameters)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Step 8: Visualize results\n",
    "\n",
    "(2) Visualize the predictions of a certain patient by trained predictive model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Each prediction of the original model\n",
      "+----------+-------+\n",
      "| id/label | death |\n",
      "+==========+=======+\n",
      "|    1     | 0.056 |\n",
      "+----------+-------+\n",
      "Each prediction of the transfer model\n",
      "+----------+-------+\n",
      "| id/label | death |\n",
      "+==========+=======+\n",
      "|    1     | 0.048 |\n",
      "+----------+-------+\n"
     ]
    }
   ],
   "source": [
    "from evaluation import print_prediction\n",
    "\n",
    "# Set the patient index for visualization\n",
    "index = [1]\n",
    "\n",
    "print('Each prediction of the original model')\n",
    "print_prediction(ori_test_y_hat[index], metric_parameters)\n",
    "print('Each prediction of the transfer model')\n",
    "print_prediction(trans_test_y_hat[index], metric_parameters)"
   ]
  },
  {
   "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": "3.7.4"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}