Tacton_MLmodel2.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This tutorial will show how to use and modify the data generator pack for Tacton.\n",
    "Data objects that are relevant for this project are:\n",
    "C_data - Combination data, \n",
    "D_data - Discount data, \n",
    "p_data - probability data, \n",
    "deal_data - Quote outcome. \n",
    "For more information regarding the algorithms behind the data generator: \n",
    "https://docs.google.com/document/d/10QUrTwTu61wxq57S-zQgsrRc2BtC6QuKa8mXj4Q4f84/edit"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "from matplotlib import pylab as plt\n",
    "from generate_cpq_data_class import *\n",
    "import psutil\n",
    "import os\n",
    "\n",
    "os.environ['KMP_DUPLICATE_LIB_OK']='True'"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "virtual memory used:  56.6 %\n"
     ]
    }
   ],
   "source": [
    "print(\"virtual memory used: \", psutil.virtual_memory()[2], \"%\")\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "The Generator_pack creates a C-gen (Combinations/dimensions-generator), D_gen (Discount-generator), a P (Probability function) and a corresponding P_gen (Probability-generator).\n",
    "\n",
    "We start this tutortial by initiate a Generator_pack G. discount_max, discount_min, discount_mean, discount_std are parameters for the distribution of the discount data (truncated normal distribution). mean_p and max_p are parameters for the probability distribution. "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 12,
   "metadata": {},
   "outputs": [],
   "source": [
    "G=Generator_pack(discount_max = 0.4, discount_min=0, discount_mean = 0.15, discount_std = 0.2, mean_p = 0.2,\n",
    "                  max_p = 0.4)\n",
    "\n",
    "#C_3 = [3,2,0,2,2,2]\n",
    "#G.set_C(C_3, p=lambda d:np.sin(d*2*np.pi/0.4), n= lambda: 1000, d_dist= lambda nr: 0.4*np.random.rand(nr))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import pickle\n",
    "\n",
    "training_data = pickle.load( open( \"training_data.p\", \"rb\" ) )\n",
    "C_train = training_data['C_train']\n",
    "D_train = training_data['D_train']\n",
    "p_train = training_data['p_train']\n",
    "deal_train = training_data['deal_train']\n",
    "\n",
    "test_data = pickle.load( open( \"test_data.p\", \"rb\" ) )\n",
    "C_test = test_data['C_test']\n",
    "D_test = test_data['D_test']\n",
    "p_test = test_data['p_test']\n",
    "deal_test = test_data['deal_test']"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Generate training and test data from G."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 13,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Generated data:  611312 / 611312\n"
     ]
    }
   ],
   "source": [
    "#import pickle\n",
    "#import os\n",
    "\n",
    "C_train, D_train, p_train, deal_train = G.generate_data(copyfunction=lambda: int(np.maximum(0,np.random.normal(100,0.2))))\n",
    "#training_data = {'C_train': C_train, \"D_train\": D_train, \"p_train\": p_train, \"deal_train\": deal_train}\n",
    "#pickle.dump( training_data, open( \"training_data.p\", \"wb\" ) )"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 14,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Generated data:  611360 / 611360\n"
     ]
    }
   ],
   "source": [
    "C_test, D_test, p_test, deal_test = G.generate_data(copyfunction=lambda: int(np.maximum(0,np.random.normal(100,0.2))))\n",
    "#test_data = {'C_test': C_test, \"D_test\": D_test, \"p_test\": p_test, \"deal_test\": deal_test}\n",
    "#pickle.dump( test_data, open( \"test_data.p\", \"wb\" ) )"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Define the model shell"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 48,
   "metadata": {},
   "outputs": [],
   "source": [
    "def custom_loss_function(y, p_y):\n",
    "    #logloss\n",
    "    return -p_y.log_prob(y)\n",
    "\n",
    "eps = 0.000001\n",
    "input_shape = (30,)\n",
    "\n",
    "def create_model2(input_shape=(30,)):\n",
    "    C = tf.keras.layers.Input(shape=input_shape, name='C_data')\n",
    "    dis = tf.keras.layers.Input(shape=(1,), name='D_data')\n",
    "\n",
    "    k_ = tf.keras.layers.Dense(20, name='k_layer')(C)\n",
    "    m_ = tf.keras.layers.Dense(20, name='m_layer')(C)\n",
    "\n",
    "    #k_ = tf.keras.layers.Dense(10, activation='relu', name='k_layer')(C)\n",
    "    #k_ = tf.keras.layers.Dense(10, name='k_layer2')(k_)\n",
    "\n",
    "    #m_ = tf.keras.layers.Dense(100, activation='relu', name='m_layer')(C)\n",
    "    #m_ = tf.keras.layers.Dense(1000, name='m_layer2')(m_)\n",
    "\n",
    "    z = tf.math.multiply(k_,dis, name='multiply_k')\n",
    "\n",
    "    z = tf.math.add(z,m_,name='add_m')\n",
    "    z = tf.keras.layers.Activation('sigmoid',name='h')(z)\n",
    "    #z = tf.keras.layers.Dense(100,activation='sigmoid')(z)\n",
    "\n",
    "    z = tf.keras.layers.Dense(1)(z)\n",
    "\n",
    "    p = tf.keras.layers.Lambda(lambda t: tf.sigmoid(t) * (1.0 - eps) + eps / 2)(z)\n",
    "    Out = tfp.layers.DistributionLambda(lambda t: tfd.Bernoulli(probs=t))(p)\n",
    "\n",
    "    model = tf.keras.models.Model(inputs=[dis,C], outputs=Out)\n",
    "    model.compile(optimizer=tf.optimizers.Adam(learning_rate=0.001, decay=0), loss=custom_loss_function, metrics=['mae'])\n",
    "    return model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 16,
   "metadata": {},
   "outputs": [],
   "source": [
    "def logloss(true_label, predicted, eps=1e-15):\n",
    "    p = np.clip(predicted, eps, 1 - eps)\n",
    "    return -np.log(p*true_label + (1-p)*(1- true_label))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Define the model M"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 49,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Model: \"model_15\"\n",
      "__________________________________________________________________________________________________\n",
      "Layer (type)                    Output Shape         Param #     Connected to                     \n",
      "==================================================================================================\n",
      "C_data (InputLayer)             [(None, 30)]         0                                            \n",
      "__________________________________________________________________________________________________\n",
      "k_layer (Dense)                 (None, 20)           620         C_data[0][0]                     \n",
      "__________________________________________________________________________________________________\n",
      "D_data (InputLayer)             [(None, 1)]          0                                            \n",
      "__________________________________________________________________________________________________\n",
      "tf_op_layer_multiply_k_15 (Tens [(None, 20)]         0           k_layer[0][0]                    \n",
      "                                                                 D_data[0][0]                     \n",
      "__________________________________________________________________________________________________\n",
      "m_layer (Dense)                 (None, 20)           620         C_data[0][0]                     \n",
      "__________________________________________________________________________________________________\n",
      "tf_op_layer_add_m_15 (TensorFlo [(None, 20)]         0           tf_op_layer_multiply_k_15[0][0]  \n",
      "                                                                 m_layer[0][0]                    \n",
      "__________________________________________________________________________________________________\n",
      "h (Activation)                  (None, 20)           0           tf_op_layer_add_m_15[0][0]       \n",
      "__________________________________________________________________________________________________\n",
      "dense_15 (Dense)                (None, 1)            21          h[0][0]                          \n",
      "__________________________________________________________________________________________________\n",
      "lambda_15 (Lambda)              (None, 1)            0           dense_15[0][0]                   \n",
      "__________________________________________________________________________________________________\n",
      "distribution_lambda_15 (Distrib ((None, 1), (None, 1 0           lambda_15[0][0]                  \n",
      "==================================================================================================\n",
      "Total params: 1,261\n",
      "Trainable params: 1,261\n",
      "Non-trainable params: 0\n",
      "__________________________________________________________________________________________________\n"
     ]
    }
   ],
   "source": [
    "from MLmodel import*\n",
    "M = MLmodel(model_shell=create_model2())\n",
    "M.model.summary()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Train the model."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "true logloss train set:  0.36471088169276145\n",
      "true logloss test set:  0.36694466644848245\n",
      "training starts\n",
      "Train on 611312 samples, validate on 50000 samples\n",
      "Epoch 1/500\n",
      "611312/611312 [==============================] - 17s 28us/sample - loss: 0.3877 - mae: 0.2421 - val_loss: 0.3808 - val_mae: 0.2365\n",
      "Epoch 2/500\n",
      "611312/611312 [==============================] - 16s 25us/sample - loss: 0.3783 - mae: 0.2359 - val_loss: 0.3790 - val_mae: 0.2418\n",
      "Epoch 3/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3769 - mae: 0.2350 - val_loss: 0.3778 - val_mae: 0.2391\n",
      "Epoch 4/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3760 - mae: 0.2347 - val_loss: 0.3771 - val_mae: 0.2360\n",
      "Epoch 5/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3753 - mae: 0.2347 - val_loss: 0.3766 - val_mae: 0.2367\n",
      "Epoch 6/500\n",
      "611312/611312 [==============================] - 15s 24us/sample - loss: 0.3747 - mae: 0.2347 - val_loss: 0.3761 - val_mae: 0.2398\n",
      "Epoch 7/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3741 - mae: 0.2346 - val_loss: 0.3757 - val_mae: 0.2341\n",
      "Epoch 8/500\n",
      "611312/611312 [==============================] - 15s 24us/sample - loss: 0.3737 - mae: 0.2343 - val_loss: 0.3752 - val_mae: 0.2373\n",
      "Epoch 9/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3731 - mae: 0.2348 - val_loss: 0.3750 - val_mae: 0.2296\n",
      "Epoch 10/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3726 - mae: 0.2346 - val_loss: 0.3748 - val_mae: 0.2297\n",
      "Epoch 11/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3721 - mae: 0.2338 - val_loss: 0.3744 - val_mae: 0.2341\n",
      "Epoch 12/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3718 - mae: 0.2337 - val_loss: 0.3740 - val_mae: 0.2345\n",
      "Epoch 13/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3714 - mae: 0.2341 - val_loss: 0.3737 - val_mae: 0.2379\n",
      "Epoch 14/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3712 - mae: 0.2340 - val_loss: 0.3736 - val_mae: 0.2342\n",
      "Epoch 15/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3710 - mae: 0.2327 - val_loss: 0.3734 - val_mae: 0.2315\n",
      "Epoch 16/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3708 - mae: 0.2335 - val_loss: 0.3731 - val_mae: 0.2320\n",
      "Epoch 17/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3706 - mae: 0.2337 - val_loss: 0.3731 - val_mae: 0.2300\n",
      "Epoch 18/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3704 - mae: 0.2326 - val_loss: 0.3731 - val_mae: 0.2309\n",
      "Epoch 19/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3703 - mae: 0.2331 - val_loss: 0.3728 - val_mae: 0.2334\n",
      "Epoch 20/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3701 - mae: 0.2337 - val_loss: 0.3728 - val_mae: 0.2315\n",
      "Epoch 21/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3700 - mae: 0.2331 - val_loss: 0.3726 - val_mae: 0.2372\n",
      "Epoch 22/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3698 - mae: 0.2326 - val_loss: 0.3728 - val_mae: 0.2293\n",
      "Epoch 23/500\n",
      "611312/611312 [==============================] - 18s 29us/sample - loss: 0.3697 - mae: 0.2329 - val_loss: 0.3723 - val_mae: 0.2369\n",
      "Epoch 24/500\n",
      "611312/611312 [==============================] - 16s 25us/sample - loss: 0.3696 - mae: 0.2334 - val_loss: 0.3725 - val_mae: 0.2327\n",
      "Epoch 25/500\n",
      "611312/611312 [==============================] - 15s 24us/sample - loss: 0.3695 - mae: 0.2321 - val_loss: 0.3722 - val_mae: 0.2338\n",
      "Epoch 26/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3695 - mae: 0.2330 - val_loss: 0.3724 - val_mae: 0.2315\n",
      "Epoch 27/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3694 - mae: 0.2331 - val_loss: 0.3719 - val_mae: 0.2330\n",
      "Epoch 28/500\n",
      "611312/611312 [==============================] - 15s 24us/sample - loss: 0.3693 - mae: 0.2328 - val_loss: 0.3726 - val_mae: 0.2327\n",
      "Epoch 29/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3692 - mae: 0.2332 - val_loss: 0.3719 - val_mae: 0.2368\n",
      "Epoch 30/500\n",
      "611312/611312 [==============================] - 15s 24us/sample - loss: 0.3692 - mae: 0.2332 - val_loss: 0.3721 - val_mae: 0.2306\n",
      "Epoch 31/500\n",
      "611312/611312 [==============================] - 15s 24us/sample - loss: 0.3691 - mae: 0.2329 - val_loss: 0.3718 - val_mae: 0.2368\n",
      "Epoch 32/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3690 - mae: 0.2329 - val_loss: 0.3720 - val_mae: 0.2326\n",
      "Epoch 33/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3690 - mae: 0.2327 - val_loss: 0.3718 - val_mae: 0.2359\n",
      "Epoch 34/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3689 - mae: 0.2327 - val_loss: 0.3717 - val_mae: 0.2370\n",
      "Epoch 35/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3688 - mae: 0.2325 - val_loss: 0.3716 - val_mae: 0.2341\n",
      "Epoch 36/500\n",
      "611312/611312 [==============================] - 15s 24us/sample - loss: 0.3688 - mae: 0.2322 - val_loss: 0.3718 - val_mae: 0.2301\n",
      "Epoch 37/500\n",
      "611312/611312 [==============================] - 15s 24us/sample - loss: 0.3687 - mae: 0.2332 - val_loss: 0.3715 - val_mae: 0.2336\n",
      "Epoch 38/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3687 - mae: 0.2325 - val_loss: 0.3715 - val_mae: 0.2353\n",
      "Epoch 39/500\n",
      "611312/611312 [==============================] - 17s 28us/sample - loss: 0.3686 - mae: 0.2327 - val_loss: 0.3716 - val_mae: 0.2304\n",
      "Epoch 40/500\n",
      "611312/611312 [==============================] - 16s 26us/sample - loss: 0.3685 - mae: 0.2327 - val_loss: 0.3713 - val_mae: 0.2343\n",
      "Epoch 41/500\n",
      "611312/611312 [==============================] - 16s 26us/sample - loss: 0.3686 - mae: 0.2326 - val_loss: 0.3715 - val_mae: 0.2358\n",
      "Epoch 42/500\n",
      "611312/611312 [==============================] - 17s 28us/sample - loss: 0.3685 - mae: 0.2333 - val_loss: 0.3713 - val_mae: 0.2330\n",
      "Epoch 43/500\n",
      "611312/611312 [==============================] - 16s 27us/sample - loss: 0.3684 - mae: 0.2321 - val_loss: 0.3713 - val_mae: 0.2339\n",
      "Epoch 44/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3684 - mae: 0.2327 - val_loss: 0.3714 - val_mae: 0.2345\n",
      "Epoch 45/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3684 - mae: 0.2332 - val_loss: 0.3712 - val_mae: 0.2324\n",
      "Epoch 46/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3683 - mae: 0.2322 - val_loss: 0.3713 - val_mae: 0.2315\n",
      "Epoch 47/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3683 - mae: 0.2319 - val_loss: 0.3712 - val_mae: 0.2321\n",
      "Epoch 48/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3682 - mae: 0.2330 - val_loss: 0.3711 - val_mae: 0.2364\n",
      "Epoch 49/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3682 - mae: 0.2325 - val_loss: 0.3711 - val_mae: 0.2318\n",
      "Epoch 50/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3681 - mae: 0.2332 - val_loss: 0.3711 - val_mae: 0.2326\n",
      "Epoch 51/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3681 - mae: 0.2318 - val_loss: 0.3712 - val_mae: 0.2317\n",
      "Epoch 52/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3680 - mae: 0.2323 - val_loss: 0.3712 - val_mae: 0.2322\n",
      "Epoch 53/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3680 - mae: 0.2315 - val_loss: 0.3708 - val_mae: 0.2337\n",
      "Epoch 54/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3680 - mae: 0.2325 - val_loss: 0.3709 - val_mae: 0.2334\n",
      "Epoch 55/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3679 - mae: 0.2321 - val_loss: 0.3707 - val_mae: 0.2374\n",
      "Epoch 56/500\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3679 - mae: 0.2328 - val_loss: 0.3708 - val_mae: 0.2340\n",
      "Epoch 57/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3679 - mae: 0.2323 - val_loss: 0.3708 - val_mae: 0.2349\n",
      "Epoch 58/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3678 - mae: 0.2322 - val_loss: 0.3708 - val_mae: 0.2328\n",
      "Epoch 59/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3678 - mae: 0.2321 - val_loss: 0.3711 - val_mae: 0.2292\n",
      "Epoch 60/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3678 - mae: 0.2321 - val_loss: 0.3708 - val_mae: 0.2341\n",
      "Epoch 61/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3678 - mae: 0.2322 - val_loss: 0.3707 - val_mae: 0.2386\n",
      "Epoch 62/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3677 - mae: 0.2327 - val_loss: 0.3707 - val_mae: 0.2349\n",
      "Epoch 63/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3677 - mae: 0.2322 - val_loss: 0.3707 - val_mae: 0.2342\n",
      "Epoch 64/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3677 - mae: 0.2320 - val_loss: 0.3710 - val_mae: 0.2311\n",
      "Epoch 65/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3676 - mae: 0.2319 - val_loss: 0.3709 - val_mae: 0.2370\n",
      "Epoch 66/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3676 - mae: 0.2325 - val_loss: 0.3708 - val_mae: 0.2319\n",
      "Epoch 67/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3676 - mae: 0.2317 - val_loss: 0.3707 - val_mae: 0.2322\n",
      "Epoch 68/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3676 - mae: 0.2322 - val_loss: 0.3705 - val_mae: 0.2362\n",
      "Epoch 69/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3675 - mae: 0.2312 - val_loss: 0.3707 - val_mae: 0.2311\n",
      "Epoch 70/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3675 - mae: 0.2324 - val_loss: 0.3705 - val_mae: 0.2343\n",
      "Epoch 71/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3675 - mae: 0.2328 - val_loss: 0.3706 - val_mae: 0.2348\n",
      "Epoch 72/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3675 - mae: 0.2333 - val_loss: 0.3706 - val_mae: 0.2347\n",
      "Epoch 73/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3674 - mae: 0.2327 - val_loss: 0.3711 - val_mae: 0.2257\n",
      "Epoch 74/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3674 - mae: 0.2322 - val_loss: 0.3705 - val_mae: 0.2315\n",
      "Epoch 75/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3674 - mae: 0.2320 - val_loss: 0.3713 - val_mae: 0.2283\n",
      "Epoch 76/500\n",
      "611312/611312 [==============================] - 16s 25us/sample - loss: 0.3674 - mae: 0.2326 - val_loss: 0.3707 - val_mae: 0.2322\n",
      "Epoch 77/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3673 - mae: 0.2324 - val_loss: 0.3709 - val_mae: 0.2380\n",
      "Epoch 78/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3673 - mae: 0.2318 - val_loss: 0.3705 - val_mae: 0.2349\n",
      "Epoch 79/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3673 - mae: 0.2324 - val_loss: 0.3705 - val_mae: 0.2320\n",
      "Epoch 80/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3673 - mae: 0.2324 - val_loss: 0.3707 - val_mae: 0.2313\n",
      "Epoch 81/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3673 - mae: 0.2323 - val_loss: 0.3704 - val_mae: 0.2335\n",
      "Epoch 82/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3672 - mae: 0.2335 - val_loss: 0.3705 - val_mae: 0.2300\n",
      "Epoch 83/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3672 - mae: 0.2314 - val_loss: 0.3703 - val_mae: 0.2316\n",
      "Epoch 84/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3672 - mae: 0.2322 - val_loss: 0.3705 - val_mae: 0.2316\n",
      "Epoch 85/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3672 - mae: 0.2323 - val_loss: 0.3705 - val_mae: 0.2328\n",
      "Epoch 86/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3672 - mae: 0.2319 - val_loss: 0.3707 - val_mae: 0.2359\n",
      "Epoch 87/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3672 - mae: 0.2326 - val_loss: 0.3704 - val_mae: 0.2343\n",
      "Epoch 88/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3671 - mae: 0.2326 - val_loss: 0.3706 - val_mae: 0.2326\n",
      "Epoch 89/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3671 - mae: 0.2315 - val_loss: 0.3704 - val_mae: 0.2291\n",
      "Epoch 90/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3671 - mae: 0.2318 - val_loss: 0.3704 - val_mae: 0.2319\n",
      "Epoch 91/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3671 - mae: 0.2316 - val_loss: 0.3704 - val_mae: 0.2357\n",
      "Epoch 92/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3670 - mae: 0.2325 - val_loss: 0.3706 - val_mae: 0.2300\n",
      "Epoch 93/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3670 - mae: 0.2318 - val_loss: 0.3703 - val_mae: 0.2325\n",
      "Epoch 94/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3670 - mae: 0.2320 - val_loss: 0.3704 - val_mae: 0.2341\n",
      "Epoch 95/500\n",
      "611312/611312 [==============================] - 16s 25us/sample - loss: 0.3670 - mae: 0.2318 - val_loss: 0.3704 - val_mae: 0.2330\n",
      "Epoch 96/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3670 - mae: 0.2315 - val_loss: 0.3708 - val_mae: 0.2270\n",
      "Epoch 97/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3670 - mae: 0.2319 - val_loss: 0.3703 - val_mae: 0.2314\n",
      "Epoch 98/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3670 - mae: 0.2320 - val_loss: 0.3704 - val_mae: 0.2356\n",
      "Epoch 99/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3669 - mae: 0.2323 - val_loss: 0.3704 - val_mae: 0.2324\n",
      "Epoch 100/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3669 - mae: 0.2320 - val_loss: 0.3701 - val_mae: 0.2339\n",
      "Epoch 101/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3669 - mae: 0.2318 - val_loss: 0.3702 - val_mae: 0.2337\n",
      "Epoch 102/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3669 - mae: 0.2315 - val_loss: 0.3705 - val_mae: 0.2366\n",
      "Epoch 103/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3669 - mae: 0.2319 - val_loss: 0.3704 - val_mae: 0.2312\n",
      "Epoch 104/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3669 - mae: 0.2317 - val_loss: 0.3702 - val_mae: 0.2365\n",
      "Epoch 105/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3669 - mae: 0.2321 - val_loss: 0.3701 - val_mae: 0.2338\n",
      "Epoch 106/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3668 - mae: 0.2318 - val_loss: 0.3703 - val_mae: 0.2318\n",
      "Epoch 107/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3668 - mae: 0.2315 - val_loss: 0.3702 - val_mae: 0.2324\n",
      "Epoch 108/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3668 - mae: 0.2317 - val_loss: 0.3701 - val_mae: 0.2359\n",
      "Epoch 109/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3668 - mae: 0.2318 - val_loss: 0.3702 - val_mae: 0.2347\n",
      "Epoch 110/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3668 - mae: 0.2312 - val_loss: 0.3702 - val_mae: 0.2337\n",
      "Epoch 111/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3668 - mae: 0.2317 - val_loss: 0.3702 - val_mae: 0.2362\n",
      "Epoch 112/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3668 - mae: 0.2323 - val_loss: 0.3702 - val_mae: 0.2371\n"
     ]
    },
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Epoch 113/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3667 - mae: 0.2324 - val_loss: 0.3702 - val_mae: 0.2319\n",
      "Epoch 114/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3668 - mae: 0.2322 - val_loss: 0.3707 - val_mae: 0.2296\n",
      "Epoch 115/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3667 - mae: 0.2316 - val_loss: 0.3702 - val_mae: 0.2304\n",
      "Epoch 116/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3667 - mae: 0.2316 - val_loss: 0.3705 - val_mae: 0.2263\n",
      "Epoch 117/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3667 - mae: 0.2318 - val_loss: 0.3700 - val_mae: 0.2331\n",
      "Epoch 118/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3667 - mae: 0.2318 - val_loss: 0.3701 - val_mae: 0.2327\n",
      "Epoch 119/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3667 - mae: 0.2319 - val_loss: 0.3701 - val_mae: 0.2366\n",
      "Epoch 120/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3667 - mae: 0.2317 - val_loss: 0.3702 - val_mae: 0.2328\n",
      "Epoch 121/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3667 - mae: 0.2316 - val_loss: 0.3701 - val_mae: 0.2323\n",
      "Epoch 122/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3667 - mae: 0.2311 - val_loss: 0.3702 - val_mae: 0.2370\n",
      "Epoch 123/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3667 - mae: 0.2314 - val_loss: 0.3701 - val_mae: 0.2314\n",
      "Epoch 124/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3666 - mae: 0.2318 - val_loss: 0.3704 - val_mae: 0.2289\n",
      "Epoch 125/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3666 - mae: 0.2327 - val_loss: 0.3703 - val_mae: 0.2357\n",
      "Epoch 126/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3666 - mae: 0.2325 - val_loss: 0.3702 - val_mae: 0.2354\n",
      "Epoch 127/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3666 - mae: 0.2325 - val_loss: 0.3702 - val_mae: 0.2293\n",
      "Epoch 128/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3666 - mae: 0.2317 - val_loss: 0.3701 - val_mae: 0.2372\n",
      "Epoch 129/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3666 - mae: 0.2320 - val_loss: 0.3700 - val_mae: 0.2332\n",
      "Epoch 130/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3666 - mae: 0.2317 - val_loss: 0.3701 - val_mae: 0.2310\n",
      "Epoch 131/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3666 - mae: 0.2325 - val_loss: 0.3701 - val_mae: 0.2334\n",
      "Epoch 132/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3666 - mae: 0.2318 - val_loss: 0.3705 - val_mae: 0.2276\n",
      "Epoch 133/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3666 - mae: 0.2326 - val_loss: 0.3701 - val_mae: 0.2336\n",
      "Epoch 134/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3666 - mae: 0.2326 - val_loss: 0.3700 - val_mae: 0.2328\n",
      "Epoch 135/500\n",
      "611312/611312 [==============================] - 16s 26us/sample - loss: 0.3666 - mae: 0.2316 - val_loss: 0.3700 - val_mae: 0.2396\n",
      "Epoch 136/500\n",
      "611312/611312 [==============================] - 16s 25us/sample - loss: 0.3666 - mae: 0.2325 - val_loss: 0.3701 - val_mae: 0.2399loss: 0.3665 - mae: 0.2\n",
      "Epoch 137/500\n",
      "611312/611312 [==============================] - 1818s 3ms/sample - loss: 0.3666 - mae: 0.2310 - val_loss: 0.3704 - val_mae: 0.2389\n",
      "Epoch 138/500\n",
      "611312/611312 [==============================] - 2905s 5ms/sample - loss: 0.3665 - mae: 0.2316 - val_loss: 0.3703 - val_mae: 0.2299\n",
      "Epoch 139/500\n",
      "611312/611312 [==============================] - 1852s 3ms/sample - loss: 0.3665 - mae: 0.2317 - val_loss: 0.3699 - val_mae: 0.2329\n",
      "Epoch 140/500\n",
      "611312/611312 [==============================] - 3748s 6ms/sample - loss: 0.3665 - mae: 0.2317 - val_loss: 0.3703 - val_mae: 0.2361\n",
      "Epoch 141/500\n",
      "611312/611312 [==============================] - 7251s 12ms/sample - loss: 0.3665 - mae: 0.2323 - val_loss: 0.3701 - val_mae: 0.2324\n",
      "Epoch 142/500\n",
      "611312/611312 [==============================] - 2242s 4ms/sample - loss: 0.3665 - mae: 0.2321 - val_loss: 0.3700 - val_mae: 0.2353\n",
      "Epoch 143/500\n",
      "611312/611312 [==============================] - 7251s 12ms/sample - loss: 0.3665 - mae: 0.2313 - val_loss: 0.3699 - val_mae: 0.2378\n",
      "Epoch 144/500\n",
      "611312/611312 [==============================] - 7251s 12ms/sample - loss: 0.3665 - mae: 0.2320 - val_loss: 0.3700 - val_mae: 0.2328\n",
      "Epoch 145/500\n",
      "611312/611312 [==============================] - 134s 220us/sample - loss: 0.3665 - mae: 0.2323 - val_loss: 0.3699 - val_mae: 0.2344\n",
      "Epoch 146/500\n",
      "611312/611312 [==============================] - 63s 103us/sample - loss: 0.3665 - mae: 0.2318 - val_loss: 0.3699 - val_mae: 0.2362\n",
      "Epoch 147/500\n",
      "611312/611312 [==============================] - 18s 29us/sample - loss: 0.3665 - mae: 0.2317 - val_loss: 0.3699 - val_mae: 0.2341\n",
      "Epoch 148/500\n",
      "611312/611312 [==============================] - 18s 30us/sample - loss: 0.3665 - mae: 0.2320 - val_loss: 0.3699 - val_mae: 0.2337\n",
      "Epoch 149/500\n",
      "611312/611312 [==============================] - 19s 31us/sample - loss: 0.3665 - mae: 0.2323 - val_loss: 0.3701 - val_mae: 0.2316\n",
      "Epoch 150/500\n",
      "611312/611312 [==============================] - 21s 34us/sample - loss: 0.3664 - mae: 0.2322 - val_loss: 0.3699 - val_mae: 0.2355\n",
      "Epoch 151/500\n",
      "611312/611312 [==============================] - 20s 33us/sample - loss: 0.3665 - mae: 0.2327 - val_loss: 0.3700 - val_mae: 0.2348\n",
      "Epoch 152/500\n",
      "611312/611312 [==============================] - 19s 31us/sample - loss: 0.3664 - mae: 0.2313 - val_loss: 0.3700 - val_mae: 0.2314\n",
      "Epoch 153/500\n",
      "611312/611312 [==============================] - 15s 24us/sample - loss: 0.3664 - mae: 0.2316 - val_loss: 0.3700 - val_mae: 0.2332\n",
      "Epoch 154/500\n",
      "611312/611312 [==============================] - 15s 24us/sample - loss: 0.3665 - mae: 0.2317 - val_loss: 0.3702 - val_mae: 0.2358\n",
      "Epoch 155/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3664 - mae: 0.2322 - val_loss: 0.3701 - val_mae: 0.2347\n",
      "Epoch 156/500\n",
      "611312/611312 [==============================] - 16s 26us/sample - loss: 0.3664 - mae: 0.2315 - val_loss: 0.3707 - val_mae: 0.2262\n",
      "Epoch 157/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3664 - mae: 0.2325 - val_loss: 0.3699 - val_mae: 0.2352\n",
      "Epoch 158/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3664 - mae: 0.2319 - val_loss: 0.3699 - val_mae: 0.2342\n",
      "Epoch 159/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3664 - mae: 0.2315 - val_loss: 0.3701 - val_mae: 0.2313\n",
      "Epoch 160/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3664 - mae: 0.2320 - val_loss: 0.3703 - val_mae: 0.2303\n",
      "Epoch 161/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3664 - mae: 0.2319 - val_loss: 0.3702 - val_mae: 0.2353\n",
      "Epoch 162/500\n",
      "611312/611312 [==============================] - 15s 25us/sample - loss: 0.3664 - mae: 0.2320 - val_loss: 0.3701 - val_mae: 0.2358\n",
      "Epoch 00162: early stopping\n"
     ]
    }
   ],
   "source": [
    "\n",
    "test_size = 50000\n",
    "C_val = C_test[:test_size]\n",
    "D_val = D_test[:test_size]\n",
    "deal_val = deal_test[:test_size]\n",
    "p_val = p_test[:test_size]\n",
    "print(\"true logloss train set: \", np.mean(logloss(deal_train,p_train)))\n",
    "print(\"true logloss validation set: \", np.mean(logloss(deal_val,p_val)))\n",
    "M.train(C_train, D_train, deal_train, C_val, D_val, deal_val, batch_size=100, verbose=1, early_stopping_patience=15)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Validation score (logloss) ´"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 56,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "true logloss validation set:  0.3669\n",
      "ML model logloss validation set:  0.3699\n"
     ]
    }
   ],
   "source": [
    "pred_p_val = M.probability(C_val,D_val)\n",
    "print(\"true logloss validation set: \", np.round(np.mean(logloss(deal_val,p_val)),4))\n",
    "print(\"ML model logloss validation set: \", np.round(np.mean(logloss(deal_val,pred_p_val)),4))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {
    "scrolled": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Generated data:  101 / 101\n",
      "Generated data:  101 / 101\n",
      "Generated data:  101 / 101\n"
     ]
    },
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "nr=101\n",
    "dl = np.linspace(0,0.4,nr)\n",
    "\n",
    "C_1 = one_hot([1,1,1,1,1,1])\n",
    "C_1_list = np.tile(np.expand_dims(np.array(C_1),0),(nr,1))\n",
    "pred_p1 = M.probability(C_1_list, dl)\n",
    "true_p1 = G.P.p(C_1_list,dl)\n",
    "\n",
    "plt.plot(dl,true_p1, c='b', label='true p(C_1,D)')\n",
    "plt.plot(dl,pred_p1, c='b', ls='--', label='pred p(C_1,D)')\n",
    "\n",
    "C_2 = [3,2,0,2,6,2]\n",
    "C_2oh = one_hot(C_2)\n",
    "C_2_list = np.tile(np.expand_dims(np.array(C_2oh),0),(nr,1))\n",
    "pred_p2 =  M.probability(C_2_list, dl)\n",
    "true_p2 = G.P.p(C_2_list,dl)\n",
    "\n",
    "plt.plot(dl,true_p2, c='r', label='true p(C_2,D)')\n",
    "plt.plot(dl,pred_p2, c='r', ls='-.', label='pred p(C_2,D)')\n",
    "\n",
    "\n",
    "C_3 = [0,2,0,1,2,2]\n",
    "C_3oh = one_hot(C_3)\n",
    "C_3_list = np.tile(np.expand_dims(np.array(C_3oh),0),(nr,1))\n",
    "pred_p3 =  M.probability(C_3_list, dl)\n",
    "true_p3 = G.P.p(C_3_list,dl)\n",
    "\n",
    "plt.plot(dl,true_p3, c='g', label='true p(C_3,D)')\n",
    "plt.plot(dl,pred_p3, c='g', ls='--', label='pred p(C_3,D)')\n",
    "\n",
    "plt.xlabel('Discount D')\n",
    "plt.ylabel('Probability')\n",
    "plt.legend()\n",
    "plt.savefig(\"plotswith\")\n",
    "plt.show()\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "\n",
    "print(\"np.expand_dims(D_train,1) shape: \", (np.expand_dims(D_train,1)).shape)\n",
    "print(\"C_train shape: \", C_train.shape)\n",
    "#p_pred_train = np.array([model([np.expand_dims(d,0), np.expand_dims(c,0)]).mean() for d,c in zip(np.expand_dims(D_train,1), C_train)])\n",
    "p_pred_train = np.squeeze(model([np.expand_dims(D_train,1), C_train]).mean())\n",
    "print(\"p_pred_train shape: \", p_pred_train.shape)\n",
    "pred_score = logloss(deal_train,p_pred_train)\n",
    "print(\"pred mean: \", np.mean(pred_score))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "score = logloss(deal_train,p_train)\n",
    "print(\"true log loss: \", np.mean(score))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "nr=101\n",
    "dl = np.linspace(0,0.4,nr)\n",
    "C_1 = one_hot([1,1,1,1,1,1])\n",
    "C_1_list = np.tile(np.expand_dims(np.array(C_1),0),(nr,1))\n",
    "p_pred = model([np.expand_dims(dl,1),C_1_list]).mean()\n",
    "p_pred = np.squeeze(p_pred)\n",
    "true_C_1_p = lambda d: np.float64(G.P.base_func(to_scalar([1,1,1,1,1,1]), d))\n",
    "true_p = true_C_1_p(dl)\n",
    "p1,deal1 = G.P.sample(C_1_list,dl)\n",
    "score = logloss(deal1,p1)\n",
    "print(\"true log loss: \", np.mean(score))\n",
    "score = logloss(deal1,p_pred)\n",
    "print(\"pred log loss: \", np.mean(score))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "def logloss_test(C, G, model):\n",
    "    nr=101\n",
    "    dl = np.linspace(0,0.4,nr)\n",
    "    C_oh = one_hot(C)\n",
    "    C_oh_list = np.tile(np.expand_dims(np.array(C_oh),0),(nr,1))\n",
    "    p_pred = model([np.expand_dims(dl,1),C_oh_list]).mean()\n",
    "    p_pred = np.squeeze(p_pred)\n",
    "    true_C_oh_p = lambda d: np.float64(G.P.base_func(to_scalar(C), d))\n",
    "    true_p = true_C_oh_p(dl)\n",
    "    p1,deal1 = G.P.sample(C_oh_list,dl)\n",
    "    true_score = np.mean( logloss(deal1,p1))\n",
    "    print(\"true log loss: \", true_score)\n",
    "    pred_score = np.mean(logloss(deal1,p_pred))\n",
    "    print(\"pred log loss: \", pred_score)\n",
    "    print(\"fraction (pred_logloss/true_logloss): \", pred_score/true_score)\n",
    "    return true_score, pred_score\n",
    "    "
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "logloss_test([0,0,0,0,0,0], G, model)"
   ]
  },
  {
   "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
}