03ST_TL.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Combination of Spatial Transformer and VGG16\n",
    "Download the VGG16 weights to the folder \"checkpoints\" from this source:\n",
    "\n",
    "https://github.com/tensorflow/models/tree/master/research/slim\n",
    "\n",
    "The data should be stored as follows\n",
    "\n",
    ".../train<pre>\n",
    "/pso\n",
    "/ecz\n",
    "</pre>\n",
    ".../validation <pre>\n",
    "/pso  \n",
    "/ecz \n",
    "</pre>\n",
    ".../test<pre>\n",
    "/pso  \n",
    "/ecz  \n",
    "</pre>\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "/home/nbip/.conda/envs/tf/lib/python3.6/importlib/_bootstrap.py:219: RuntimeWarning: compiletime version 3.5 of module 'tensorflow.python.framework.fast_tensor_util' does not match runtime version 3.6\n",
      "  return f(*args, **kwds)\n"
     ]
    }
   ],
   "source": [
    "# for specifying which GPU to use\n",
    "import os\n",
    "os.environ[\"CUDA_DEVICE_ORDER\"]=\"PCI_BUS_ID\"\n",
    "os.environ[\"CUDA_VISIBLE_DEVICES\"]=\"4\"\n",
    "\n",
    "import tensorflow as tf"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using TensorFlow backend.\n"
     ]
    }
   ],
   "source": [
    "import glob\n",
    "%matplotlib inline\n",
    "import numpy as np\n",
    "import matplotlib.pyplot as plt\n",
    "from keras.preprocessing.image import load_img, img_to_array\n",
    "from skimage.transform import resize\n",
    "from random import shuffle\n",
    "import pickle\n",
    "import datetime\n",
    "from keras.preprocessing.image import ImageDataGenerator\n",
    "from keras.applications.vgg16 import preprocess_input"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Visualize the input"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "DIM = 224\n",
    "nchannels = 3\n",
    "img_height, img_width = DIM, DIM\n",
    "NUM_CLASSES = 2\n",
    "NUM_EPOCHS = 50\n",
    "BATCH_SIZE = 100 # batch size of 128 is too big when training the full model (the FC or STN alone is okay)\n",
    "# channel means subtracted in the prerocessing step can be found here \n",
    "# https://github.com/tensorflow/models/blob/master/research/slim/preprocessing/vgg_preprocessing.py"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "checkpoints_dir =          # path to the VGG16 weights\n",
    "train_data_dir =           # path to the train data directory\n",
    "validation_data_dir =      # path to the validation data directory\n",
    "test_data_dir =            # path to the test data directory\n",
    "tensorboard_dir =          # path to the tensorboard directory\n",
    "\n",
    "train_pso = glob.glob(train_data_dir + \"/pso/*.png\")\n",
    "train_ecz = glob.glob(train_data_dir + \"/ecz/*.png\")\n",
    "val_pso = glob.glob(validation_data_dir + \"/pso/*.png\")\n",
    "val_ecz = glob.glob(validation_data_dir + \"/ecz/*.png\")"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 11,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Found 40000 images belonging to 2 classes.\n",
      "Found 4000 images belonging to 2 classes.\n"
     ]
    }
   ],
   "source": [
    "datagen = ImageDataGenerator(\n",
    "    rescale=1. / 255,\n",
    "    preprocessing_function = preprocess_input\n",
    ")\n",
    "\n",
    "train_generator = datagen.flow_from_directory(\n",
    "        train_data_dir,  \n",
    "        target_size=(img_height, img_width),  \n",
    "        batch_size=BATCH_SIZE,\n",
    "        class_mode='categorical')  \n",
    "\n",
    "validation_generator = datagen.flow_from_directory(\n",
    "        validation_data_dir,\n",
    "        target_size=(img_height, img_width),\n",
    "        batch_size=BATCH_SIZE,\n",
    "        class_mode='categorical')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Visualize the input"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "n = 4\n",
    "dummy_batch_size = n*n\n",
    "\n",
    "x_batch, y_batch = train_generator.next()\n",
    "\n",
    "idx = 0\n",
    "canvas = np.zeros((DIM*n, n*DIM, nchannels))\n",
    "for i in range(n):\n",
    "    for j in range(n):\n",
    "        img = x_batch[idx]\n",
    "        mi = np.min(img)\n",
    "        img = img - mi\n",
    "        ma = np.max(img)\n",
    "        img = img / ma\n",
    "        canvas[i*DIM:(i+1)*DIM, j*DIM:(j+1)*DIM] = img\n",
    "        idx += 1\n",
    "plt.figure(figsize=(10, 10))\n",
    "plt.imshow(canvas, cmap='gray')\n",
    "plt.title('canvas of {0}x{0} images'.format(DIM))\n",
    "plt.axis('off')\n",
    "\n",
    "plt.show()\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "# Dependencies and supporting functions"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "import os\n",
    "import matplotlib\n",
    "np.random.seed(123)\n",
    "import tensorflow as tf\n",
    "\n",
    "import os\n",
    "import sys\n",
    "sys.path.append(os.path.join('.', '..')) \n",
    "from spatial_transformer import transformer\n",
    "\n",
    "from tensorflow.contrib.layers import fully_connected, convolution2d, flatten, max_pool2d, dropout\n",
    "pool = max_pool2d\n",
    "conv = convolution2d\n",
    "dense = fully_connected\n",
    "from tensorflow.python.ops.nn import relu, softmax\n",
    "from tensorflow.python.framework.ops import reset_default_graph\n",
    "from tensorflow.contrib import slim\n",
    "from tensorflow.contrib.slim.nets import vgg\n",
    "slim = tf.contrib.slim"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Build the Spatial Transformer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def build_STN_model(x_pl, input_width, input_height, output_dim):\n",
    "    \n",
    "    # set up weights for transformation (notice we always need 6 output neurons)\n",
    "    with tf.name_scope(\"localization\"):\n",
    "        \n",
    "        # convert from RGB to HSV\n",
    "        hsv_image = tf.image.rgb_to_hsv(x_pl)\n",
    "    \n",
    "        # make distributed representation of input image for localization network\n",
    "        loc_l1 = pool(hsv_image, kernel_size=[2, 2], scope=\"localization_l1\")\n",
    "        loc_l2 = conv(loc_l1, num_outputs=8, kernel_size=[5, 5], stride=[1, 1], padding=\"SAME\", scope=\"localization_l2\")\n",
    "        loc_l3 = pool(loc_l2, kernel_size=[2, 2], scope=\"localization_l3\")\n",
    "        loc_l4 = conv(loc_l3, num_outputs=16, kernel_size=[5, 5], stride=[1, 1], padding=\"SAME\", scope=\"localization_l4\")\n",
    "        loc_l4_flatten = flatten(loc_l4, scope=\"localization_l4-flatten\")\n",
    "        loc_l5 = dense(loc_l4_flatten, num_outputs=50, activation_fn=relu, scope=\"localization_l5\")\n",
    "\n",
    "\n",
    "        W_loc_out = tf.get_variable(\"localization_loc-out\", [50, 6], initializer=tf.constant_initializer(0.0))\n",
    "        initial = np.array([[1, 0, 0], [0, 1, 0]])\n",
    "        initial = initial.astype('float32')\n",
    "        initial = initial.flatten()\n",
    "        b_loc_out = tf.Variable(initial_value=initial, name='b-loc-out')\n",
    "        loc_out = tf.matmul(loc_l5, W_loc_out) + b_loc_out\n",
    "\n",
    "    # spatial transformer\n",
    "    l_trans1 = transformer(x_pl, loc_out, out_size=(DIM, DIM))\n",
    "    l_trans1.set_shape([None, DIM, DIM, nchannels])\n",
    "\n",
    "    print( \"Transformer network output shape: \", l_trans1.get_shape())\n",
    "\n",
    "    return l_trans1, loc_out, hsv_image"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Build the VGG16 model"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 21,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def build_VGG_model(inputs, is_training=True):\n",
    "    inputs = tf.cast(inputs, tf.float32)\n",
    "    #Use Pretrained Base Model\n",
    "    with tf.variable_scope(\"vgg_16\"):\n",
    "        with slim.arg_scope(vgg.vgg_arg_scope()):\n",
    "            net = slim.repeat(inputs, 2, slim.conv2d, 64, [3, 3], scope='conv1')\n",
    "            net = slim.max_pool2d(net, [2, 2], scope='pool1')\n",
    "            net = slim.repeat(net, 2, slim.conv2d, 128, [3, 3], scope='conv2')\n",
    "            net2 = slim.max_pool2d(net, [2, 2], scope='pool2')\n",
    "            net = slim.repeat(net2, 3, slim.conv2d, 256, [3, 3], scope='conv3')\n",
    "            net = slim.max_pool2d(net, [2, 2], scope='pool3')\n",
    "            net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv4')\n",
    "            net = slim.max_pool2d(net, [2, 2], scope='pool4')\n",
    "            net = slim.repeat(net, 3, slim.conv2d, 512, [3, 3], scope='conv5')\n",
    "            net = slim.max_pool2d(net, [2, 2], scope='pool5')\n",
    "    #Append fully connected layer\n",
    "    fc = slim.flatten(net)\n",
    "    fc = slim.fully_connected(fc, 256,\n",
    "            weights_initializer=tf.contrib.layers.xavier_initializer(),\n",
    "            weights_regularizer=slim.l2_regularizer(0.5),\n",
    "            scope='finetune/fc1')\n",
    "    fc = slim.dropout(fc, 0.5)\n",
    "    fc_out= slim.fully_connected(fc, 2,\n",
    "            activation_fn=tf.nn.softmax,\n",
    "            weights_initializer=tf.contrib.layers.xavier_initializer(),\n",
    "            weights_regularizer=slim.l2_regularizer(0.5),\n",
    "            scope='finetune/fc2')\n",
    "    print( \"VGG model output shape: \", net.get_shape())\n",
    "    return net, fc_out, net2\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "def loss_and_acc(logits, labels):\n",
    "    with tf.name_scope(\"loss\"):\n",
    "        # weight the logits by the class imbalance\n",
    "        # computing cross entropy per sample\n",
    "        cross_entropy = -tf.reduce_sum(labels * tf.log(logits+1e-8), reduction_indices=[1])\n",
    "\n",
    "        # averaging over samples\n",
    "        cross_entropy = tf.reduce_mean(cross_entropy)\n",
    "        \n",
    "        # accuracy\n",
    "        argmax_logits = tf.to_int32(tf.argmax(logits, axis=1))\n",
    "        argmax_labels = tf.to_int32(tf.argmax(labels, axis=1))\n",
    "        correct = tf.to_float(tf.equal(argmax_logits, argmax_labels))\n",
    "        accuracy = tf.reduce_mean(correct)\n",
    "        \n",
    "        # precision\n",
    "        prec = tf.metrics.precision(argmax_labels, argmax_logits)\n",
    "        \n",
    "        # recall\n",
    "        rec = tf.metrics.recall(argmax_labels, argmax_logits)\n",
    "        \n",
    "        # AUC of the ROC\n",
    "        auc, update_op_auc = tf.metrics.auc(labels, logits)\n",
    "\n",
    "\n",
    "    return cross_entropy, accuracy, argmax_logits, prec, rec, update_op_auc\n",
    "        \n",
    "def optimize(losses, lr_pl, trainable_vars, g_step):\n",
    "    \n",
    "    # defining our optimizer\n",
    "    optimizer = tf.train.AdamOptimizer(learning_rate=lr_pl)\n",
    "\n",
    "    # applying the gradients\n",
    "    train_op = optimizer.minimize(losses, var_list=trainable_vars, global_step=g_step)\n",
    "    \n",
    "    return train_op"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Stitch it together"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 23,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Transformer network output shape:  (?, 224, 224, 3)\n",
      "VGG model output shape:  (?, 7, 7, 512)\n"
     ]
    },
    {
     "data": {
      "text/plain": [
       "<tf.Tensor 'Evaluation/accuracy:0' shape=() dtype=string>"
      ]
     },
     "execution_count": 23,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "tf.reset_default_graph()\n",
    "\n",
    "# input images\n",
    "x_pl = tf.placeholder(tf.float32, [None, DIM, DIM, nchannels], name=\"input\")\n",
    "# corresponding labels as onehot encoding\n",
    "y_pl = tf.placeholder(tf.float32, shape=[None, NUM_CLASSES], name=\"output\")\n",
    "# learning rate placeholder\n",
    "lr_pl = tf.placeholder(tf.float32, shape=[], name=\"learning-rate\")\n",
    "# keep track of training steps. Incremements by 1 each time tf.optimizer.minimize is called\n",
    "global_step = tf.Variable(initial_value=0, trainable=False)  \n",
    "\n",
    "# the result of the transform, the weights used for the transform and the HSV image\n",
    "l_transform, w_transform, hsv = build_STN_model(x_pl, DIM, DIM, NUM_CLASSES)\n",
    "\n",
    "# The VGG16 model output and the logits from the fully connected layer on top\n",
    "vgg16_out, logits, vgg16_mid = build_VGG_model(l_transform)\n",
    "\n",
    "# define losses and training operation\n",
    "loss, accuracy, prediction, precision, recall, auc_op = loss_and_acc(logits, y_pl)\n",
    "\n",
    "# training fully connected layers \n",
    "trainable_vars1 = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,\"finetune\")\n",
    "train_op1 = optimize(loss, lr_pl, trainable_vars1, global_step)\n",
    "\n",
    "# training fc + fine tuning VGG16 top layers\n",
    "trainable_vars2 = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,\"vgg_16/conv5\")\n",
    "trainable_vars2.extend(trainable_vars1)\n",
    "train_op2 = optimize(loss, lr_pl, trainable_vars2, global_step)\n",
    "\n",
    "# training fc + STN\n",
    "trainable_vars3 = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,\"localization\")\n",
    "trainable_vars3.extend(trainable_vars1)\n",
    "train_op3 = optimize(loss, lr_pl, trainable_vars3, global_step)\n",
    "\n",
    "# training STN + finetune + fc\n",
    "trainable_vars4 = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,\"localization\")\n",
    "trainable_vars4.extend(trainable_vars2)\n",
    "train_op4 = optimize(loss, lr_pl, trainable_vars4, global_step)\n",
    "\n",
    "# training only STN\n",
    "trainable_vars5 = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES,\"localization\")\n",
    "train_op5 = optimize(loss, lr_pl, trainable_vars5, global_step)\n",
    "\n",
    "# add TensorBoard summaries for all variables\n",
    "tf.summary.merge_all()\n",
    "tf.summary.scalar('Evaluation/loss', loss)\n",
    "tf.summary.scalar('Evaluation/accuracy', accuracy)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 24,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "INFO:tensorflow:Restoring parameters from ./tmp/checkpoints/vgg_16.ckpt\n"
     ]
    }
   ],
   "source": [
    "# restricting memory usage, TensorFlow is greedy and will use all memory otherwise\n",
    "gpu_opts = tf.GPUOptions(per_process_gpu_memory_fraction=1)\n",
    "\n",
    "# initialize the Session\n",
    "sess = tf.Session(config=tf.ConfigProto(gpu_options=gpu_opts))\n",
    "sess.run(tf.global_variables_initializer())\n",
    "sess.run(tf.local_variables_initializer())  # this is needed for the precision and recall for some reason\n",
    "\n",
    "# restorethe VGG_16 weights from file\n",
    "restore = slim.assign_from_checkpoint_fn(\n",
    "               os.path.join(checkpoints_dir, 'vgg_16.ckpt'),\n",
    "               slim.get_model_variables(\"vgg_16\"))\n",
    "restore(sess)\n",
    "\n",
    "# Tensorboard\n",
    "file_writer = tf.summary.FileWriter(tensorboard_dir, sess.graph)\n",
    "# Make run-options\n",
    "run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE,\n",
    "                            output_partition_graphs=True)\n",
    "run_metadata = tf.RunMetadata()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Test the forward pass"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 25,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Forward pass successful!\n"
     ]
    }
   ],
   "source": [
    "x_batch, y_batch = train_generator.next()\n",
    "res = sess.run(fetches=logits, feed_dict={x_pl: x_batch})\n",
    "assert res.shape == (len(x_batch), 2), \"ERROR the output shape is not as expected!\" \\\n",
    "        + \" Output shape should be \" + str(y_pl.shape) + ' but was ' + str(res.shape)\n",
    "\n",
    "print('Forward pass successful!')"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Do the training"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 29,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# setup and write summaries\n",
    "timestamp = datetime.datetime.now().strftime(\"%Y%m%d-%H%M%S\")\n",
    "summaries_path = \"tensorboard/%s\" % (timestamp)\n",
    "summaries = tf.summary.merge_all()\n",
    "\n",
    "summarywriter_train = tf.summary.FileWriter(summaries_path + '/train', sess.graph)\n",
    "summarywriter_valid = tf.summary.FileWriter(summaries_path + '/valid', sess.graph)\n",
    "\n",
    "def train_epoch(learning_rate, training_op):\n",
    "    num_samples = train_generator.n\n",
    "    num_batches = int(np.ceil(num_samples / float(BATCH_SIZE)))\n",
    "    costs = []\n",
    "    accs = []\n",
    "    precs = []\n",
    "    recs = []\n",
    "    for i in range(num_batches):\n",
    "        if i % 10 == 0:\n",
    "            print(i, end=', ')\n",
    "        X_batch_tr, y_batch_tr = train_generator.next()\n",
    "        fetches_tr = [training_op, loss, accuracy, prediction, precision, recall, summaries, global_step, l_transform, w_transform, auc_op]\n",
    "        feed_dict_tr = {x_pl: X_batch_tr, y_pl: y_batch_tr, lr_pl: learning_rate}\n",
    "        res = sess.run(fetches=fetches_tr, feed_dict=feed_dict_tr)\n",
    "        costs += [res[1]]\n",
    "        accs += [res[2]]\n",
    "        recs += [res[4]]\n",
    "        precs += [res[5]]\n",
    "        summarywriter_train.add_summary(res[6], res[7])\n",
    "    print(\"\")    \n",
    "    auc = res[10] # the auc_op is calculated a running estimate of the auc, therefore we do not have to keep track of this for each batch\n",
    "    sess.run(tf.local_variables_initializer())  # since the auc is a running estimate we need to reset it after each epoch\n",
    "    return np.mean(costs), np.mean(accs), np.mean(precs), np.mean(recs), summaries, auc\n",
    "\n",
    "\n",
    "def eval_epoch():\n",
    "    num_samples = validation_generator.n\n",
    "    num_batches = int(np.ceil(num_samples / float(BATCH_SIZE)))\n",
    "    costs = []\n",
    "    accs = []\n",
    "    preds = []\n",
    "    precs = []\n",
    "    recs = []\n",
    "    \n",
    "    pred_list = []\n",
    "    y_list = []\n",
    "    transform_list = []\n",
    "    weight_list = []\n",
    "    for i in range(num_batches):\n",
    "        if i % 10 == 0:\n",
    "            print(i, end=', ')\n",
    "        X_batch_val, y_batch_val = validation_generator.next()\n",
    "        fetches_val = [loss, accuracy, prediction, precision, recall, summaries, global_step, l_transform, w_transform, auc_op]\n",
    "        feed_dict_val = {x_pl: X_batch_val, y_pl: y_batch_val}\n",
    "        res = sess.run(fetches=fetches_val, feed_dict=feed_dict_val)\n",
    "        costs += [res[0]]\n",
    "        accs += [res[1]]\n",
    "        preds += [res[2]]\n",
    "        precs += [res[3]]\n",
    "        recs += [res[4]]\n",
    "        summarywriter_valid.add_summary(res[5], res[6])\n",
    "    preds_ = np.concatenate(preds, axis=0)\n",
    "    auc = res[9] # the auc_op is calculated a running estimate of the auc, therefore we do not have to keep track of this for each batch\n",
    "    sess.run(tf.local_variables_initializer())  # since the auc is a running estimate we need to reset it after each epoch\n",
    "    print('')\n",
    "    return np.mean(costs), np.mean(accs), preds_, np.mean(precs), np.mean(recs), summaries, auc"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 30,
   "metadata": {
    "collapsed": false,
    "scrolled": true
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Number of epochs: 100\n",
      "Epoch 0:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 1.2, train acc 0.51, val cost 0.69, val acc 0.54\n",
      "train auc 0.51, val auc 0.55\n",
      "train prec 0.76, train rec 0.5, val prec 0.54, val rec 0.52\n",
      "Epoch 1:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.67, train acc 0.58, val cost 0.66, val acc 0.6\n",
      "train auc 0.62, val auc 0.64\n",
      "train prec 0.58, train rec 0.57, val prec 0.6, val rec 0.54\n",
      "Epoch 2:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.63, train acc 0.64, val cost 0.67, val acc 0.62\n",
      "train auc 0.69, val auc 0.65\n",
      "train prec 0.63, train rec 0.62, val prec 0.68, val rec 0.42\n",
      "Epoch 3:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.6, train acc 0.67, val cost 0.64, val acc 0.64\n",
      "train auc 0.74, val auc 0.69\n",
      "train prec 0.69, train rec 0.67, val prec 0.67, val rec 0.52\n",
      "Epoch 4:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.56, train acc 0.71, val cost 0.67, val acc 0.64\n",
      "train auc 0.78, val auc 0.68\n",
      "train prec 0.71, train rec 0.7, val prec 0.68, val rec 0.48\n",
      "Epoch 5:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.52, train acc 0.74, val cost 0.66, val acc 0.66\n",
      "train auc 0.82, val auc 0.71\n",
      "train prec 0.75, train rec 0.73, val prec 0.68, val rec 0.56\n",
      "Epoch 6:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.48, train acc 0.76, val cost 0.66, val acc 0.66\n",
      "train auc 0.85, val auc 0.71\n",
      "train prec 0.78, train rec 0.75, val prec 0.67, val rec 0.62\n",
      "Epoch 7:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.44, train acc 0.79, val cost 0.65, val acc 0.67\n",
      "train auc 0.88, val auc 0.73\n",
      "train prec 0.81, train rec 0.77, val prec 0.68, val rec 0.62\n",
      "Epoch 8:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.4, train acc 0.81, val cost 0.67, val acc 0.67\n",
      "train auc 0.9, val auc 0.73\n",
      "train prec 0.83, train rec 0.8, val prec 0.67, val rec 0.67\n",
      "Epoch 9:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.37, train acc 0.83, val cost 0.71, val acc 0.68\n",
      "train auc 0.92, val auc 0.73\n",
      "train prec 0.85, train rec 0.82, val prec 0.66, val rec 0.69\n",
      "Epoch 10:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.32, train acc 0.86, val cost 0.73, val acc 0.67\n",
      "train auc 0.94, val auc 0.73\n",
      "train prec 0.88, train rec 0.85, val prec 0.66, val rec 0.67\n",
      "Epoch 11:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.29, train acc 0.88, val cost 0.79, val acc 0.67\n",
      "train auc 0.95, val auc 0.73\n",
      "train prec 0.89, train rec 0.87, val prec 0.66, val rec 0.7\n",
      "Epoch 12:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.27, train acc 0.89, val cost 0.8, val acc 0.68\n",
      "train auc 0.96, val auc 0.74\n",
      "train prec 0.9, train rec 0.87, val prec 0.67, val rec 0.67\n",
      "Epoch 13:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.24, train acc 0.9, val cost 0.84, val acc 0.67\n",
      "train auc 0.97, val auc 0.73\n",
      "train prec 0.92, train rec 0.9, val prec 0.66, val rec 0.62\n",
      "Epoch 14:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.21, train acc 0.92, val cost 0.97, val acc 0.65\n",
      "train auc 0.98, val auc 0.72\n",
      "train prec 0.93, train rec 0.91, val prec 0.64, val rec 0.66\n",
      "Epoch 15:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.19, train acc 0.93, val cost 0.99, val acc 0.66\n",
      "train auc 0.98, val auc 0.72\n",
      "train prec 0.94, train rec 0.92, val prec 0.68, val rec 0.59\n",
      "Epoch 16:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.17, train acc 0.94, val cost 1.0, val acc 0.66\n",
      "train auc 0.98, val auc 0.72\n",
      "train prec 0.95, train rec 0.93, val prec 0.69, val rec 0.55\n",
      "Epoch 17:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.15, train acc 0.94, val cost 0.96, val acc 0.67\n",
      "train auc 0.99, val auc 0.73\n",
      "train prec 0.95, train rec 0.94, val prec 0.63, val rec 0.77\n",
      "Epoch 18:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.14, train acc 0.95, val cost 1.1, val acc 0.67\n",
      "train auc 0.99, val auc 0.73\n",
      "train prec 0.95, train rec 0.94, val prec 0.66, val rec 0.66\n",
      "Epoch 19:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.12, train acc 0.96, val cost 1.1, val acc 0.67\n",
      "train auc 0.99, val auc 0.73\n",
      "train prec 0.96, train rec 0.96, val prec 0.68, val rec 0.61\n",
      "New LR: 7e-06\n",
      "Epoch 20:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.093, train acc 0.97, val cost 1.2, val acc 0.67\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 0.97, train rec 0.97, val prec 0.68, val rec 0.57\n",
      "Epoch 21:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.083, train acc 0.97, val cost 1.2, val acc 0.67\n",
      "train auc 1.0, val auc 0.73\n",
      "train prec 0.98, train rec 0.97, val prec 0.7, val rec 0.59\n",
      "Epoch 22:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.078, train acc 0.98, val cost 1.4, val acc 0.66\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 0.98, train rec 0.97, val prec 0.71, val rec 0.55\n",
      "Epoch 23:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.07, train acc 0.98, val cost 1.3, val acc 0.68\n",
      "train auc 1.0, val auc 0.73\n",
      "train prec 0.98, train rec 0.98, val prec 0.69, val rec 0.64\n",
      "Epoch 24:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.064, train acc 0.98, val cost 1.4, val acc 0.67\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 0.98, train rec 0.98, val prec 0.71, val rec 0.56\n",
      "Epoch 25:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.064, train acc 0.98, val cost 1.5, val acc 0.66\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 0.98, train rec 0.98, val prec 0.7, val rec 0.55\n",
      "Epoch 26:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.059, train acc 0.98, val cost 1.5, val acc 0.66\n",
      "train auc 1.0, val auc 0.71\n",
      "train prec 0.98, train rec 0.98, val prec 0.71, val rec 0.54\n",
      "Epoch 27:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.053, train acc 0.98, val cost 1.7, val acc 0.65\n",
      "train auc 1.0, val auc 0.69\n",
      "train prec 0.99, train rec 0.98, val prec 0.72, val rec 0.47\n",
      "Epoch 28:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.045, train acc 0.99, val cost 1.8, val acc 0.65\n",
      "train auc 1.0, val auc 0.7\n",
      "train prec 0.99, train rec 0.99, val prec 0.71, val rec 0.44\n",
      "Epoch 29:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.043, train acc 0.99, val cost 1.6, val acc 0.67\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 0.99, train rec 0.99, val prec 0.69, val rec 0.59\n",
      "Epoch 30:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.035, train acc 0.99, val cost 1.6, val acc 0.67\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 0.99, train rec 0.99, val prec 0.69, val rec 0.57\n",
      "Epoch 31:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.031, train acc 0.99, val cost 1.8, val acc 0.66\n",
      "train auc 1.0, val auc 0.71\n",
      "train prec 0.99, train rec 0.99, val prec 0.71, val rec 0.53\n",
      "Epoch 32:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.035, train acc 0.99, val cost 1.7, val acc 0.67\n",
      "train auc 1.0, val auc 0.73\n",
      "train prec 0.99, train rec 0.99, val prec 0.69, val rec 0.6\n",
      "Epoch 33:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.03, train acc 0.99, val cost 1.6, val acc 0.67\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 0.99, train rec 0.99, val prec 0.69, val rec 0.58\n",
      "Epoch 34:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.028, train acc 0.99, val cost 1.7, val acc 0.67\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 0.99, train rec 0.99, val prec 0.69, val rec 0.6\n",
      "Epoch 35:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.025, train acc 0.99, val cost 1.8, val acc 0.67\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 0.99, train rec 0.99, val prec 0.72, val rec 0.56\n",
      "Epoch 36:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.024, train acc 0.99, val cost 2.0, val acc 0.66\n",
      "train auc 1.0, val auc 0.71\n",
      "train prec 0.99, train rec 0.99, val prec 0.74, val rec 0.51\n",
      "Epoch 37:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.022, train acc 0.99, val cost 1.8, val acc 0.68\n",
      "train auc 1.0, val auc 0.73\n",
      "train prec 0.99, train rec 0.99, val prec 0.69, val rec 0.61\n",
      "Epoch 38:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.022, train acc 0.99, val cost 1.8, val acc 0.67\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 0.99, train rec 0.99, val prec 0.71, val rec 0.56\n",
      "Epoch 39:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.022, train acc 0.99, val cost 1.8, val acc 0.67\n",
      "train auc 1.0, val auc 0.73\n",
      "train prec 0.99, train rec 0.99, val prec 0.69, val rec 0.61\n",
      "New LR: 4.9e-06\n",
      "Epoch 40:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.018, train acc 1.0, val cost 1.8, val acc 0.67\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 0.99, train rec 0.99, val prec 0.7, val rec 0.57\n",
      "Epoch 41:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.015, train acc 1.0, val cost 1.8, val acc 0.68\n",
      "train auc 1.0, val auc 0.73\n",
      "train prec 1.0, train rec 1.0, val prec 0.68, val rec 0.65\n",
      "Epoch 42:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.011, train acc 1.0, val cost 2.1, val acc 0.67\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 1.0, train rec 1.0, val prec 0.71, val rec 0.56\n",
      "Epoch 43:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.01, train acc 1.0, val cost 1.9, val acc 0.69\n",
      "train auc 1.0, val auc 0.73\n",
      "train prec 1.0, train rec 1.0, val prec 0.69, val rec 0.62\n",
      "Epoch 44:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.011, train acc 1.0, val cost 2.1, val acc 0.66\n",
      "train auc 1.0, val auc 0.7\n",
      "train prec 1.0, train rec 1.0, val prec 0.7, val rec 0.5\n",
      "Epoch 45:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.015, train acc 1.0, val cost 2.1, val acc 0.67\n",
      "train auc 1.0, val auc 0.71\n",
      "train prec 1.0, train rec 0.99, val prec 0.69, val rec 0.54\n",
      "Epoch 46:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.0099, train acc 1.0, val cost 2.1, val acc 0.66\n",
      "train auc 1.0, val auc 0.71\n",
      "train prec 1.0, train rec 1.0, val prec 0.69, val rec 0.53\n",
      "Epoch 47:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.0084, train acc 1.0, val cost 2.0, val acc 0.68\n",
      "train auc 1.0, val auc 0.72\n",
      "train prec 1.0, train rec 1.0, val prec 0.7, val rec 0.58\n",
      "Epoch 48:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, 100, 110, 120, 130, 140, 150, 160, 170, 180, 190, 200, 210, 220, 230, 240, 250, 260, 270, 280, 290, 300, 310, 320, 330, 340, 350, 360, 370, 380, 390, \n",
      "valid:, 0, 10, 20, 30, \n",
      "train cost 0.013, train acc 1.0, val cost 2.2, val acc 0.66\n",
      "train auc 1.0, val auc 0.7\n",
      "train prec 1.0, train rec 0.99, val prec 0.74, val rec 0.5\n",
      "Epoch 49:, \n",
      "train:, 0, 10, 20, 30, 40, 50, 60, 70, 80, 90, "
     ]
    }
   ],
   "source": [
    "NUM_EPOCHS = 50\n",
    "\n",
    "valid_accs, train_accs, test_accs = [], [], []\n",
    "valid_precs, train_precs, valid_recs, train_recs = [], [], [], []\n",
    "train_costs, valid_costs = [], []\n",
    "train_aucs, valid_aucs = [], []\n",
    "\n",
    "learning_rate=0.00001\n",
    "# define which variables are included in the training\n",
    "tr_op = train_op4\n",
    "print('Number of epochs: %s' % NUM_EPOCHS)\n",
    "\n",
    "try:\n",
    "    for n in range(NUM_EPOCHS):\n",
    "        print(\"Epoch %d:\" % n, end=', ')\n",
    "        print('\\ntrain:', end=', ')\n",
    "        train_cost, train_acc, train_prec, train_rec, train_summary, train_auc = train_epoch(learning_rate, tr_op)\n",
    "        train_costs += [train_cost]\n",
    "        train_accs += [train_acc]\n",
    "        train_aucs += [train_auc]\n",
    "        train_precs += [train_prec]\n",
    "        train_recs += [train_rec]\n",
    "        print('valid:', end=', ')\n",
    "        valid_cost, valid_acc, preds, valid_prec, valid_rec, valid_summary, valid_auc = eval_epoch()\n",
    "        valid_costs += [valid_cost]\n",
    "        valid_accs += [valid_acc]\n",
    "        valid_aucs += [valid_auc]\n",
    "        valid_precs += [valid_prec]\n",
    "        valid_recs += [valid_rec]\n",
    "\n",
    "        print(\"train cost {0:.2}, train acc {1:.2}, val cost {2:.2}, val acc {3:.2}\".format(\n",
    "                train_cost, train_acc, valid_cost, valid_acc))\n",
    "        print(\"train auc {0:.2}, val auc {1:.2}\". format(train_auc, valid_auc))\n",
    "        print(\"train prec {0:.2}, train rec {1:.2}, val prec {2:.2}, val rec {3:.2}\". format(\n",
    "                train_prec, train_rec, valid_prec, valid_rec))\n",
    "        \n",
    "        # learning rate annealing\n",
    "        if (n+1) % 20 == 0:\n",
    "            learning_rate = learning_rate * 0.7\n",
    "            print(\"New LR:\", learning_rate)      \n",
    "        \n",
    "except KeyboardInterrupt:\n",
    "    pass"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 31,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# save session?\n",
    "saver = tf.train.Saver()\n",
    "save_path = saver.save(sess, \"output/03ST_TL.ckpt\")\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Plot errors and zoom"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 54,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f2ad42fb438>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.figure(figsize=(9,9))\n",
    "plt.plot(np.array(train_accs), label='Training Acc')\n",
    "plt.plot(np.array(valid_accs), label='Validation Acc')\n",
    "plt.plot(np.array(train_aucs), label='Train AUC')\n",
    "plt.plot(np.array(valid_aucs), label='Validation AUC')\n",
    "plt.legend(fontsize=20)\n",
    "plt.xlabel('Epoch', fontsize=20)\n",
    "plt.savefig('output/TL_STN_ACC_AUC')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Show original image, transformed image and the weights used for transformation"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "X_batch_val, y_batch_val = train_generator.next()\n",
    "fetches_val = [prediction, l_transform, w_transform, hsv, vgg16_mid]\n",
    "feed_dict_val = {x_pl: X_batch_val}\n",
    "res = sess.run(fetches=fetches_val, feed_dict=feed_dict_val)\n",
    "output_eval, transform_eval, weight_eval, hsv_im, convnet = tuple(res)\n",
    "\n",
    "print(X_batch_val.shape)\n",
    "\n",
    "plt.figure(figsize=(10,8))\n",
    "\n",
    "offset = 0\n",
    "\n",
    "for i in range(3):\n",
    "    img = X_batch_val[i+offset]\n",
    "    mi = np.min(img)\n",
    "    img = img - mi\n",
    "    ma = np.max(img)\n",
    "    img = img / ma\n",
    "    trans = transform_eval[i+offset]\n",
    "    mi = np.min(trans)\n",
    "    trans = trans - mi\n",
    "    ma = np.max(trans)\n",
    "    trans = trans / ma\n",
    "\n",
    "    plt.subplot(331+i*3)\n",
    "    plt.imshow(img, interpolation='none')\n",
    "    if i == 0:\n",
    "        plt.title('Original {0}x{0}'.format(X_batch_val.shape[1]), fontsize=10)\n",
    "    plt.axis('off')\n",
    "    plt.subplot(332+i*3)\n",
    "    plt.imshow(trans, interpolation='none')\n",
    "    if i == 0:\n",
    "        plt.title('Transformed {0}x{0}'.format(l_transform.shape[1]), fontsize=10)\n",
    "    plt.axis('off')\n",
    "    plt.subplot(333+i*3)\n",
    "    plt.imshow(weight_eval[i+offset].reshape(2,3), interpolation='nearest', cmap='jet')\n",
    "    plt.colorbar()\n",
    "    if i == 0:\n",
    "        plt.title('Transformation weights', fontsize=10)\n",
    "    plt.axis('off')\n",
    "    \n",
    "plt.tight_layout()\n",
    "# plt.savefig('output/TL_STN_weights')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Test data"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 46,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Found 617 images belonging to 2 classes.\n",
      "batch no 1, number of samples: 100\n",
      "batch no 2, number of samples: 100\n",
      "batch no 3, number of samples: 100\n",
      "batch no 4, number of samples: 100\n",
      "batch no 5, number of samples: 100\n",
      "batch no 6, number of samples: 100\n",
      "batch no 7, number of samples: 17\n"
     ]
    }
   ],
   "source": [
    "test_generator = datagen.flow_from_directory(\n",
    "    test_data_dir,\n",
    "    target_size=(img_height, img_width),\n",
    "    batch_size=BATCH_SIZE,\n",
    "    class_mode='categorical',\n",
    "    shuffle=False)\n",
    "\n",
    "num_samples = test_generator.n\n",
    "num_batches = int(np.ceil(num_samples / float(BATCH_SIZE)))\n",
    "\n",
    "test_logits = np.zeros((num_samples, 2))\n",
    "y = np.zeros((num_samples, 2))\n",
    "\n",
    "for i in range(num_batches):\n",
    "\n",
    "    x_batch_test, y_batch_test = test_generator.next()\n",
    "    n_batch = x_batch_test.shape[0]\n",
    "    print(\"batch no {}, number of samples: {}\".format(i+1, n_batch))\n",
    "    fetches_test = [logits, prediction, accuracy, auc_op]\n",
    "    feed_dict_test= {x_pl: x_batch_test, y_pl: y_batch_test}\n",
    "    res = sess.run(fetches=fetches_test, feed_dict=feed_dict_test)\n",
    "    \n",
    "    test_logits[i*BATCH_SIZE: i*BATCH_SIZE + n_batch, :] = res[0]\n",
    "    y[i*BATCH_SIZE: i*BATCH_SIZE + n_batch, :] = y_batch_test"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 50,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Sklearn auc: 0.7305633147113593\n"
     ]
    }
   ],
   "source": [
    "# For the ROC curve\n",
    "from sklearn import metrics\n",
    "fpr, tpr, thresholds = metrics.roc_curve(y[:, 0], test_logits[:, 0])\n",
    "auc = metrics.auc(fpr, tpr)\n",
    "print(\"Sklearn auc: {}\".format(auc))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 51,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "image/png": "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\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x7f2ad42c90f0>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "plt.plot(fpr,tpr)\n",
    "plt.plot(fpr, tpr, color='darkorange',\n",
    "         lw=2, label='ROC curve (area = {0:0.2f})'.format(auc))\n",
    "plt.plot([0, 1], [0, 1], color='navy', lw=2, linestyle='--')\n",
    "plt.xlim([0.0, 1.0])\n",
    "plt.ylim([0.0, 1.05])\n",
    "plt.xlabel('False Positive Rate')\n",
    "plt.ylabel('True Positive Rate')\n",
    "plt.title('Receiver operating characteristic')\n",
    "plt.legend(loc=\"lower right\")\n",
    "plt.savefig('output/TL_STN_ROC')\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 39,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "acc: 0.72, auc: 0.73\n"
     ]
    }
   ],
   "source": [
    "acc = np.sum(np.equal(np.argmax(y, axis=1), np.argmax(test_logits, axis=1)))/num_samples\n",
    "\n",
    "print(\"acc: {0:0.2f}, auc: {1:0.2f}\".format(acc, auc))"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 40,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "397 45 35 140\n",
      "Precision, Recall, TPR, FPR\n",
      "0.918981481481 0.739292364991 0.739292364991 0.4375\n"
     ]
    }
   ],
   "source": [
    "# indices for positive and negative examples\n",
    "idxp = np.argmax(y, axis=1) == 0\n",
    "idxn = np.argmax(y, axis=1) == 1\n",
    "\n",
    "tp = np.sum(np.equal(np.argmax(y[idxp, :], axis=1), np.argmax(test_logits[idxp, :], axis=1)))\n",
    "tn = np.sum(np.equal(np.argmax(y[idxn, :], axis=1), np.argmax(test_logits[idxn, :], axis=1)))\n",
    "fp = np.sum(np.argmax(y[idxn, :], axis=1) != np.argmax(test_logits[idxn, :], axis=1))\n",
    "fn = np.sum(np.argmax(y[idxp, :], axis=1) != np.argmax(test_logits[idxp, :], axis=1))\n",
    "\n",
    "print(tp, tn, fp, fn)\n",
    "\n",
    "prec = tp / (tp + fp)\n",
    "rec = tp / (tp + fn)\n",
    "tpr = tp / (tp + fn)\n",
    "fpr = fp / (fp + tn)\n",
    "\n",
    "print(\"Precision, Recall, TPR, FPR\")\n",
    "print(prec, rec, tpr, fpr)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Inspect classifications on the test set"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "n = 17\n",
    "offset = 0\n",
    "plt.figure(figsize=(100,100))\n",
    "for i in range(n):\n",
    "    plt.subplot(n,1,i+1)\n",
    "    img = x_batch_test[i + offset]\n",
    "    mi = np.min(img)\n",
    "    img = img - mi\n",
    "    ma = np.max(img)\n",
    "    img = img / ma\n",
    "    plt.imshow(img)\n",
    "    plt.title('logits: {0:3.2}  {1:3.2}, truth: {2}'.format(res[0][i][0],res[0][i][1], y_batch_test[i+offset]))"
   ]
  }
 ],
 "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": "3.5.2"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}