MNIST-DANN.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "Extracting MNIST_data/train-images-idx3-ubyte.gz\n",
      "Extracting MNIST_data/train-labels-idx1-ubyte.gz\n",
      "Extracting MNIST_data/t10k-images-idx3-ubyte.gz\n",
      "Extracting MNIST_data/t10k-labels-idx1-ubyte.gz\n"
     ]
    }
   ],
   "source": [
    "%matplotlib inline\n",
    "\n",
    "import tensorflow as tf\n",
    "import numpy as np\n",
    "import pickle as pkl\n",
    "from sklearn.manifold import TSNE\n",
    "\n",
    "from flip_gradient import flip_gradient\n",
    "from utils import *\n",
    "\n",
    "from tensorflow.examples.tutorials.mnist import input_data\n",
    "mnist = input_data.read_data_sets('MNIST_data', one_hot=True)\n",
    "\n",
    "# Process MNIST\n",
    "mnist_train = (mnist.train.images > 0).reshape(55000, 28, 28, 1).astype(np.uint8) * 255\n",
    "mnist_train = np.concatenate([mnist_train, mnist_train, mnist_train], 3)\n",
    "mnist_test = (mnist.test.images > 0).reshape(10000, 28, 28, 1).astype(np.uint8) * 255\n",
    "mnist_test = np.concatenate([mnist_test, mnist_test, mnist_test], 3)\n",
    "\n",
    "# Load MNIST-M\n",
    "mnistm = pkl.load(open('mnistm_data.pkl','rb'))\n",
    "mnistm_train = mnistm['train']\n",
    "mnistm_test = mnistm['test']\n",
    "mnistm_valid = mnistm['valid']\n",
    "\n",
    "# Compute pixel mean for normalizing data\n",
    "pixel_mean = np.vstack([mnist_train, mnistm_train]).mean((0, 1, 2))\n",
    "\n",
    "# Create a mixed dataset for TSNE visualization\n",
    "num_test = 500\n",
    "combined_test_imgs = np.vstack([mnist_test[:num_test], mnistm_test[:num_test]])\n",
    "combined_test_labels = np.vstack([mnist.test.labels[:num_test], mnist.test.labels[:num_test]])\n",
    "combined_test_domain = np.vstack([np.tile([1., 0.], [num_test, 1]),\n",
    "        np.tile([0., 1.], [num_test, 1])])\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXQAAABzCAYAAACIEflfAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nrkJggg==\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x1c148bfc88>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    },
    {
     "data": {
      "image/png": "iVBORw0KGgoAAAANSUhEUgAAAXQAAABzCAYAAACIEflfAAAABHNCSVQICAgIfAhkiAAAAAlwSFlz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\nHlpooYW2QOz/ATxqm3MpgmuoAAAAAElFTkSuQmCC\n",
      "text/plain": [
       "<matplotlib.figure.Figure at 0x1c148bfc50>"
      ]
     },
     "metadata": {},
     "output_type": "display_data"
    }
   ],
   "source": [
    "imshow_grid(mnist_train)\n",
    "imshow_grid(mnistm_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [],
   "source": [
    "batch_size = 128\n",
    "\n",
    "class MNISTModel(object):\n",
    "    \"\"\"Simple MNIST domain adaptation model.\"\"\"\n",
    "    def __init__(self):\n",
    "        self._build_model()\n",
    "    \n",
    "    def _build_model(self):\n",
    "        \n",
    "        self.X = tf.placeholder(tf.uint8, [None, 28, 28, 3])\n",
    "        self.y = tf.placeholder(tf.float32, [None, 10])\n",
    "        self.domain = tf.placeholder(tf.float32, [None, 2])\n",
    "        self.l = tf.placeholder(tf.float32, [])\n",
    "        self.train = tf.placeholder(tf.bool, [])\n",
    "        \n",
    "        X_input = (tf.cast(self.X, tf.float32) - pixel_mean) / 255.\n",
    "        \n",
    "        # CNN model for feature extraction\n",
    "        with tf.variable_scope('feature_extractor'):\n",
    "\n",
    "            W_conv0 = weight_variable([5, 5, 3, 32])\n",
    "            b_conv0 = bias_variable([32])\n",
    "            h_conv0 = tf.nn.relu(conv2d(X_input, W_conv0) + b_conv0)\n",
    "            h_pool0 = max_pool_2x2(h_conv0)\n",
    "            \n",
    "            W_conv1 = weight_variable([5, 5, 32, 48])\n",
    "            b_conv1 = bias_variable([48])\n",
    "            h_conv1 = tf.nn.relu(conv2d(h_pool0, W_conv1) + b_conv1)\n",
    "            h_pool1 = max_pool_2x2(h_conv1)\n",
    "            \n",
    "            # The domain-invariant feature\n",
    "            self.feature = tf.reshape(h_pool1, [-1, 7*7*48])\n",
    "            \n",
    "        # MLP for class prediction\n",
    "        with tf.variable_scope('label_predictor'):\n",
    "            \n",
    "            # Switches to route target examples (second half of batch) differently\n",
    "            # depending on train or test mode.\n",
    "            all_features = lambda: self.feature\n",
    "            source_features = lambda: tf.slice(self.feature, [0, 0], [(int)(batch_size / 2), -1])\n",
    "            classify_feats = tf.cond(self.train, source_features, all_features)\n",
    "            \n",
    "            all_labels = lambda: self.y\n",
    "            source_labels = lambda: tf.slice(self.y, [0, 0], [int(batch_size / 2), -1])\n",
    "            self.classify_labels = tf.cond(self.train, source_labels, all_labels)\n",
    "            \n",
    "            W_fc0 = weight_variable([7 * 7 * 48, 100])\n",
    "            b_fc0 = bias_variable([100])\n",
    "            h_fc0 = tf.nn.relu(tf.matmul(classify_feats, W_fc0) + b_fc0)\n",
    "\n",
    "            W_fc1 = weight_variable([100, 100])\n",
    "            b_fc1 = bias_variable([100])\n",
    "            h_fc1 = tf.nn.relu(tf.matmul(h_fc0, W_fc1) + b_fc1)\n",
    "\n",
    "            W_fc2 = weight_variable([100, 10])\n",
    "            b_fc2 = bias_variable([10])\n",
    "            logits = tf.matmul(h_fc1, W_fc2) + b_fc2\n",
    "            \n",
    "            self.pred = tf.nn.softmax(logits)\n",
    "            self.pred_loss = tf.nn.softmax_cross_entropy_with_logits(logits=logits, labels=self.classify_labels)\n",
    "\n",
    "        # Small MLP for domain prediction with adversarial loss\n",
    "        with tf.variable_scope('domain_predictor'):\n",
    "            \n",
    "            # Flip the gradient when backpropagating through this operation\n",
    "            feat = flip_gradient(self.feature, self.l)\n",
    "            \n",
    "            d_W_fc0 = weight_variable([7 * 7 * 48, 100])\n",
    "            d_b_fc0 = bias_variable([100])\n",
    "            d_h_fc0 = tf.nn.relu(tf.matmul(feat, d_W_fc0) + d_b_fc0)\n",
    "            \n",
    "            d_W_fc1 = weight_variable([100, 2])\n",
    "            d_b_fc1 = bias_variable([2])\n",
    "            d_logits = tf.matmul(d_h_fc0, d_W_fc1) + d_b_fc1\n",
    "            \n",
    "            self.domain_pred = tf.nn.softmax(d_logits)\n",
    "            self.domain_loss = tf.nn.softmax_cross_entropy_with_logits(logits=d_logits, labels=self.domain)\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Build the model graph\n",
    "graph = tf.get_default_graph()\n",
    "with graph.as_default():\n",
    "    model = MNISTModel()\n",
    "    \n",
    "    learning_rate = tf.placeholder(tf.float32, [])\n",
    "    \n",
    "    pred_loss = tf.reduce_mean(model.pred_loss)\n",
    "    domain_loss = tf.reduce_mean(model.domain_loss)\n",
    "    total_loss = pred_loss + domain_loss\n",
    "\n",
    "    regular_train_op = tf.train.MomentumOptimizer(learning_rate, 0.9).minimize(pred_loss)\n",
    "    dann_train_op = tf.train.MomentumOptimizer(learning_rate, 0.9).minimize(total_loss)\n",
    "    \n",
    "    # Evaluation\n",
    "    correct_label_pred = tf.equal(tf.argmax(model.classify_labels, 1), tf.argmax(model.pred, 1))\n",
    "    label_acc = tf.reduce_mean(tf.cast(correct_label_pred, tf.float32))\n",
    "    correct_domain_pred = tf.equal(tf.argmax(model.domain, 1), tf.argmax(model.domain_pred, 1))\n",
    "    domain_acc = tf.reduce_mean(tf.cast(correct_domain_pred, tf.float32))\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "# Params\n",
    "num_steps = 8600\n",
    "\n",
    "def train_and_evaluate(training_mode, graph, model, verbose=False):\n",
    "    \"\"\"Helper to run the model with different training modes.\"\"\"\n",
    "\n",
    "    with tf.Session(graph=graph) as sess:\n",
    "        tf.global_variables_initializer().run()\n",
    "\n",
    "        # Batch generators\n",
    "        gen_source_batch = batch_generator(\n",
    "            [mnist_train, mnist.train.labels], batch_size / 2)\n",
    "        gen_target_batch = batch_generator(\n",
    "            [mnistm_train, mnist.train.labels], batch_size / 2)\n",
    "        gen_source_only_batch = batch_generator(\n",
    "            [mnist_train, mnist.train.labels], batch_size)\n",
    "        gen_target_only_batch = batch_generator(\n",
    "            [mnistm_train, mnist.train.labels], batch_size)\n",
    "\n",
    "        domain_labels = np.vstack([np.tile([1, 0], [(int)(batch_size / 2), 1]),\n",
    "                                   np.tile([0, 1], [(int)(batch_size / 2), 1])])\n",
    "\n",
    "        # Training loop\n",
    "        for i in range(num_steps):\n",
    "            \n",
    "            # Adaptation param and learning rate schedule as described in the paper\n",
    "            p = float(i) / num_steps\n",
    "            l = 2. / (1. + np.exp(-10. * p)) - 1\n",
    "            lr = 0.01 / (1. + 10 * p)**0.75\n",
    "\n",
    "            # Training step\n",
    "            if training_mode == 'dann':\n",
    "\n",
    "                X0, y0 = gen_source_batch.__next__()\n",
    "                X1, y1 = gen_target_batch.__next__()\n",
    "                X = np.vstack([X0, X1])\n",
    "                y = np.vstack([y0, y1])\n",
    "\n",
    "                _, batch_loss, dloss, ploss, d_acc, p_acc = \\\n",
    "                    sess.run([dann_train_op, total_loss, domain_loss, pred_loss, domain_acc, label_acc],\n",
    "                             feed_dict={model.X: X, model.y: y, model.domain: domain_labels,\n",
    "                                        model.train: True, model.l: l, learning_rate: lr})\n",
    "\n",
    "                if verbose and i % 100 == 0:\n",
    "                    print('loss: %f  d_acc: %f  p_acc: %f  p: %f  l: %f  lr: %f' % \\\n",
    "                            (batch_loss, d_acc, p_acc, p, l, lr))\n",
    "\n",
    "            elif training_mode == 'source':\n",
    "                X, y = gen_source_only_batch.__next__()\n",
    "                _, batch_loss = sess.run([regular_train_op, pred_loss],\n",
    "                                     feed_dict={model.X: X, model.y: y, model.train: False,\n",
    "                                                model.l: l, learning_rate: lr})\n",
    "\n",
    "            elif training_mode == 'target':\n",
    "                X, y = gen_target_only_batch.__next__()\n",
    "                _, batch_loss = sess.run([regular_train_op, pred_loss],\n",
    "                                     feed_dict={model.X: X, model.y: y, model.train: False,\n",
    "                                                model.l: l, learning_rate: lr})\n",
    "\n",
    "        # Compute final evaluation on test data\n",
    "        source_acc = sess.run(label_acc,\n",
    "                            feed_dict={model.X: mnist_test, model.y: mnist.test.labels,\n",
    "                                       model.train: False})\n",
    "\n",
    "        target_acc = sess.run(label_acc,\n",
    "                            feed_dict={model.X: mnistm_test, model.y: mnist.test.labels,\n",
    "                                       model.train: False})\n",
    "        \n",
    "        test_domain_acc = sess.run(domain_acc,\n",
    "                            feed_dict={model.X: combined_test_imgs,\n",
    "                                       model.domain: combined_test_domain, model.l: 1.0})\n",
    "        \n",
    "        test_emb = sess.run(model.feature, feed_dict={model.X: combined_test_imgs})\n",
    "        \n",
    "    return source_acc, target_acc, test_domain_acc, test_emb\n",
    "\n",
    "\n",
    "print('\\nSource only training')\n",
    "source_acc, target_acc, _, source_only_emb = train_and_evaluate('source', graph, model)\n",
    "print('Source (MNIST) accuracy:', source_acc)\n",
    "print('Target (MNIST-M) accuracy:', target_acc)\n",
    "\n",
    "print('\\nDomain adaptation training')\n",
    "source_acc, target_acc, d_acc, dann_emb = train_and_evaluate('dann', graph, model)\n",
    "print('Source (MNIST) accuracy:', source_acc)\n",
    "print('Target (MNIST-M) accuracy:', target_acc)\n",
    "print('Domain accuracy:', d_acc)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=3000)\n",
    "source_only_tsne = tsne.fit_transform(source_only_emb)\n",
    "\n",
    "tsne = TSNE(perplexity=30, n_components=2, init='pca', n_iter=3000)\n",
    "dann_tsne = tsne.fit_transform(dann_emb)\n",
    "        \n",
    "plot_embedding(source_only_tsne, combined_test_labels.argmax(1), combined_test_domain.argmax(1), 'Source only')\n",
    "plot_embedding(dann_tsne, combined_test_labels.argmax(1), combined_test_domain.argmax(1), 'Domain Adaptation')"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": []
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3",
   "language": "python",
   "name": "python3"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.6.3"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 1
}