{ "cells": [ { "cell_type": "markdown", "metadata": {}, "source": [ "Deep Learning Models -- A collection of various deep learning architectures, models, and tips for TensorFlow and PyTorch in Jupyter Notebooks.\n", "- Author: Sebastian Raschka\n", "- GitHub Repository: https://github.com/rasbt/deeplearning-models" ] }, { "cell_type": "code", "execution_count": 1, "metadata": {}, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Sebastian Raschka \n", "\n", "CPython 3.6.1\n", "IPython 6.0.0\n", "\n", "tensorflow 1.2.0\n" ] } ], "source": [ "%load_ext watermark\n", "%watermark -a 'Sebastian Raschka' -v -p tensorflow" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "# Model Zoo -- Siamese Network with Multilayer Perceptrons" ] }, { "cell_type": "code", "execution_count": 2, "metadata": { "scrolled": true }, "outputs": [ { "name": "stdout", "output_type": "stream", "text": [ "Extracting ./train-images-idx3-ubyte.gz\n", "Extracting ./train-labels-idx1-ubyte.gz\n", "Extracting ./t10k-images-idx3-ubyte.gz\n", "Extracting ./t10k-labels-idx1-ubyte.gz\n", "Initializing variables:\n", "\n", "\n", "\n", "\n", "\n", "\n", "Epoch: 001 | AvgCost: 0.472\n", "Epoch: 002 | AvgCost: 0.258\n", "Epoch: 003 | AvgCost: 0.250\n", "Epoch: 004 | AvgCost: 0.250\n", "Epoch: 005 | AvgCost: 0.250\n" ] } ], "source": [ "import numpy as np\n", "import tensorflow as tf\n", "from tensorflow.examples.tutorials.mnist import input_data\n", "\n", "\n", "##########################\n", "### SETTINGS\n", "##########################\n", "\n", "# General settings\n", "\n", "random_seed = 0\n", "\n", "# Hyperparameters\n", "learning_rate = 0.001\n", "training_epochs = 5\n", "batch_size = 100\n", "margin = 1.0\n", "\n", "# Architecture\n", "n_hidden_1 = 256\n", "n_hidden_2 = 256\n", "n_input = 784\n", "n_classes = 1 # for 'true' and 'false' matches\n", "\n", "\n", "def fully_connected(inputs, output_nodes, activation=None, seed=None):\n", "\n", " input_nodes = inputs.get_shape().as_list()[1]\n", " weights = tf.get_variable(name='weights', \n", " shape=(input_nodes, output_nodes),\n", " initializer=tf.truncated_normal_initializer(\n", " mean=0.0,\n", " stddev=0.001,\n", " dtype=tf.float32,\n", " seed=seed))\n", "\n", " biases = tf.get_variable(name='biases', \n", " shape=(output_nodes,),\n", " initializer=tf.constant_initializer(\n", " value=0.0, \n", " dtype=tf.float32))\n", " \n", " act = tf.matmul(inputs, weights) + biases\n", " if activation is not None:\n", " act = activation(act)\n", " return act\n", "\n", "\n", "def euclidean_distance(x_1, x_2):\n", " return tf.sqrt(tf.maximum(tf.sum(\n", " tf.square(x - y), axis=1, keepdims=True), 1e-06))\n", "\n", "def contrastive_loss(x_1, x_2, margin=1.0):\n", " return (x_1 * tf.square(x_2) +\n", " (1.0 - x_1) * tf.square(tf.maximum(margin - x_2, 0.)))\n", "\n", "\n", "##########################\n", "### GRAPH DEFINITION\n", "##########################\n", "\n", "g = tf.Graph()\n", "with g.as_default():\n", " \n", " tf.set_random_seed(random_seed)\n", "\n", " # Input data\n", " tf_x_1 = tf.placeholder(tf.float32, [None, n_input], name='inputs_1')\n", " tf_x_2 = tf.placeholder(tf.float32, [None, n_input], name='inputs_2')\n", " tf_y = tf.placeholder(tf.float32, [None], \n", " name='targets') # here: 'true' or 'false' valuess\n", "\n", " # Siamese Network\n", " def build_mlp(inputs):\n", " with tf.variable_scope('fc_1'):\n", " layer_1 = fully_connected(inputs, n_hidden_1, \n", " activation=tf.nn.relu)\n", " with tf.variable_scope('fc_2'):\n", " layer_2 = fully_connected(layer_1, n_hidden_2, \n", " activation=tf.nn.relu)\n", " with tf.variable_scope('fc_3'):\n", " out_layer = fully_connected(layer_2, n_classes, \n", " activation=tf.nn.relu)\n", "\n", " return out_layer\n", " \n", " \n", " with tf.variable_scope('siamese_net', reuse=False):\n", " pred_left = build_mlp(tf_x_1)\n", " with tf.variable_scope('siamese_net', reuse=True):\n", " pred_right = build_mlp(tf_x_2)\n", " \n", " # Loss and optimizer\n", " loss = contrastive_loss(pred_left, pred_right)\n", " cost = tf.reduce_mean(loss, name='cost')\n", " optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate)\n", " train = optimizer.minimize(cost, name='train')\n", " \n", "##########################\n", "### TRAINING & EVALUATION\n", "##########################\n", "\n", "np.random.seed(random_seed) # set seed for mnist shuffling\n", "mnist = input_data.read_data_sets(\"./\", one_hot=False)\n", "\n", "with tf.Session(graph=g) as sess:\n", " \n", " print('Initializing variables:')\n", " sess.run(tf.global_variables_initializer())\n", " for i in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,\n", " scope='siamese_net'):\n", " print(i)\n", "\n", " for epoch in range(training_epochs):\n", " avg_cost = 0.\n", " \n", " total_batch = mnist.train.num_examples // batch_size // 2\n", "\n", " for i in range(total_batch):\n", " \n", " batch_x_1, batch_y_1 = mnist.train.next_batch(batch_size)\n", " batch_x_2, batch_y_2 = mnist.train.next_batch(batch_size)\n", " batch_y = (batch_y_1 == batch_y_2).astype('float32')\n", " \n", " _, c = sess.run(['train', 'cost:0'], feed_dict={'inputs_1:0': batch_x_1,\n", " 'inputs_2:0': batch_x_2,\n", " 'targets:0': batch_y})\n", " avg_cost += c\n", "\n", " print(\"Epoch: %03d | AvgCost: %.3f\" % (epoch + 1, avg_cost / (i + 1)))" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "- Todo: add embedding visualization" ] } ], "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": "3.7.1" } }, "nbformat": 4, "nbformat_minor": 2 }