Faster 'same' mode convolutions in Lasagne, for even filter sizes too!

Fast padding.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "Using gpu device 1: GeForce GTX 780 Ti (CNMeM is disabled)\n"
     ]
    }
   ],
   "source": [
    "import lasagne\n",
    "import theano\n",
    "import numpy as np"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "1 loops, best of 3: 198 ms per loop\n",
      "10 loops, best of 3: 133 ms per loop\n",
      "10 loops, best of 3: 82.3 ms per loop\n",
      "True\n",
      "True\n"
     ]
    }
   ],
   "source": [
    "filter_size = (5, 12)\n",
    "input_shape = (100, 3, 30, 40)\n",
    "n_filters = 16\n",
    "dummy_input = np.random.standard_normal(input_shape).astype(theano.config.floatX)\n",
    "\n",
    "# To produce the same results for each network\n",
    "lasagne.random.get_rng().seed(1234)\n",
    "l_in = lasagne.layers.InputLayer(input_shape)\n",
    "# Use 'full' convolution (relying on Theano to do the padding)\n",
    "l_conv = lasagne.layers.Conv2DLayer(l_in, n_filters, filter_size, pad='full')\n",
    "# Crop out the 'same' portion of the convolution using SliceLayers\n",
    "shift_x = (filter_size[0] - 1) // 2\n",
    "shift_y = (filter_size[1] - 1) // 2\n",
    "l_conv = lasagne.layers.SliceLayer(l_conv, slice(shift_x, input_shape[2] + shift_x), 2)\n",
    "l_conv = lasagne.layers.SliceLayer(l_conv, slice(shift_y, input_shape[3] + shift_y), 3)\n",
    "output_a = lasagne.layers.get_output(l_conv).eval({l_in.input_var: dummy_input})\n",
    "%timeit lasagne.layers.get_output(l_conv).eval({l_in.input_var: dummy_input})\n",
    "\n",
    "lasagne.random.get_rng().seed(1234)\n",
    "l_in = lasagne.layers.InputLayer(input_shape)\n",
    "# This mode will no longer work for even filter sizes\n",
    "l_conv = lasagne.layers.Conv2DLayer(l_in, n_filters, filter_size, pad='same')\n",
    "output_b = lasagne.layers.get_output(l_conv).eval({l_in.input_var: dummy_input})\n",
    "%timeit lasagne.layers.get_output(l_conv).eval({l_in.input_var: dummy_input})\n",
    "\n",
    "lasagne.random.get_rng().seed(1234)\n",
    "l_in = lasagne.layers.InputLayer(input_shape)\n",
    "# Not specifying a pad argument corresponds to pad=0, or 'valid' mode\n",
    "l_conv = lasagne.layers.Conv2DLayer(l_in, n_filters, filter_size)\n",
    "# Pre-pad the input with zeros, which will be cropped by the convolution operation\n",
    "shift_x = int(np.ceil((filter_size[0] - 1) / 2.))\n",
    "shift_y = int(np.ceil((filter_size[1] - 1) / 2.))\n",
    "dummy_input_padded = np.zeros((input_shape[0], input_shape[1],\n",
    "                               input_shape[2] + filter_size[0] - 1,\n",
    "                               input_shape[3] + filter_size[1] - 1),\n",
    "                              dtype=theano.config.floatX)\n",
    "dummy_input_padded[:, :, shift_x:input_shape[2] + shift_x,\n",
    "                   shift_y:input_shape[3] + shift_y] = dummy_input\n",
    "output_c = lasagne.layers.get_output(l_conv).eval({l_in.input_var: dummy_input_padded})\n",
    "%timeit lasagne.layers.get_output(l_conv).eval({l_in.input_var: dummy_input_padded})\n",
    "\n",
    "print np.allclose(output_a, output_b)\n",
    "print np.allclose(output_b, output_c)"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 2",
   "language": "python",
   "name": "python2"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 2
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython2",
   "version": "2.7.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}