Colin's example.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 15,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import theano.tensor as T\n",
    "import theano\n",
    "import lasagne\n",
    "try:\n",
    "    from lasagne.layers.dnn import Conv2DDNNLayer as Conv2DLayer\n",
    "    from lasagne.layers.dnn import MaxPool2DDNNLayer as MaxPool2DLayer\n",
    "except ImportError:\n",
    "    from lasagne.layers import Conv2DLayer, MaxPool2DLayer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 126,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "# Hyperparameters\n",
    "num_filters = [1, 1]\n",
    "\n",
    "# First layer is 8 frames wide and 12 semitones tall (for example)\n",
    "filter_size = [(4, 12), (4, 3)]\n",
    "\n",
    "# Pooling sizes\n",
    "ds = [(1, 4), (1, 4)]\n",
    "\n",
    "hidden_layer_sizes = [128, 128]\n",
    "\n",
    "output_dim = 3\n",
    "\n",
    "batch_size = 50\n",
    "\n",
    "sequence_length = 128\n",
    "\n",
    "learning_rate = 10**-3\n",
    "\n",
    "momentum = .5"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 141,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "num_channels = 1\n",
    "num_features = 216"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 142,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Create network, starting with input layer\n",
    "layers = [lasagne.layers.InputLayer(shape=(None,\n",
    "                                           num_channels,\n",
    "                                           None,\n",
    "                                           num_features))]\n",
    "\n",
    "# Add each convolutional and pooling layer recursively\n",
    "for n in range(len(num_filters)):\n",
    "    layers.append(Conv2DLayer(layers[-1],\n",
    "                              num_filters=num_filters[n],\n",
    "                              filter_size=filter_size[n],\n",
    "                              nonlinearity=lasagne.nonlinearities.rectify,\n",
    "                              border_mode='same'))\n",
    "    \n",
    "    layers.append(MaxPool2DLayer(layers[-1], ds[n]))\n",
    "    \n",
    "# The next few lines basically take the time step dimension and squash it into\n",
    "# the first (batch) dimension, which is what we want for dense layers.\n",
    "layers.append(lasagne.layers.DimshuffleLayer(layers[-1], (0, 2, 1, 3)))\n",
    "\n",
    "conv_output_shape = layers[-1].get_output_shape()\n",
    "\n",
    "layers.append(lasagne.layers.ReshapeLayer(layers[-1],\n",
    "                                          (-1, conv_output_shape[2] * conv_output_shape[3])))\n",
    "\n",
    "# Add dense hidden layers\n",
    "for hidden_layer_size in hidden_layer_sizes:\n",
    "    layers.append(lasagne.layers.DenseLayer(layers[-1],\n",
    "                                            num_units=hidden_layer_size,\n",
    "                                            nonlinearity=lasagne.nonlinearities.rectify))\n",
    "\n",
    "    # Add output layer (there are other nonlinearities depending on what you want)\n",
    "layers.append(lasagne.layers.DenseLayer(layers[-1], num_units=output_dim,\n",
    "                                        nonlinearity=lasagne.nonlinearities.sigmoid))\n",
    "\n",
    "\n",
    "# Now, define a cost using theano... something like\n",
    "Y = T.imatrix('target')\n",
    "\n",
    "cost = T.nnet.binary_crossentropy(layers[-1].get_output(), Y).mean()\n",
    "\n",
    "# Compute updates, there's also SGD with momentum, Adagrad, etc.\n",
    "updates = lasagne.updates.rmsprop(cost,\n",
    "                                  lasagne.layers.get_all_params(layers[-1]),\n",
    "                                  learning_rate, momentum)\n",
    "\n",
    "# Compile theano functions for train/test\n",
    "# input_var is a theano variable automatically created by the InputLayer\n",
    "train = theano.function([layers[0].input_var, Y], cost, updates=updates)\n",
    "\n",
    "# Other useful functions\n",
    "compute_cost = theano.function([layers[0].input_var, Y], cost)\n",
    "\n",
    "output = theano.function([layers[0].input_var], layers[-1].get_output())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 188,
   "metadata": {
    "collapsed": false
   },
   "outputs": [],
   "source": [
    "# Load in training data here as X_train; a list of np.ndarrays each with shape\n",
    "# (num_channels, num_frames, num_features)\n",
    "# num_channels is probably 1 in your case\n",
    "\n",
    "X_train = np.random.randn(*(5, num_channels, sequence_length, num_features))\n",
    "Y_train = np.random.randint(0, high=2, size=(len(X_train), X_train.shape[2], output_dim)).astype(np.int32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 189,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "(5, 1, 128, 216) (5, 128, 3)\n"
     ]
    }
   ],
   "source": [
    "print X_train.shape, Y_train.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 190,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "ename": "TypeError",
     "evalue": "('Bad input argument to theano function with name \"<ipython-input-142-e5b726cf2525>:52\"  at index 1(0-based)', 'Wrong number of dimensions: expected 2, got 3 with shape (5, 128, 3).')",
     "output_type": "error",
     "traceback": [
      "\u001b[1;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[1;31mTypeError\u001b[0m                                 Traceback (most recent call last)",
      "\u001b[1;32m<ipython-input-190-2f6a822651f9>\u001b[0m in \u001b[0;36m<module>\u001b[1;34m()\u001b[0m\n\u001b[1;32m----> 1\u001b[1;33m \u001b[0mcompute_cost\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0mX_train\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0mY_train\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m",
      "\u001b[1;32m/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.pyc\u001b[0m in \u001b[0;36m__call__\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m    511\u001b[0m                     \u001b[1;32mtry\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m    512\u001b[0m                         s.storage[0] = s.type.filter(arg, strict=s.strict,\n\u001b[1;32m--> 513\u001b[1;33m                                 allow_downcast=s.allow_downcast)\n\u001b[0m\u001b[0;32m    514\u001b[0m \u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m    515\u001b[0m                     \u001b[1;32mexcept\u001b[0m \u001b[0mException\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0me\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
      "\u001b[1;32m/usr/local/lib/python2.7/dist-packages/theano/tensor/type.pyc\u001b[0m in \u001b[0;36mfilter\u001b[1;34m(self, data, strict, allow_downcast)\u001b[0m\n\u001b[0;32m    167\u001b[0m             raise TypeError(\"Wrong number of dimensions: expected %s,\"\n\u001b[0;32m    168\u001b[0m                             \" got %s with shape %s.\" % (self.ndim, data.ndim,\n\u001b[1;32m--> 169\u001b[1;33m                                                         data.shape))\n\u001b[0m\u001b[0;32m    170\u001b[0m         \u001b[1;32mif\u001b[0m \u001b[1;32mnot\u001b[0m \u001b[0mdata\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mflags\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0maligned\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m    171\u001b[0m             \u001b[1;32mtry\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
      "\u001b[1;31mTypeError\u001b[0m: ('Bad input argument to theano function with name \"<ipython-input-142-e5b726cf2525>:52\"  at index 1(0-based)', 'Wrong number of dimensions: expected 2, got 3 with shape (5, 128, 3).')"
     ]
    }
   ],
   "source": [
    "compute_cost(X_train, Y_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 182,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "ename": "TypeError",
     "evalue": "('Bad input argument to theano function with name \"<ipython-input-142-e5b726cf2525>:49\"  at index 1(0-based)', 'Wrong number of dimensions: expected 2, got 3 with shape (5, 128, 3).')",
     "output_type": "error",
     "traceback": [
      "\u001b[1;31m---------------------------------------------------------------------------\u001b[0m",
      "\u001b[1;31mTypeError\u001b[0m                                 Traceback (most recent call last)",
      "\u001b[1;32m<ipython-input-182-9aa45b9a33d3>\u001b[0m in \u001b[0;36m<module>\u001b[1;34m()\u001b[0m\n\u001b[1;32m----> 1\u001b[1;33m \u001b[0mtrain\u001b[0m\u001b[1;33m(\u001b[0m\u001b[0mX_train\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0mY_train\u001b[0m\u001b[1;33m)\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0m",
      "\u001b[1;32m/usr/local/lib/python2.7/dist-packages/theano/compile/function_module.pyc\u001b[0m in \u001b[0;36m__call__\u001b[1;34m(self, *args, **kwargs)\u001b[0m\n\u001b[0;32m    511\u001b[0m                     \u001b[1;32mtry\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m    512\u001b[0m                         s.storage[0] = s.type.filter(arg, strict=s.strict,\n\u001b[1;32m--> 513\u001b[1;33m                                 allow_downcast=s.allow_downcast)\n\u001b[0m\u001b[0;32m    514\u001b[0m \u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m    515\u001b[0m                     \u001b[1;32mexcept\u001b[0m \u001b[0mException\u001b[0m\u001b[1;33m,\u001b[0m \u001b[0me\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
      "\u001b[1;32m/usr/local/lib/python2.7/dist-packages/theano/tensor/type.pyc\u001b[0m in \u001b[0;36mfilter\u001b[1;34m(self, data, strict, allow_downcast)\u001b[0m\n\u001b[0;32m    167\u001b[0m             raise TypeError(\"Wrong number of dimensions: expected %s,\"\n\u001b[0;32m    168\u001b[0m                             \" got %s with shape %s.\" % (self.ndim, data.ndim,\n\u001b[1;32m--> 169\u001b[1;33m                                                         data.shape))\n\u001b[0m\u001b[0;32m    170\u001b[0m         \u001b[1;32mif\u001b[0m \u001b[1;32mnot\u001b[0m \u001b[0mdata\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0mflags\u001b[0m\u001b[1;33m.\u001b[0m\u001b[0maligned\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n\u001b[0;32m    171\u001b[0m             \u001b[1;32mtry\u001b[0m\u001b[1;33m:\u001b[0m\u001b[1;33m\u001b[0m\u001b[0m\n",
      "\u001b[1;31mTypeError\u001b[0m: ('Bad input argument to theano function with name \"<ipython-input-142-e5b726cf2525>:49\"  at index 1(0-based)', 'Wrong number of dimensions: expected 2, got 3 with shape (5, 128, 3).')"
     ]
    }
   ],
   "source": [
    "train(X_train, Y_train)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 183,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "5 (1, 128, 216)\n",
      "[(128, 3), (128, 3), (128, 3), (128, 3), (128, 3)]\n"
     ]
    }
   ],
   "source": [
    "print len(X_train), X_train[-1].shape\n",
    "print [output([x]).shape for x in X_train]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 161,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(1, 128, 216)"
      ]
     },
     "execution_count": 161,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "X_train[-1].shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 159,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(128, 3)"
      ]
     },
     "execution_count": 159,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "output(X_train[-1:]).shape"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "---"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 174,
   "metadata": {
    "collapsed": true
   },
   "outputs": [],
   "source": [
    "X_train = np.asarray([np.random.randn(*(num_channels, sequence_length*3, num_features)) for _ in range(5)])\n",
    "Y_train = np.random.randint(0, high=2, size=(len(X_train), X_train.shape[2], output_dim)).astype(np.int32)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 175,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "5 (1, 384, 216)\n",
      "[(384, 3), (384, 3), (384, 3), (384, 3), (384, 3)]\n"
     ]
    }
   ],
   "source": [
    "print len(X_train), X_train[-1].shape\n",
    "print [output([x]).shape for x in X_train]"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 176,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(5, 1, 384, 216)"
      ]
     },
     "execution_count": 176,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "X_train.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 177,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(5, 384, 3)"
      ]
     },
     "execution_count": 177,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "Y_train.shape"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 179,
   "metadata": {
    "collapsed": false
   },
   "outputs": [
    {
     "data": {
      "text/plain": [
       "(384, 3)"
      ]
     },
     "execution_count": 179,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "output(X_train[-1:]).shape"
   ]
  }
 ],
 "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": "2.7.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 0
}