model_hack_notebook.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 17,
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "\n",
    "from tensorflow.python.layers import base as base_layer\n",
    "from tensorflow.python.ops.rnn_cell_impl import _LayerRNNCell\n",
    "from tensorflow.python.ops.math_ops import tanh, sigmoid\n",
    "from tensorflow.python.ops.init_ops import constant_initializer, zeros_initializer"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 9,
   "metadata": {},
   "outputs": [],
   "source": []
  },
  {
   "cell_type": "code",
   "execution_count": 19,
   "metadata": {},
   "outputs": [],
   "source": [
    "tf.reset_default_graph()\n",
    "\n",
    "class WaveRNN(_LayerRNNCell):\n",
    "    \n",
    "    def __init__(self, \n",
    "                 number_units,\n",
    "                 kernel_intialiser=None,\n",
    "                 bias_initialiser=None,\n",
    "                 reuse=None, \n",
    "                 name=None):\n",
    "        \n",
    "        super(WaveRNN, self).__init__(_reuse=reuse, \n",
    "                                      name=name)\n",
    "        \n",
    "        # Inputs must be two dimensional.\n",
    "        self.input_spec = base_layer.InputSpec(ndim=2)\n",
    "        \n",
    "        self.number_units = number_units\n",
    "        self.kernel_initialiser = kernel_intialiser\n",
    "        self.bias_initialiser = bias_initialiser\n",
    "        \n",
    "    @property\n",
    "    def state_size(self):\n",
    "        return self.number_units\n",
    "    \n",
    "    @property\n",
    "    def output_size(self):\n",
    "        return self.number_units\n",
    "    \n",
    "    def build(self, inputs_shape):\n",
    "        input_depth = inputs_shape[1].value\n",
    "        \n",
    "        # TO DO: Change kernels and biases to match the vars needed in equation (2)\n",
    "        #        from the paper.\n",
    "        \n",
    "        # Gate kernel.\n",
    "        gate_kernel_name = 'WaveRNN/gate/kernel'\n",
    "        gate_kernel_shape = [input_depth + self.number_units, 2 * self.number_units]\n",
    "        gate_kernel_initialiser = self.kernel_initialiser\n",
    "        self.gate_kernel = self.add_variable(gate_kernel_name, \n",
    "                                             shape=gate_kernel_shape, \n",
    "                                             initializer=gate_kernel_initialiser)\n",
    "        # Gate Bias.\n",
    "        gate_bias_name = 'WaveRNN/gate/bias'\n",
    "        gate_bias_shape = [2 * self.number_units]\n",
    "        gate_bias_initialiser = self.bias_initialiser \\\n",
    "                                if self.bias_initialiser is not None \\\n",
    "                                else constant_initializer(1.0, dtype=self.dtype)\n",
    "                             \n",
    "        self.gate_bias = self.add_variable(gate_bias_name, \n",
    "                                           shape=gate_bias_shape, \n",
    "                                           initializer=gate_bias_initialiser)\n",
    "        # Candidate Kernel.\n",
    "        candidate_kernel_name = 'WaveRNN/candidate/kernel'\n",
    "        candidate_kernel_shape = [input_depth + self.number_units, self.number_units]\n",
    "        candidate_kernel_initialiser = self.kernel_initialiser\n",
    "        self.candidate_kernel = self.add_variable(candidate_kernel_name, \n",
    "                                                  shape=candidate_kernel_shape, \n",
    "                                                  initializer=candidate_kernel_initialiser)\n",
    "        # Candidate Bias.\n",
    "        candidate_bias_name = 'WaveRNN/candidate/bias'\n",
    "        candidate_bias_shape = [self.number_units]\n",
    "        candidate_bias_initialiser = self.bias_initialiser \\\n",
    "                                     if self.bias_initialiser is not None \\\n",
    "                                     else zeros_initializer(dtype=self.dtype)\n",
    "        self.candidate_bias = self.add_variable(candidate_bias_name, \n",
    "                                                shape=candidate_bias_shape, \n",
    "                                                initializer=candidate_bias_initialiser)\n",
    "        \n",
    "        self.built = True\n",
    "        \n",
    "        \n",
    "    def call(self, inputs, state):\n",
    "        \n",
    "        # Create Equation (2) from paper here.\n",
    "        pass"
   ]
  }
 ],
 "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
}