performance at mvn

mvn1.0.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "import torch.distributions as dist\n",
    "\n",
    "class MultivariateNormalFixed(dist.MultivariateNormal):\n",
    "    def expand(self, batch_shape, _instance=None):\n",
    "        new = self._get_checked_instance(MultivariateNormalFixed, _instance)\n",
    "        batch_shape = torch.Size(batch_shape)\n",
    "        loc_shape = batch_shape + self.event_shape\n",
    "        cov_shape = batch_shape + self.event_shape + self.event_shape\n",
    "        new.loc = self.loc.expand(loc_shape)\n",
    "        # this is where to fix\n",
    "        new._unbroadcasted_scale_tril = self._unbroadcasted_scale_tril\n",
    "        if 'covariance_matrix' in self.__dict__:\n",
    "            new.covariance_matrix = self.covariance_matrix.expand(cov_shape)\n",
    "        if 'scale_tril' in self.__dict__:\n",
    "            new.scale_tril = self.scale_tril.expand(cov_shape)\n",
    "        if 'precision_matrix' in self.__dict__:\n",
    "            new.precision_matrix = self.precision_matrix.expand(cov_shape)\n",
    "        super(dist.MultivariateNormal, new).__init__(batch_shape,\n",
    "                                                self.event_shape,\n",
    "                                                validate_args=False)\n",
    "        new._validate_args = self._validate_args\n",
    "        return new"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "loc = torch.randn(3)\n",
    "A = torch.randn(3, 3)\n",
    "cov = A.matmul(A.t())\n",
    "d1 = dist.MultivariateNormal(loc, cov).expand([1000, 3])\n",
    "d2 = MultivariateNormalFixed(loc, cov).expand([1000, 3])\n",
    "x = torch.randn(3)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "70.2 ms ± 3.09 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)\n"
     ]
    }
   ],
   "source": [
    "%timeit d1.log_prob(x)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "488 µs ± 28.8 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)\n"
     ]
    }
   ],
   "source": [
    "%timeit d2.log_prob(x)"
   ]
  }
 ],
 "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": "3.6.6"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}