{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "import torch\n",
    "import torchvision\n",
    "import torch.nn as nn\n",
    "from torch.autograd import Variable\n",
    "\n",
    "# Relative import of code from distiller, w/o installing the package\n",
    "import os\n",
    "import sys\n",
    "import pandas as pd\n",
    "import distiller\n",
    "import distiller.models as models\n",
    "from distiller.apputils import *"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Performance overview\n",
    "\n"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "model = models.create_model(pretrained=False, dataset='imagenet', arch='resnet50', parallel=False)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dummy_input = Variable(torch.randn(1, 3, 224, 224), requires_grad=False)\n",
    "df = distiller.model_performance_summary(model, dummy_input, batch_size=1)\n",
    "display(df)\n",
    "\n",
    "total_macs = df['MACs'].sum()\n",
    "\n",
    "print(\"Total MACs: \" + \"{:,}\".format(total_macs))"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Let's take a look at how our compute is distibuted:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(\"MAC distribution:\")\n",
    "counts = df['MACs'].value_counts()\n",
    "print(counts)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Let's look at which convolutions kernel sizes we're using, and how many instances:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "print(\"Convolution kernel size distribution:\")\n",
    "counts = df['Attrs'].value_counts()\n",
    "print(counts)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### Let's look at how the MACs are distributed between the layers and the convolution kernel sizes"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "scrolled": false
   },
   "outputs": [],
   "source": [
    "def get_layer_color(layer_type, attrs):\n",
    "    if layer_type == \"Conv2d\":\n",
    "        if attrs == 'k=(1, 1)':\n",
    "            return 'tomato'\n",
    "        elif attrs == 'k=(3, 3)':\n",
    "            return 'limegreen'\n",
    "        else:\n",
    "            return 'steelblue'\n",
    "    return 'indigo'\n",
    "\n",
    "df_compute = df['MACs']\n",
    "ax = df_compute.plot.bar(figsize=[15,10], title=\"MACs\", \n",
    "                         color=[get_layer_color(layer_type, attrs) for layer_type,attrs in zip(df['Type'], df['Attrs'])])\n",
    "\n",
    "ax.set_xticklabels(df.Name, rotation=90);"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### How do the Weights and Feature-maps footprints distribute across the layers:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "df['FM volume'] = df['IFM volume'] + df['OFM volume']\n",
    "df_footprint = df[['FM volume', 'Weights volume']]\n",
    "ax = df_footprint.plot.bar(figsize=[15,10], title=\"Footprint\");\n",
    "ax.set_xticklabels(df.Name, rotation=90);"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "### How the Arithmetic Intensity distributes across the layers:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "df_performance = df\n",
    "df_performance['raw traffic'] = df_footprint['FM volume'] + df_footprint['Weights volume']\n",
    "df_performance['arithmetic intensity'] = df['MACs'] / df_performance['raw traffic']\n",
    "df_performance2 = df_performance['arithmetic intensity']\n",
    "ax = df_performance2.plot.bar(figsize=[15,10], title=\"Arithmetic Intensity\");\n",
    "ax.set_xticklabels(df.Name, rotation=90);"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## ResNet20 channel pruning using SSL\n",
    "\n",
    "Let's see how many MACs we saved by using SSL to prune filters from ResNet20:"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "resnet20_dense = models.create_model(pretrained=False, dataset='cifar10', arch='resnet20_cifar', parallel=True)\n",
    "resnet20_sparse = models.create_model(pretrained=False, dataset='cifar10', arch='resnet20_cifar', parallel=True)\n",
    "checkpoint_file = \"../examples/ssl/checkpoints/checkpoint_trained_channel_regularized_resnet20_finetuned.pth.tar\" \n",
    "load_checkpoint(resnet20_sparse, checkpoint_file);"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {},
   "outputs": [],
   "source": [
    "dummy_input = Variable(torch.randn(1, 3, 32, 32), requires_grad=False)\n",
    "df_dense = distiller.model_performance_summary(resnet20_dense, dummy_input, batch_size=1)\n",
    "df_sparse = distiller.model_performance_summary(resnet20_sparse, dummy_input, batch_size=1)\n",
    "\n",
    "dense_macs = df_dense['MACs'].sum()\n",
    "sparse_macs = df_sparse['MACs'].sum()\n",
    "\n",
    "print(\"Dense MACs: \" + \"{:,}\".format(int(dense_macs)))\n",
    "print(\"Sparse MACs: \" + \"{:,}\".format(int(sparse_macs)))\n",
    "print(\"Saved MACs: %.2f%%\" % ((1 - sparse_macs/dense_macs)*100))"
   ]
  }
 ],
 "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
}