lecture_20_ssl_autoencoders_gans.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": null,
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "outputs": [],
   "source": [
    "%load_ext autoreload\n",
    "%autoreload 2\n",
    "\n",
    "%matplotlib inline"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Basic setup"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Create anaconda environment\n",
    "<br>\n",
    "```bash\n",
    "conda create -n ml python=3.7.5 jupyter\n",
    "```\n",
    "Install fastai library\n",
    "<br>\n",
    "```bash\n",
    "conda install -c pytorch -c fastai fastai\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Generative adversarial network (GAN)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Before considered different approaches for training models\n",
    "- Train model with one objective\n",
    "- Train model with multi objectives (multi-task learning) localization"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Multi-task learning\n",
    "<img src=\"images/od/localization_classification.png\" height=\"1000\" width=\"1000\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Now let's consider multi-model learning\n",
    "<br>\n",
    "We train different models simultaneously"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## GAN"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Let's we have one model which generates data from input vector $z$\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Recall auto-encoders:\n",
    "<img src=\"images/ssl/ae_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "But let's generate directly from the latent vectors - from the vectors from noise\n",
    "<br>\n",
    "But how can we be sure, that generation is \"appropriate\""
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Consider model:\n",
    "$$G:Z \\to X$$\n",
    "<br>\n",
    "Call this model the <b>generator</b> which generates (for instance ConvNet with up-sampling layers)\n",
    "To estimate that model is generates something we should estimate generated data\n",
    "<br>\n",
    "Consider the second model \n",
    "$$D:X \\to \\{0, 1\\}$$\n",
    "<br>\n",
    "Call this model the <b>discriminator</b> and train this model to classify between real and generated images\n",
    "<br>\n",
    "In other words model consumes real and generated data and discriminates them"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "In other words: discriminator model is a binary classifier\n",
    "<br>\n",
    "Now how can we train the both model?\n",
    "<br>\n",
    "Lets play a game:\n",
    "<br>\n",
    "Train generator to generate more realistic data and train discriminator to distinguish between generated and the real data"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Consider the objective:\n",
    "$$\n",
    "\\min_{G}\\max_{D}\\mathbb{E}_{x~p_{data}}[\\log(D(x))] + \\mathbb{E}_{z~p_{z}}[\\log(1 - D(G(z)))]\n",
    "$$\n",
    "<br>\n",
    "Min the generator value for generated images and max the discriminators value"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "For more details:\n",
    "<img src=\"images/ssl/gan_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Let's look into the pseudo-code:\n",
    "<img src=\"images/ssl/gan_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Results of generated images:\n",
    "<img src=\"images/ssl/gan_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Deep convolutional GANs (DCGAN)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "First big results with GANs was deep convolutional GANs:\n",
    "<br>\n",
    "It does not mean that before GANs where not using ConvNets\n",
    "<br>\n",
    "But DCGAN came up with several engineering ideas"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/dcgan_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/dcgan_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/dcgan_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/dcgan_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/dcgan_5.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/dcgan_6.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/dcgan_7.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/dcgan_8.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## More different techniques"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "More techniques where applied to the GANs\n",
    "- Label smoothing\n",
    "- Feature matching"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Label smoothing:\n",
    "<br>\n",
    "Instead of using the hard-labels\n",
    "$$\n",
    "\\{0, 1\\}\n",
    "$$\n",
    "<br>\n",
    "Use smooth labels:\n",
    "$$\n",
    "\\{0,01, 0.99\\}\n",
    "$$"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Feature matching:\n",
    "$$\n",
    "||\\mathbb{E}_{x~p_{data}}f(x) - \\mathbb{E}_{z~p(x)}f(G(z))||^2\n",
    "$$\n",
    "<br>\n",
    "Here $f$ goes for features\n",
    "<br>\n",
    "Make layer-wise feature matching is the <b>perceptual loss</b>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Progressive growing GAN"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "The training techniques which we talked about\n",
    "- Train with small images\n",
    "- After some iteration, increase image size and train more\n",
    "- Repeat the process several times\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Multi-objectives"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "As well we can add more layers and layer-blocks as \"lego\" and experiment\n",
    "<br>\n",
    "We can have different models and costs and experiments with them"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "For instance we can have auto-encoder plus GAN generator and discriminators\n",
    "<br>\n",
    "We can have different generators and discriminators\n",
    "<br>\n",
    "We can have a cycles for transfer images between domains"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Cycle GAN"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "One of the techniques of multi-generator and discriminators is the CycleGAN approach\n",
    "- We have to domains $X$ and $Y$\n",
    "- We have two generators $G:X \\to Y$ and $F:Y \\to X$\n",
    "- Two discriminators $D_X$ and $D_Y$"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "let's look at the results of different domains:\n",
    "<img src=\"images/ssl/cg_1.jpg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Cycle consistency loss:\n",
    "$$\n",
    "C = \\mathbb{E}_{x~p_{data}(x)}||F(G(x)) - x||^1 + \\mathbb{E}_{y~p_{data}(y)}||G(F(y)) - y||^1\n",
    "$$"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "The major idea is that it does not need pairing:\n",
    "<br>\n",
    "We can just have two domains and make style transfer without additional labeling"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cgpr_1.jpeg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Worth to mention"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "BiGAN when GAN is learning representations"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Use GAN for classification or as an additional loss"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Self-supervised learning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Self supervised learning is ML learning when labels are simple generated by the modification of original data, without human-in-the loop:\n",
    "- Generative models:\n",
    " - Auto-encoders when labels are the original images\n",
    "- Discriminative models:\n",
    " - When labels are the meta information about original data modification"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Self-supervised representation learning is a self-supervised learning, when model is trained for pretext task and for downstream task"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Representation learning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Deep models for feature extraction.\n",
    "<br>\n",
    "Train deep model on modified data (pretext task)\n",
    "<br>\n",
    "Use it without last layers as feature extractors and train other model on top of it (downstream task)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Pluses of self-supervised learning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "- Labeling is expensive and time-consuming, mistakes are almost inevitable.\n",
    "- Labeled data for classification contains small amount of information in comparison with images which has other different objects\n",
    "- Multi-task learning adds more difficulties and hyper-parameters\n",
    "- Multi object classifiers need more labeling"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "For deep learning:\n",
    "<img src=\"images/ssl/dl_cacke_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Generative models"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Generative models are learning to generate data from the representation (latent vectors)\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Auto-encoders"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Auto-encoders learn from the original images, they encode and reconstruct data\n",
    "<img src=\"images/ssl/ae_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Denoising auto-encoders"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "If we add augmentation and make the model extract features from images:\n",
    "<img src=\"images/ssl/dae_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Context encoders"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/ce_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/ce_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/ce_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/ce_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/ce_5.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/ce_6.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Predicting one view from another"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/pv_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/pv_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/pv_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/pv_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/pv_5.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/pv_6.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/pv_7.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/pv_8.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/pv_9.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "DeOldify\n",
    "<img src=\"images/ssl/pv_10.jpg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Relative position of image patches"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/rp_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/rp_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/rp_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/rp_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Solving jigsaw puzzles"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/jp_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/jp_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Rotation classifier"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Learn many features to understand position\n",
    "<img src=\"images/ssl/rc_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/rc_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/rc_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/rc_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/rc_5.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/rc_6.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Deep methic learning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Instead of just classify among different classes:\n",
    "$$\n",
    "P(x|c)\n",
    "$$\n",
    "<br>\n",
    "Let's train model to distinguish between classes in terms of metrics (recall the kernel methods):\n",
    "<br>\n",
    "Model outputs vectors $v$ and we train model to make $v_1, v_2$ \"closer\" (with different) metric if inputs are in the same classes and make them \"far\" otherwise"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Siamese network"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Model outputs $d$ dimensional (not-probability) vector\n",
    "$$\n",
    "m:X\\to\\mathbb{R}^d\n",
    "$$\n",
    "<br>\n",
    "The idea was use two samples: anchor and positive or anchor and negative\n",
    "<br>\n",
    "Extract vectors by the model\n",
    "<br>\n",
    "Make anchors and positives closer and anchors and negatives far"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Model consumes two images (with the shared weights) and outputs the vectors:\n",
    "<img src=\"images/ssl/siames_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Optimize for Euclidean distance\n",
    "<img src=\"images/ssl/siames_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Objective for the Siamese model:\n",
    "$$\n",
    "C = \\sum_{i=1}^{N}[(1 - y_i)\\frac{1}{2}(d_i)^2 + y_i(\\max(0, m-d_i))^2]\n",
    "$$\n",
    "<br>\n",
    "Where $m$ is a margin"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "```python\n",
    "euclidean_distance = d(v1, v2)\n",
    "loss = torch.mean((1-label) * torch.pow(euclidean_distance, 2) +\n",
    "(label) * torch.pow(torch.clamp(self.margin - euclidean_distance, min=0.0), 2)\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Then we can use the single model to generate (extract) vector from the image and compare it with existing generated vectors\n",
    "<br>\n",
    "If distance is more then margin then they belong to the different classes\n",
    "<br>\n",
    "Else they are the same classes"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Triplet loss"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "The FaceNet was the paper for the distance (metric) learning with different approach:\n",
    "<br>\n",
    "Use triplets instead of duples: anchors, positives and negatives\n",
    "<br>\n",
    "make anchors and positives closed and anchors and negatives far at the same time"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/triplet_1.jpeg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/triplet_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Objective for the triplet model:\n",
    "$$\n",
    "C = \\sum_{i=1}^{N}[||f^a_i - f^p_i||^2 - ||f^a_i - f^n_i|| + \\alpha]\n",
    "$$\n",
    "<br>\n",
    "Where $\\alpha$ is a margin"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "- We need labeled data\n",
    "- Data is tripled\n",
    "<br>\n",
    "- We don't need re-training for new class\n",
    "- It's an example of so called one-shot learning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "There where techniques for training:\n",
    "<br>\n",
    "Hard mining\n",
    "<br>\n",
    "Offline vs online hard mining, etc"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Predicting neighbouring context"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/nc_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Word embeddings with dimensionality reduction"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/we_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/we_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/we_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/we_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/we_5.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Word embeddings (Word2Vec)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/wv_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/wv_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/wv_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/wv_5.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/wv_6.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/wv_7.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/wv_8.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Contrastive predicting coding (CPC)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_5.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_6.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_7.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_8.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_9.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_10.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_11.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc1_12.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## CPC2"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_5.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_6.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_7.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_8.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_9.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_10.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_11.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_12.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_13.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_14.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_15.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_16.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_17.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/cpc2_18.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Instance discrimination"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/id_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/id_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Momentum contrast (MoCo)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/moco1_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/moco1_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/moco1_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/moco1_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/moco1_5.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/moco1_6.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## SimCLR"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/simclr_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/simclr_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/simclr_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/simclr_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## MoCo2"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/moco2_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/moco2_2.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/moco2_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Image super-resolution"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Recall denoising autoencoders\n",
    "<br>\n",
    "Let's now upscale image from the small resolution to the higher resolution"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/sr_1.jpeg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/sr_2.jpg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Use the high resolution images for the training\n",
    "<br>\n",
    "Downscale high resolution images\n",
    "<br>\n",
    "Use downscaled images as training data and original images as labels for autoencoders"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/sr_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Another approach would be:\n",
    "<br>\n",
    "Down-sample high-resolution images\n",
    "<br>\n",
    "Up-sample them with chosen interpolation\n",
    "<br>\n",
    "Use up-sampled images for training set\n",
    "<br>\n",
    "Use original (high-resolution) images as labels"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/sr_4.jpg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Perceptual loss"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Use pre-trained model:\n",
    "<br>\n",
    "ResNet50\n",
    "<br>\n",
    "InceptionV3\n",
    "<br>\n",
    "VGG16"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Calculate features of the label images on the several layers:\n",
    "$$p:X \\to Y$$\n",
    "<br>\n",
    "$$\n",
    "p = p_n\\circ p_{n_1} \\circ \\dots \\circ p_1\n",
    "$$"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Use some of this features for the estimation of the up-sampling layers\n",
    "<br>\n",
    "$$\n",
    "\\sum_{i=1}^{k}||u^l_i(X) - p^s_i(X)||^2\n",
    "$$\n",
    "<br>\n",
    "add this value to the last-layer loss ($L_2$ distance with the original image)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Model VGG (visual geometry group) was first presented at 2013 ImageNet competition and got high results (not winner ZENet was winner) but it was fast and had a simple architecture.\n",
    "VGG-16 is still used in\n",
    "- Backbone for many models\n",
    "- Perceptual loss for style transfer and auto-encoder (super-resolution, variational auto-encoder) models\n",
    "- From my experience, VGG-16 has a place, color and size invariant feature extraction property because of pooling layers and convolutional stack."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "One of the best features extractor for background jobs:\n",
    "- High dimension\n",
    "- Spark job for image clustering"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Perceptual loss:\n",
    "<img src=\"images/od/perceptual_1.jpg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "VGG-16 architecture:\n",
    "<img src=\"images/od/vgg_16_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/pl_1.jpg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/pl_2.jpg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Perceptual loss was developed for the style-transfer models\n",
    "<img src=\"images/ssl/st_1.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/st_2.jpeg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/st_3.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=\"images/ssl/st_4.png\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Questions\n",
    "\n",
    "<img src=\"images/od/questions_2.jpg\" height=\"800\" width=\"800\">"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Thank you"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": []
  }
 ],
 "metadata": {
  "celltoolbar": "Slideshow",
  "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": "3.7.4"
  },
  "toc": {
   "base_numbering": 1,
   "nav_menu": {},
   "number_sections": true,
   "sideBar": true,
   "skip_h1_title": false,
   "title_cell": "Table of Contents",
   "title_sidebar": "Contents",
   "toc_cell": false,
   "toc_position": {},
   "toc_section_display": true,
   "toc_window_display": false
  },
  "varInspector": {
   "cols": {
    "lenName": 16,
    "lenType": 16,
    "lenVar": 40
   },
   "kernels_config": {
    "python": {
     "delete_cmd_postfix": "",
     "delete_cmd_prefix": "del ",
     "library": "var_list.py",
     "varRefreshCmd": "print(var_dic_list())"
    },
    "r": {
     "delete_cmd_postfix": ") ",
     "delete_cmd_prefix": "rm(",
     "library": "var_list.r",
     "varRefreshCmd": "cat(var_dic_list()) "
    }
   },
   "types_to_exclude": [
    "module",
    "function",
    "builtin_function_or_method",
    "instance",
    "_Feature"
   ],
   "window_display": false
  }
 },
 "nbformat": 4,
 "nbformat_minor": 2
}