ImageWang & Self-Supervised Learning

ai_padawan · August 8, 2020, 7:19am

Hey everyone, so I was looking at the ImageNette, ImageWoof repo and came across an ImageWang dataset for self-supervised learning. There’s also some implementations for fastai v2:

github.com

WAMRI-AI/imagewang/blob/d575c7e6d531ec14ec10e545c6e672f87e3d5953/03_ImageWang_ContrastLearning_final_224.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Image网 Submission `224x224`"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "This contains a submission for the Image网 leaderboard in the `224x224` category.\n",
    "\n",
    "Inspired by the paper: [A Simple Framework for Contrastive Learning of Visual Representations](https://arxiv.org/abs/2002.05709)\n",
    "\n",
    "Code run on Titan RTX\n",
    "\n",
    "In this notebook we:\n",

This file has been truncated. show original

It looks similar to SimCLR?

Has anyone done any work on using unlabeled images to pretrain the backbone? The one thing I’m stuck on is how/why is the output set to 128 in that notebook. I thought the contrastive loss thing is like a softmax, but it “softmax-es” various patches of images as to how similar they are to each other. So shouldn’t the output be 2?

shimsan · August 10, 2020, 5:08am

Maybe this would be of interest to you:

ai_padawan · August 10, 2020, 7:38am

Thanks, but this tutorial is for rotation. I’m referring to the contrastive loss method. In that method they have 8 rotations, so 8 outputs make sense. I still don’t get where the 128 outputs is coming from in the SimCLR fastai repo…