A walk with fastai2 - Vision - Study Group and Online Lectures Megathread

muellerzr · December 14, 2019, 9:37am

Just did looks to be good, only bit I read was about mobile streaming I’d need over 1k subscribers or more

iamgianluca · December 14, 2019, 1:15pm

Very exciting! I’m particularly interested in the tabular part.

Moody · December 15, 2019, 11:02am

You never know. I never dream of my local study group grows over 2.3K members and is still growing.

muellerzr · December 15, 2019, 2:03pm

You’ll certainly be more intrigued by the fourth lesson (in the tabular block)! I revised it to try to let us experiment with the new architectures that have come out over the last year or so (NODE, TabNet, etc).

jankelowitz · December 15, 2019, 9:32pm

@muellerzr excited to watch this. Do you happen to have a step by step for fastai v2 for GCP? I followed the steps outlined here:https://course.fast.ai/start_gcp.html and successfully setup the server with v1 but can’t seem to wrap my head around getting v2 running.

much appreciated and looking forward to class.

muellerzr · December 15, 2019, 9:33pm

I don’t use GCP, so I’m unsure about that. What have you tried so far?

jankelowitz · December 15, 2019, 9:37pm

I tried conda install -c fastai -c pytorch fastai but couldn’t find where fastai v2 would be in the vm terminal.

and am confused about the other suggestion of this. because I don’t see where fastai2 would be in the terminal
cd fastai2
conda env create -f environment.yml
source activate fastai2

muellerzr · December 15, 2019, 9:39pm

You’d need to either git clone the fastai2 repo (to get fastai2) or do a pip install fastai2. Try the install directions here:

github.com

fastai/fastai2/blob/master/README.md

<!--

#################################################
### THIS FILE WAS AUTOGENERATED! DO NOT EDIT! ###
#################################################
# file to edit: nbs/index.ipynb
# command to build the docs after a change: nbdev_build_docs

-->

# Welcome to fastai v2

> NB: This is still in early development. Use v1 unless you want to contribute to the next version of fastai


## Installing

You can get all the necessary dependencies by simply installing fastai v1: `conda install -c fastai -c pytorch fastai`. Or alternatively you can automatically install the dependencies into a new environment:

```bash

This file has been truncated. show original

jankelowitz · December 15, 2019, 9:55pm

I guess my confusion with this, is where I am supposed to be changing the directory into fastai2, or am I supposed to create it first?

Thanks for the help.

muellerzr · December 15, 2019, 9:57pm

From what I can read (@sgugger correct me if I’m wrong here):

Clone the repo into some base directory to run out of
cd and Conda activate

sgugger · December 15, 2019, 9:59pm

The conda install of fastai2 is not super well tested, I would strongly encourage using a developer install using pip for now.

jankelowitz · December 15, 2019, 10:11pm

cool that did the trick! thanks @sgugger and @muellerzr for the help.

For anyone other newbies out there, I followed this:

pip install packaging
git clone https://github.com/fastai/fastai2
cd fastai2
pip install -e .[dev]

muellerzr · December 15, 2019, 10:12pm

I’ll also add that in the Wiki as GCP instructions

Edit: done. Thanks for trying that out @jankelowitz!

jankelowitz · December 15, 2019, 10:14pm

no problem, going through dev pets tutorial to make sure I actually did it right

howtodowtle · December 16, 2019, 2:21pm

@muellerzr Thanks for the initiative! I’m in, although I’ll be watching the recordings since 5 pm CST is midnight CET. But I already blocked my Thursday nights for that.

Pak · December 16, 2019, 4:00pm

That is supercool. Looks like I’m living under the rock and still under the impression that no architectures except fully-connected layers is applicable here.
That do you recommend to try at first NODE, TabNet or maybe something else? I don’t think I will be able to overcome all of them
What worked best in your cases, if you was already able to try it (or what is your intuition on what will play better/have more potential)

muellerzr · December 16, 2019, 4:03pm

@Pak NODE looks to be the “easiest” (I’m reworking my implementation to solve some headaches). There’s some cost benefit with both (that I plan on going over). TabNet I’m yet to play with but it’s one that I saw that looked interesting. I have my own intuition about a paper that came out in which the data was embedded in a 2d space (aka they wrote the variables on a picture) and I think we can push that further, I’ll be testing that out this week or so (and if it works you’ll see it here, if not you won’t!)

Pak · December 16, 2019, 4:16pm

Am I right that you mean these notebooks https://github.com/muellerzr/Practical-Deep-Learning-for-Coders-2.0 (tabular in this repository)?
Awesome I will definitely watch your github on that.
As well as will watch your stream.

muellerzr · December 16, 2019, 4:18pm

Not yet, I’m working on them now. I’ll announce the course repo once it’s closer (so there’s little questions on them beforehand and cause I’m in the middle of making a few of them right now). I don’t have have the Tabular implementations done yet (I’m finishing vision block this week). I believe there’s an old NODE thread where I posted an example (very very messy example) with fastai v1

pnvijay · December 17, 2019, 12:11pm

@muellerzr I was trying the second notebook 02_Custom_Image_Classification.ipynb via colab and had the following error.

The file is not being read as a file but as a string. I tried with the following

file = open(path/files[i])
download_images(file, path/folders[i], max_pics=50)

This time it downloads the files but none of them are proper images and therefore the verify_images deletes all of them. I can attach the url_text files here if you want. Wanted your help to know if it is an issue with the code or the way I am doing things here.