Lesson 1 - Official topic

Part 1 (2020)
104

What are the fastai team’s thoughts on contenders for different types of stacks? The previous class was Swift and Tensorflow 2.0 has been released recently.

1 Like
105

@Jess The Auto-setting does a great job at fixing the res as well.

Source: Experience of studying my first fastai course from a choppy college dorm wifi

1 Like
106

Right here: https://www.amazon.com/BlueParrott-B450-XT-Canceling-Bluetooth-Headset/dp/B01M3MMFA5

4 Likes
107

At fast.ai, we were early adopters of PyTorch. Here is Jeremy’s blog post from 2017 announcing when we switched from TensorFlow to PyTorch, and our reasons for doing so:

2 Likes
108

We will learn how to turn one model in a web app, but the main focus of fastai is on training, so the rest will be student projects.

6 Likes
109

I think there’s gonna be lecture parts talking about putting fastai models into production. How deep are you looking to go though? If it’s just one model as a webapp, the course should cover that.

If you need to deploy in PyTorch alone, there’s a few forum posts that can help

2 Likes
111

I took fastai last year. Definitely work on projects! I spent too much time on theory as well. Find a project and work on that. You will eventually get to a problem where you are literally realizing that you need softmax instead of sigmoid, I need a different type of normalization, I need more weight decay, I can split linear layers to save memory in the adam optimization step, etc. Inspect everything when you are running into problems and it will just come naturally as you need it.

2 Likes
113

@ganesh.bhat

As promised in Youtube thread, article from Jeremy on why they switched to Pytorch.

3 Likes
114

@rachel, in a nutshell, what is the main difference between Tensorflow and Pytorch?

115

For now, PyTorch is the clear winner in terms of ease-to-use and flexibility. Swift is not there just yet, but we are still very excited about it.

4 Likes
120

PAY A LOT OF ATTENTION TO WHAT JEREMY JUST SAID (yes I know I’m yelling).
Do not try to set up your own machine yet. Use one of our free/cheap options for the duration of the course. Don’t waste time on setup right now, spend time training models.

18 Likes
123

100% yes, I find many people in the study group I run all year long making the mistake of trying to use their own machine. Dealing with the complexities of getting an environment setup is a good way to get way behind on the course.

4 Likes
124

Would you recommend just using e.g. free Colab now vs Google Compute Platform? (I’m in the process of setting up GCP and was planning to complete this after today’s course.)

Thanks :slight_smile:

125

For folks who are interested in being able to switch to a different GPU (should they have the luxury of doing) and also print out the version of fastai2 they are using (helpful as the framework evolves esp. as you encounter errors that have been addressed in later versions)

import torch
from fastai2 import __version__ as fa2_version

device = 'cuda:1'
torch.cuda.set_device(device)

print(f'fastai version: {fa2_version}')
print(f'Using cuda:{torch.cuda.current_device()} - {torch.cuda.get_device_name()}')

# fastai version: 0.0.13
# Using cuda:1 - GeForce GTX 1080 Ti

This is right at the top of every notebook I work on.

5 Likes
126

I’m not Sylvain but I’ve used Colab since the dev dev version of fastai2 and highly recommend it :slight_smile:

1 Like
128

Is there a benchmark about the use of :
Azure / Colab / etc to see which one perform better?

129

Anything you feel comfortable with (and don’t spend more than an hour trying to setup just now). The whole point of the top-down approach is to get you to do things first, then get down the rabbit hole.

So if you get frustrated during the setup of GCP try one of the easier approaches (some are free) and then come back to it later.

4 Likes
130

Thanks - I’ve played around in Colab before and it’s pretty intuitive. So I might just stick with that then. Any benefit of doing paid version now or is free good enough? Thx

132

Thanks – I was making good progress on setting up GCP but got distracted with work :slight_smile: haha. So perhaps I’ll just focus on Colab for now and come back to GCP later.

133

I immediately jumped on the paid and have no complaints. The free version acts like the old Colab just you have a small chance of not getting a GPU. I decided paid because longer instances, and higher storage and better GPU’s (you mostly get K80’s without Pro)