Lesson 3 In-Class Discussion ✅

(Jeremy Howard (Admin)) #1

Use this topic for in-class discussion. Live stream and other links:

Lesson 3 Advanced Discussion:
https://forums.fast.ai/t/lesson-3-advanced-discussion/29750

5 Likes

📝 Deep Learning Lesson 3 Notes
Fast.ai v3 2019课程中文版笔记
📝 Deep Learning Lesson 3 Notes
(Rick N) #3

I’ve noticed that most of the examples and most of my models result in train_loss > valid_loss. What are the best ways to correct that?

I should add that this still happens after trying many variations on # of epochs and lr

15 Likes

(Will) #4

Ooops initially misread your > symbol. train > valid is a sign of underfitting and most likely you could train for another epoch or so.

1 Like

(Rick N) #6

actually I’ve run 16-24 epochs in some cases and it rarely seems to get there (meaning to train_loss < valid_loss. I’m referring mainly to lesson 2

1 Like

(Abhishek Sharma) #9

You should try increasing your learning rate. It must be too small.

1 Like

(Rick N) #14

nope I’ve tried many learning rates also

0 Likes

(Jesús Pérez) #16

The FileDeleter still exist in the new library version?

I was running lesson 2 (download images notebook) with the new fastai version and got this error. Does anyone has any idea?

1 Like

(Sven Becker) #24

I think it is called ImageDeleter now

0 Likes

(Brian Holland) #25

What version of FastAI are you running?

0 Likes

(Fernando Melo) #27

Questions about nlp Imdb notebook
1- Let’s say I have already prepared train and valid datasets? how to pass them to tokenizer/numericalizer?
2- How to set a different language to the Tokenizer? Like lang=”pt”?

1 Like

(Jesús Pérez) #28

version 1.0.21

0 Likes

(windalfin) #30

It is updated with ImageDeleter function

https://forums.fast.ai/t/lesson-2-further-discussion/28706/62?u=win

2 Likes

(Ravi S. Rāmphal) #33

Is there a repo to contribute to http://videos.fast.ai/?

5 Likes

#38

Password is deeplearningSF2018 (do not share outside the forum group)

9 Likes

(Matthew Rosenthal) #41

Is there a suggested way to to prediction on batches of images? Not from the web but just scripted?

1 Like

(Md Mofijul Islam) #44

But I think @rrmphal is looking for a code repo to contribute.

0 Likes

#45

Is the recommendation for Andrew Ng’s original Coursera course or the new @ deeplearning.ai or both?

2 Likes

(Gregory PAKOSZ) #46

The original one

0 Likes

(Zach) #47

PRs welcome: https://github.com/zcaceres/fastai-video-browser

14 Likes

(hector) #48

The fastai ML course covers random forest in detail…can we expect ML part2 covering GBMs?

5 Likes