Pardon for the direct tag @jeremy and @yinterian, just a quick question: last year’s lesson 7 dealt with the fish classification + regression for bounding box estimation (plus heatmaps and so on).
I see there’s a fish.ipynb in the repo, but it’s not been updated in 2 months, and for L7 I see there’s some work on cifar10 planned. Any chance we will tackle fisheries again tonight?
1 Like
Nope no fish.
1 Like
Damn, was hoping some cool tricks on bounding boxes with Pytorch from you! Well, CIFAR10 it is then! Thanks for answering!
Bounding boxes will be in part 2 - see you there! 
14 Likes
such a teaser! 
1 Like
That I am. But - I think you’d be more than ready to have a go at this yourself. @yinterian has already done all the heavy lifting to make transforms work with coordinates (which I believe make fastai the first library to support this!)
5 Likes
oh great to know! - hurt my brain when affine transforming the BB coordinates on my synthetic data 
Thanks Jeremy, I’m planning to try fastai for these things as soon as Pytorch’s Windows support is fast, stable and reliable as with Linux!
Thanks for the tease, I’ll apply for sure for a fellowship for Part2!
2 Likes
Pytorch for windows seems OK now - I ran all of lesson 1 yesterday in it.
1 Like
Interesting @jeremy, did you need to make any changes to the code? Have you installed with CUDA9 or CUDA8?
I tried running the Pytorch Windows (0.2) back when we were doing lesson 1 and it gobbled up all of my memory and crashed, whereas doing the same from my WSL, though not supporting the GPU, it worked OK.
The error I got was “DLL load failed: The paging file is too small for this operation to complete”.
So perhaps the latest iteration has fixed that - or perhaps it was a setting that I needed?
See: Windows 10 Installation Notes (Windows command and WSL bash)
CUDA 9. No changes. 0.3 on windows is totally different to 0.2 for memory and performance
Thanks @jeremy. Is Cuda 9 required? If so, is it compatible with other libraries, especially Tensorflow? I am a bit nervous about upgrading from 8 unless its absolutely required for Pytorch Windows…
CUDA 9 is not required.
wow thats impressive, I remember hating that I couldn’t use augmentation with bounding boxes (in Keras)…kudos to @yinterian for figuring this out!
I was actually about to start working on my own solution to this, so this is great to hear!
Are there any examples in the fastai library right now that demonstrate this in action?