There is no value in “test” dataset which is not labelled 
you can’t test the model using it because you won’t know error if you don’t have correct labels
Yes it is.Tried some basic tutorials on fastai docs yesterday itself.
Validation set
2 Likes
When using transfer learning, how important is it that your data be normalized in the same way as the data used in the original training set?
1 Like
Thanks for posting!
1 Like
Try adding num_workers=0. It’s a know problem in pytorch …
data = ImageDataBunch.from_name_re(path_img, fnames, pat, ds_tfms=get_transforms(), size=224, num_workers=0)
4 Likes
Are we going to go deeper into Pytorch or is it going to be more about the fastai library?
4 Likes
Question: according to this article inceptionv4 is best, and better than resent, why not inception was chosen for fastai library?
17 Likes
It supports it. I’ve installed fastai v1 and trained CIFAR yesterday using the example of Jeremy and get the 94% accuracy but it took about 1hour.
2 Likes
It’s due to a major outage of GitHub yesterday & has just recovered but fast.ai site didn’t. They are aware and it will be fixed soon.
Thank you very much for the beautiful FastAI package, which can allow us to do image classification very fast/easy. But it seems you linked the models with your data, so it is very easy to run. So my questions is:
Do you have any suggestions and tutorials for us to apply your package for our own problems?
Thank you very much.
2 Likes
Can you please explain more about the RESNET Part and how its is working? Why resnet there are many more models available.
Thanks
The fastai library has been deployed to production at several companies, including Github and at a Fortune 500 company.
What do you mean exactly? Getting the dataset or running the model? The model works for any image dataset with arbitrary categories.
Is there a catalog of existing models for all different types of problems?
1 Like
Stay tuned for the rest of this course
3 Likes
We do have the option of using inceptionv4, resnet is just a standard cnn architecture. One other reason might be that, we would be building resnet form scratch, which is easier to understand and code compared to inceptionv4.
2 Likes
Yeah, but Jeremy are those numbers really an apples to apples comparison? I mean are you comparing the same Keras models to fastai models? If so, why is fastai any faster? It’s the same model?
5 Likes