-
Notifications
You must be signed in to change notification settings - Fork 9
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Resolve memory allocation issue + minor updates #2
Conversation
images can be load and saved in multiple shards better way of loading and saving files update method of showing examples
Hi @nrasadi, thanks for the contribution! It looks like there is an issue with the labels, the example images at the end of the notebook do not match up with their labels: can you investigate this? (just a quick guess, |
Also, this change will break the second notebook -- it expects |
Oh! Sorry. There was a little mistake. I will take care of it. |
Sure. I will fix this. |
Modify path to be compatible with the readme guide
…otebook However the models accuracy are not close to the reported ones
It looks like the validation accuracies do not match up now -- my original notebook has top-1 of 0.71256 while yours has 0.61778. |
I updated the second notebook to be compatible with the first one. However, I don't get the same results.
|
Right, can you look into this? I think they should match up |
Yes. You're right. But I don't know why! |
Yes, of course. But I think the results should be the same. Could it be due to the difference between the Keras versions? |
I'd be surprised as this is a large difference. Are images in the correct order with the labels? Can you confirm that the preprocessing is working in the same way? |
Could you please have a look at this issue? |
images are saved in uint8 rather than float32.
I double checked the notebooks. They don't seem different than the original ones. |
Something has to be different if the results are different |
I'm gonna give up. As the last try, can you find the Keras version you were used to get those results? |
Good question, I'll try to reproduce my results now. |
@nrasadi, I'm able to reproduce the results as is with tensorflow=2.3.1 and keras=2.4.3 |
Thank you very much. |
I was able to reproduce the results for VGG19 model with Tensorflow==2.3.0. |
I will gradually check the other available models in the next few days if it would help. |
Just in case, I've added a tensor-based top_k_accuracy calculation but I mistakenly haven't created a seperate PR for that. |
Thanks, great stuff @nrasadi ! Just a few changes if you don't mind and I'll be happy to merge! For
For
|
Remove tensorflow imports Hardcode data paths Define humansize function inside the notebook
Fix the behavior of percent_completed messages
I fixed the mentioned issues. Please review them. |
Awesome, thanks for the contribution! This repo will definitely be more useful to others now. Best, |
As depicted in the original notebook, the user requires at least 28GB free memory space, which may not be available in many cases.
This PR resolves this issue.
Moreover, it comes with minor updates including the way of loading and saving files and showing examples.