Resolve memory allocation issue + minor updates

[nrasadi]

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.

[calebrob6]

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, x_val = np.load(str(path_imagenet_val_dataset/"x_val_1.npy")) may should be x_val = np.load(str(path_imagenet_val_dataset/"x_val_0.npy"))?)

[calebrob6]

Also, this change will break the second notebook -- it expects x_val.npy to exist. Do you mind changing to second notebook to run on these batched samples and verify that you get the same accuracy results?

[nrasadi]

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, x_val = np.load(str(path_imagenet_val_dataset/"x_val_1.npy")) may should be x_val = np.load(str(path_imagenet_val_dataset/"x_val_0.npy"))?)

Oh! Sorry. There was a little mistake. I will take care of it.

[nrasadi]

Also, this change will break the second notebook -- it expects x_val.npy to exist. Do you mind changing to second notebook to run on these batched samples and verify that you get the same accuracy results?

Sure. I will fix this.

[calebrob6]

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.

[nrasadi]

I updated the second notebook to be compatible with the first one. However, I don't get the same results.

	Reported Accuracy (%)	Obtained Results
Top-1	71.3	61.8
Top-5	90.0	84

[calebrob6]

Right, can you look into this? I think they should match up

[nrasadi]

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.

Yes. You're right. But I don't know why!

[nrasadi]

Right, can you look into this? I think they should match up

Yes, of course. But I think the results should be the same. Could it be due to the difference between the Keras versions?

[calebrob6]

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?

[nrasadi]

Could you please have a look at this issue?

[nrasadi]

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?

I double checked the notebooks. They don't seem different than the original ones.

[calebrob6]

Something has to be different if the results are different

[nrasadi]

I'm gonna give up. As the last try, can you find the Keras version you were used to get those results?
Just to make sure nothing is changed since then.

[calebrob6]

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.

[calebrob6]

@nrasadi, I'm able to reproduce the results as is with tensorflow=2.3.1 and keras=2.4.3

[nrasadi]

@nrasadi, I'm able to reproduce the results as is with tensorflow=2.3.1 and keras=2.4.3

Thank you very much.

[nrasadi]

I was able to reproduce the results for VGG19 model with Tensorflow==2.3.0.
Top-1 = 71.248 %
Top-5 = 89.986 %

[nrasadi]

I will gradually check the other available models in the next few days if it would help.

[nrasadi]

Just in case, I've added a tensor-based top_k_accuracy calculation but I mistakenly haven't created a seperate PR for that.

[calebrob6]

Thanks, great stuff @nrasadi ! Just a few changes if you don't mind and I'll be happy to merge!

For 1. Preprocess ImageNet validation set.ipynb:

• Can you remove tensorflow imports (just to clean it up -- tensorflow isn't used here)
• Can you hardcode the paths to ILSVRC2012_validation_ground_truth.txt, meta.mat, ... to the data/ directory
• ut.humansize is not defined

For 2. Benchmark Keras pretrained models on ImageNet:

• The % complete messages are not behaving correctly.

[nrasadi]

Just a few changes if you don't mind and I'll be happy to merge!

I fixed the mentioned issues. Please review them.
Again, thanks for your really helpful notebooks and spending your precious time on this PR. 🙏

[calebrob6]

Awesome, thanks for the contribution! This repo will definitely be more useful to others now.

Best,
Caleb