Batching on .extract_faces to improve GPU utilization and throughput

[galthran-wq]

Tickets

#105

What has been done

With this PR, .extract_faces can accept a list of images and processed them in a batch

How to test

make lint && make test

[galthran-wq]

several problems:

• currently input images have to be of the same size (hence resize in the test), otherwise it raises an uninformative error. maybe do the same resizing in the .extract_faces, but I am not quite sure if it would hurt performance
• I am not yet unable to reproduce a performance improvement in colab, but here is what my local 4090 says when processing 1k images:

Batch Size: 1, Total Time: 101.9
Batch Size: 10, Total Time: 61.3
Batch Size: 20, Total Time: 59.9

I'll update on this

@serengil what do you think?

[serengil]

several problems:

• currently input images have to be of the same size (hence resize in the test), otherwise it raises an uninformative error. maybe do the same resizing in the .extract_faces, but I am not quite sure if it would hurt performance
• I am not yet unable to reproduce a performance improvement in colab, but here is what my local 4090 says when processing 1k images:

Batch Size: 1, Total Time: 101.9
Batch Size: 10, Total Time: 61.3
Batch Size: 20, Total Time: 59.9

I'll update on this

@serengil what do you think?

IMO, batch should not be supported at all :) You should raise a ticket first to discuss this.

[galthran-wq]

why not if it improves performance without hurting anything else (I hope)?

[serengil]

why not if it improves performance without hurting anything else (I hope)?

My personal opinion, I will not close this because of the time you spend

[serengil]

would you please run this test for 2 images where 1st one has 1 face and 2nd one has many faces.

i want to see result is size of 2 in this case, and 2nd item of that list is having many items.

you can add this image to dataset folder and use it.

[serengil]

do you mind to add tests for detect_faces also? at least its return type for single item and many items.

[galthran-wq]

this is the key thing to make batching work. It is required to be able to concat images' tensors. It will not currently work it we do not call this function beforehand.

basically, it adds black strips to each image to be max width and max height of the batch (therefore all of the same size)

ideally, I think, we'd like for this operation to happen behind the scenes on .extract_faces and the results (facial area, eyes, nose) be accordingly transformed after prediction in accordance with the original image.