Batching on .represent to improve performance and utilize GPU in full
#1433
Conversation
|
Would you please write unit test to make it clearer while reviewing? |
| @@ -81,3 +83,49 @@ def test_max_faces(): | |||
| max_faces = 1 | |||
| results = DeepFace.represent(img_path="dataset/couple.jpg", max_faces=max_faces) | |||
| assert len(results) == max_faces | |||
|
|
|||
|
|
|||
| @pytest.mark.parametrize("model_name", [ | |||
There was a problem hiding this comment.
please do the with only one model - e.g. Facenet
I excluded some of those models from test otherwise it will take too long tests to be performed in github
There was a problem hiding this comment.
it's just that some models have custom forward logic, which I also had to tweak a little bit.
maybe still keep those models(Dlib, SFace, VGGFace), along with some keras one, like, Facenet?
There was a problem hiding this comment.
okay but some are optional (e.g. dlib), these are not installed in github actions. so, your tests will be failed.
|
There are some linting issues broken the actions |
|
LGTM Thank you for your contribution |
| "dataset/img2.jpg", | ||
| "dataset/img3.jpg", | ||
| "dataset/img4.jpg", | ||
| "dataset/img5.jpg", |
There was a problem hiding this comment.
if you add couple.jpg here, there are 6 input images but in the response we will have 7 items.
the bad part, we cannot understand that which image has 2 faces. i am creating a PR to store input image's index in the response payload.
we may consider to have List of List of Dict response type for batch inputs in the future.
There was a problem hiding this comment.
I think it is a good idea to have List of List of Dict. I could make a PR now, or perhaps later, when batched detection is merged, because I also had in mind to use batched detection for .represent (now it is done in the for loop)
There was a problem hiding this comment.
I will do the initial changes. PRs are always welcome!
There was a problem hiding this comment.
i merged a workaround PR for this
There was a problem hiding this comment.
will optimize this because if batch size is long, this approach gives O(n^2) complexity.
I plan to do something with dict, which decrease the complexity O(n)
Tickets
#1101
What has been done
With this PR,
.representis able to accept a list of paths/numpy arrays and process them all in a batch.How to test
I've made a collab notebook
https://colab.research.google.com/drive/1bV0yyrdT1a0a4dyemoeL5xf28w3ql1fd#scrollTo=lVndmFF5Kls7
which shows the >10x performance improvement and also the fact, that with batch_size=1 the GPU is almost not utilized