Multiple images at input

[VladVin]

There's a need to send multiple images per request item, for example, when we are running a model for image retrieval task which accepts images of a product from different angles of view. The amount of images of the product is arbitrary. So, it looks like we can make a request object with a list of tensors of different sizes

[wselfjes]

Actually, there is a possibility to send multiple views images as an input.
I wrote a simple model that accepts multiple image and return embedding and it is working
https://gist.github.com/wselfjes/379a604f67d662b775021390b36519c1

[VladVin]

Thank you for the sample. Yes, I understand that we can stack images into one tensor, but the problem is in different sizes of images that might come (btw, this was mentioned in the comment above). For example, different photos of the same product.

Currently I see two options:

1. Resize image to common size which is not good because it assumes the middleware that works with image tensors. Also, this might reduce the quality of a model because usually there's a different preprocessing applied in each model to the original image.
2. Pass tensors as a meta parameter. This looks like a hack and I think it would be better to support multiple images per item by design

[wselfjes]

Another solution is VID2IMG mode. You can build the stateful model that wait for the flag “end” and then performs inference.