This message was deleted BentoML #ask-for-help

Join Slack

This message was deleted.

# ask-for-help

Slackbot

04/12/2023, 4:05 PM

This message was deleted.

Chaoyu

04/12/2023, 10:25 PM

I’m thinking about building a BentoML demo project for this actually!

Chaoyu

04/12/2023, 10:26 PM

I haven’t got a chance to try it myself tho, but would love to chat

Sylvain LeBeux

04/19/2023, 3:22 PM

@James-Andrew Sarmiento @Chaoyu I've implemented a version of SAM for the masks generation but unfortunately, I am running into a weird issue where the service works if I run it locally but starts failing silently once I containerize it. Any further tips appreciated since I am out of ideas here. Below are the

bentofile.yaml

and

service.py

I am using for this The error I am getting is the server disconnect after a few seconds into the request. Nothing more in

--debug

mode, it just fails silently and server restarts. ah, and also two rather annoying hacks I had to put in place for it to work. • My image shape is (550, 512, 3) (W, H, C) it works fine with pytorch but I have to setup

max_batch_size

to 1000 because bentoml thinks the first dim is the batch dim (àla tensorflow) • The return value of SAM is a list of dict containing all info about the objects masks found in the image, and nested in there are numpy array for the segmentation values. This list is thus not JSON serializable, hence my conversion .., but I have to do it back again on the receiving side. There is obviously a better version I am not aware of as of now. Thanks for reading

Sylvain LeBeux

04/19/2023, 3:22 PM

The `service.py``

Sylvain LeBeux

04/19/2023, 3:23 PM

the `bentofile.yaml``

Sylvain LeBeux

04/19/2023, 3:23 PM

~service: "service:sam_masks" labels: owner: "PLACEHOLDER" stage: dev include: - "*.py" docker: system_packages: - git - libgl1-mesa-glx - libglib2.0-0 python: packages: - torch - torchvision - numpy - opencv-python - git+https://github.com/facebookresearch/segment-anything.git~

Sylvain LeBeux

04/19/2023, 3:29 PM

If I do : $ bentoml serve service.py It works and the processing of my image takes about 45 seconds

Sylvain LeBeux

04/19/2023, 3:50 PM

UPDATE: it actually works as long as you allow docker sufficient resources I currently have 12Gb Memory and 2Gb Swap configured

🙌 1

Chaoyu

04/19/2023, 10:37 PM

@Sylvain LeBeux that’s great to see! we were actually thinking about building an official demo project for SAM, along with a couple other example projects we are building. would you be interested in helping to contribute this?

Sylvain LeBeux

04/20/2023, 1:28 PM

Hi @Chaoyu. Sure ! I'd be glad to help at my own level, let me know what it would entail so we could further discuss the process. Cheers !

Tim Liu

04/20/2023, 7:59 PM

@Floris Alexandrou Some insight into deploying SAM

🙌 1

Chaoyu

04/21/2023, 10:33 PM

@Sylvain LeBeux let’s chat!

Sylvain LeBeux

05/01/2023, 12:45 PM

@Chaoyu Sorry, I was on vacation last week ...

Tim Liu

05/01/2023, 6:57 PM

Hi @Sylvain LeBeux! Our Head of Engineering @Sean actually put together this example to deploy "Segment Anything". We're likely going to promote it in the coming weeks, but thought you might be interested in a sneak peak

Sylvain LeBeux

05/01/2023, 7:12 PM

Hi @Tim Liu. Sure, let me know of the details and I'll be happy to comment or help

Tim Liu

05/01/2023, 7:18 PM

@Sylvain LeBeux Whoops, so sorry, I actually forgot to include the link 😂 https://github.com/yuqwu/Replace-Anything

Sylvain LeBeux

05/01/2023, 7:19 PM

Thanks

Sylvain LeBeux

05/01/2023, 7:25 PM

Nice, you actually merged SAM with stable-diffusion ! There are definitely some features I was missing, such as cuda computation. Thanks a lot for sharing and nice job @Sean!

🙏 1

Sylvain LeBeux

05/01/2023, 7:55 PM

Funny enough, I just stumbled upon this a few minutes ago, I guessed you would find it interesting https://github.com/IDEA-Research/Grounded-Segment-Anything

👀 1

Open in Slack

Previous Next