Icevision: Streamlit integration (local)

[lgvaz]

🚀 Feature

We need a quick demo on how to use streamlit with mantisshrimp. The first step in to get it running locally.

We should also add helpers functions to make the process easier.

[paras-jain]

I'm running late because of STOC, will push code as soon as possible.

[lgvaz]

Don't worry!

I created the issue just to make it more organized

[oke-aditya]

Should help #99 and #83

Something to get strated
https://github.com/streamlit/demo-self-driving

[oke-aditya]

Minimalistic (non-inference code) is working fine locally. We need to add model and run it. Repo is here.

You can simply give me a s3 bucket link/dropbox file link. I have made a progress bar loader too for download.

[lgvaz]

I'll try this tomorrow!!

[ai-fast-track]

Let's try to use Google Drive as a first attempt and take it from there: The notebooks + datasets will be hosted at the same place avoiding to deal with 2 separate providers.

[lgvaz]

I'm thinking the datasets are not going to be hosted by us, instead we should use the original source whenever posssible

[ai-fast-track]

Agree! I actually referring more the the weights files.

[oke-aditya]

Almost there with this repo. We just need to use the predict function of mantisshrimp. And it should display the results to the user. The other NMS and searching via the class needs a bit of work. Please try to run this repo and let me know the changes.

You can even raise a PR if something is not working or you have an improvement and update.

[lgvaz]

I ran it yesterday and it worked perfectly (apart from what we already discussed) 😄

Please let us know again when you add prediction. For now, don't bother trying to fix the deployment issues, let's first get this going locally and then we figure out

[oke-aditya]

• I fixed that small thing. Prediction will be done hopefully this week. Then we can see about the sidebar thingie of how much customization we can give to user.

[oke-aditya]

Almost done. A bit of final touch ups are left. I guess we can add this demo to read me too.

[oke-aditya]

Can you guys please test this repo?

pip install -r requirements.txt
streamlit run src/object_detection_app.py

[lgvaz]

It works!

Prediction is wrong, but I think that's caused by the low image resolution

[ai-fast-track]

It works!

[oke-aditya]

Great. Please provide me some images from Pets dataset itself so that we can show it works well.
I will polish up the app once more and then we can close this issue..

[paras-jain]

Works well!🎉
Have sent you some images from pets.

[lgvaz]

Please provide me some images from Pets dataset itself so that we can show it works well.

It's a better idea just to google some images, I used a random seed and I don't know what was in train/valid

Just google some of the classes we have in the dataset

[oke-aditya]

Yep, Paras helped me to fix it up. Now the predictions look better. We can improve the app further and refine for the sidebar UX. But I will leave this as scope for future improvement for now.
I am updating the readme to have link to this repo. We can close the issue with this.

[oke-aditya]

I Guess we can close this.