londons_explorer

joined 11 months ago
[–] londons_explorer@alien.top 1 points 9 months ago

Some people would love the opportunity to have their own self contained project with no hard deadline and get to learn as they built it. Many opensource projects are similar for example.

But it's only right for a certain kind of person, and it sounds like it isn't right for you. Management should have noticed that and reassigned you. The fact they didn't is a bit of a red flag. (although note that it's really hard to judge progress in software development, and therefore really hard to judge if someone is taking 10x longer than needed for a task, especially a one-person task like this).

Here are your options:

  • Quit and move. You are under no obligation to finish these projects before taking on a better job at another company. Go looking.

  • Speak to your manager saying you want to bring another person on to help with these tasks. Suggest hiring someone new if nobody else in the org looks spare. Then you get to progress up to a 'team lead' type position on your two tasks. You can hire someone who has the skills you are lacking.

  • Speak to your manager and say you don't have the necessary skills to complete these tasks without guidance. Ask for a mentor to be assigned, and do 'pair coding' for a few hours a week. You'll learn a lot and it might give you the knowledge needed to finish the job.

[–] londons_explorer@alien.top 1 points 10 months ago

Use pytorch - tensorflow is pretty much dead.

I'd also use google colab (the free version is fine).

Start from someone elses colab that already uses the pretrained models you need, and then nearly everything is already set up for you, and you won't spend a day wrestling with GPU drivers.

closest

L2 norm is fine, yes, although you might get better results with cosine similarity.

If I were you, I would start here:

https://colab.research.google.com/github/huggingface/notebooks/blob/main/examples/image_similarity.ipynb

Hit the little play buttons to edit and run the code yourself.

[–] londons_explorer@alien.top 1 points 10 months ago

For detecting "bogus image from the internet", just search your own database of submitted images. If you see the same image (either by md5 or by very similar embedding vector), then it's probably one from the internet. Before long, you'll have collected all the ones users can easily find from Google Images etc. (I assume this is to prevent fraud where a user says "look, my packet was damaged", when in fact they just searched for a picture of a damaged pack online)

[–] londons_explorer@alien.top 1 points 10 months ago (2 children)

using a pretrained model like CLIP and just train the final

Yes, do this. Should be quick and easy, and you should get decent results with only ~10 examples per class.