search

erikalu / class-agnostic-counting

63 stars 14 forks source link

Sample data to perform few-shot training #5

Closed darsh10 closed 3 years ago

darsh10 commented 4 years ago

Hey Authors,

I very recently learned about this implementation and I find it very exciting for medical research applications. Would it be possible for the authors or someone to share what sample (fine-tuning) training data looks like?

Thanks, Darsh.

erikalu commented 4 years ago

Hi Darsh, if I'm understanding you correctly, and you want an example of the labeled data we used for fine-tuning the model, take a look at the cell dataset here: http://www.robots.ox.ac.uk/~vgg/software/cell_detection/. We used it for one of our fine-tuning experiments.

darsh10 commented 4 years ago

Hi Erika,

I am actually wondering how I can use my own (custom) labels. Is that possible? I am fine if I need to train a model from scratch.

Thank you, Darsh.

On Sat, 1 Aug 2020 at 5:20 PM, Erika Lu notifications@github.com wrote:

Hi Darsh, if I'm understanding you correctly, and you want an example of the labeled data we used for fine-tuning the model, take a look at the cell dataset here: http://www.robots.ox.ac.uk/~vgg/software/cell_detection/. We used it for one of our fine-tuning experiments.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/erikalu/class-agnostic-counting/issues/5#issuecomment-667588502, or unsubscribe https://github.com/notifications/unsubscribe-auth/ACXM57YR7TD24ABLSLEIRM3R6SBIHANCNFSM4PMPT66Q .

erikalu commented 4 years ago

Yes, you can use your own custom data to finetune/adapt the pretrained model. Just download the pretrained model weights and set up your custom dataset in the same way as the example datasets. For minimal changes to the code, you'll want to modify src/data_loader.py and also create a new .npz file with the filenames. Please take a look at that data_loader.py file to see what is expected of the dataset format.