search

csndl-iitd / visual-attention

Distinguish bottom up vs top down visual attention based on eye tracking
MIT License
0 stars 0 forks source link

Run the Deepgaze III code, first on the original dataset then on our dataset #6

Open shreyagakhar opened 1 month ago

shreyagakhar commented 1 month ago

Current progress: trying to run the original code as it is but it is giving a lot of errors. Trying to resolve them

shreyagakhar commented 1 month ago

GitHub link for the model: https://github.com/matthias-k/DeepGaze.git

gsaurabhr commented 1 month ago

Current progress: trying to run the original code as it is but it is giving a lot of errors. Trying to resolve them

Try with a fresh install of python 3.9 (from python 3.10 onwards collections changed to collections.abc). There might be some other similar version issues. I have raised an issue in the original repository requesting the authors to add version information. In the meantime, just try with python 3.9

How to use a different python version in colab? https://gist.github.com/kargaranamir/e0b7910fed0a3189563d9254c7a2c439

shreyagakhar commented 1 month ago

We figured it out using the latest python version too.

Colab link: https://colab.research.google.com/drive/19KUQ-KwD-6gHLMdhFjDUT-v2AulBEB5T?usp=sharing

gsaurabhr commented 1 month ago

Extracting features / activations from intermediate layers: https://kozodoi.me/blog/20210527/extracting-features

shreyagakhar commented 1 month ago

Deepgaze III original paper: https://jov.arvojournals.org/article.aspx?articleid=2778776#247364950

shreyagakhar commented 3 weeks ago

Colab notebook link for running the layers of Deepgaze 2e: https://colab.research.google.com/drive/1heJfnjyge4XQ2MZbKvTrH6zvJwFQsSVQ?usp=sharing

The layers are running successfully, but I am encountering some issues while trying to use hooks to extract the activations from each layer; trying to resolve the same.