The official repository for Visual Modelling.
Find the datasets used here:
- Python3 Virtual Enviroment
mkdir python_venvs
cd python_venvs
virtualenv venv
cd ..
source python_venvs/venv/bin/activate
pip install -r requirements.txt
Dataset Put the downloaded dataset here in the root directory i.e. the root dir "data" should be here in the repo root dir.
Results directory See the results directory for all checkpoints.
cd THIS_REPOs_ROOT_DIR
mkdir .results
- Modelling Experiments As simple as running scripts here. 'mixed' scripts are not currently supported.
scripts/runs/modelling
- Test Tasks Modelling tasks are a touch more complicated.
scripts/runs/tasks
The simple ones first.
- Random runs are the scripts prepended with "random_"
- Unfrozen with no pretraining are prepended with "no_" For runs that use pretrained models. You will need to add the ".ckpt" object generated by the above modelling experiments (see ".results") to the "--model_path" argument in these test-task scripts.
- Example scripts for probing pre-trained models are those that are prepended with the numer of epochs they were trained on; e.g. 137-3db_200 ... means that the pretrained model is one that was pretrained on 137 epochs of 3d Bouncing.
- The finetuning experiments are the same as the above probing ones, but further preprended with "ft_"
We host our self-output gifs for anyone to insepct. Find them here.