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.