Scripts to fine-tune a depth estimation model.
APACHE-2.0 License
Ensure you're within this project root.
wget https://huggingface.co/datasets/sayakpaul/diode-subset-train/resolve/main/train_subset.tar.gz -O train_subset.tar.gz
wget http://diode-dataset.s3.amazonaws.com/val.tar.gz -O val.tar.gz
tar xf train_subset.tar.gz
tar xf val.tar.gz
pip install -r requirements.txt
(Assumes a latest stable torch
CUDA enabled environment)
huggingface-cli login
wandb login
Since the code also pushes the checkpoints to Hub, you would need to install Git LFS and configure it if not done already.
python run_depth_estimation.py --head_init --log_code
Consult the other supported CLI arguments by running python run_depth_estimation.py -h
.
The script is integrated with Weights and Biases (WandB) which can automatically keep track of the Git state of the project. So, it's recommended to first create a branch if there are any code changes, commit the changes to the branch, and then launch the experiment. This way we can easily track the changes from the WandB console.