CSCI-GA-2572-Deep-Learning-Final-Competition-Dragonfruit

Team12: DragonFruit - Yuanhe Guo, Jianing Zhang, Haotong Wu

Ranked 2nd in 2024 Spring final competition

Getting Started

Step1: Clone Repo

git clone https://github.com/RicercarG/CSCI-GA-2572-Deep-Learning-Final-Competition-Dragonfruit.git

cd CSCI-GA-2572-Deep-Learning-Final-Competition-Dragonfruit

Step2: Prepare Dataset

Put dataset inside dataset folder, with structure being

CSCI-GA-2572-Deep-Learning-Final-Competition-Dragonfruit
└── dataset
    ├── hidden
    ├── train
    ├── unlabeled
    └── val

Step3: Prepare conda environment

conda env create -n dfvp -f base_environment.yml
conda activate dfvp

Step4: Train the U-Net

python dfUNet_train.py

Weights will be saved to weights_hub/unet

Step5: Generate labels for unlabeled dataset

python dfLabeler.py --dataset_path './dataset/unlabeled' --unet_weight './weights_hub/unet/best_model.pth'

Step6: Train SimMP2

All configurations could be found in dragonfruitvp/custom_configs. Set test: False, submission: False in training config. Adjust gpus and num_workers based on your gpu numbers.

python dfMP_train.py --model_config_file 'dragonfruitvp/custom_configs/model_configs/mpl_gsta.yaml' --training_config_file 'dragonfruitvp/custom_configs/training_configs/mptrain_e10lr3oc.yaml'

Weights and logs will be saved to './weights_hub/<training_config>_<model_config>'. If vis_val: True in training config, then images for visualization during validation and test epochs will be saved to vis_* folder correspondingly.

Step7: Get prediction results for hidden set

Label the hidden set

python dfLabeler.py --dataset_path './dataset/hidden' --unet_weight './weights_hub/unet/best_model.pth'

Go to the desired training config yaml file ('dragonfruitvp/custom_configs/training_configs/mptrain_e10lr3oc.yaml'), and set test: True, submission: True. Then run dfMP_train again

python dfMP_train.py --model_config_file 'dragonfruitvp/custom_configs/model_configs/mpl_gsta.yaml' --training_config_file 'dragonfruitvp/custom_configs/training_configs/mptrain_e10lr3oc.yaml'

Result will be saved as team_12.pt.

File Structure

• weights_hub: All model weights will be saved here. Could be changed in configuration.
• lightning_logs: Logs saved by pytorch lightning during training
• vis_*: Images and Masks for visualization.

GCP Notes

Connet to GCP

Login burst

ssh burst

Interactive runtime

CPU:

srun --partition=interactive --account csci_ga_2572_002-2024sp-x --pty /bin/bash

GPU:

srun --partition=n1s8-v100-1 --gres=gpu:1 --account csci_ga_2572_002-2024sp-x --time=04:00:00 --pty /bin/bash

Send files

scp <NetID>@greene-dtn.hpc.nyu.edu:/path/to/files /home/<NetID>/

For sending the whole folder:

scp -r <NetID>@greene-dtn.hpc.nyu.edu:/path/to/folder /home/<NetID>/