PMF Meta-Dataset DINO

Introduction

The PMF Meta-Dataset DINO project involves model checkpoints designed for meta-testing using different configurations of the DINO architecture. The project includes specific setups for testing across various domains within the Meta-Dataset.

Architecture

The models utilize the DINO architecture, specifically the dino_small_patch16 variant. Requirements for hardware include peak VRAM of approximately 32GB for DINO-small and 42GB for DINO-base.

Training

Meta-testing is conducted using a distributed setup with 8 processes per node. Two main configurations are highlighted: training on the full Meta-Dataset and training on the ImageNet domain within the Meta-Dataset. The models are fine-tuned using specific learning rates and evaluated across multiple domains.

Guide: Running Locally

1. Prerequisites: Ensure the availability of a suitable environment with Python and PyTorch installed. A distributed setup is necessary for running the tests.

2. Setup:

   • Clone the repository containing the code and datasets.
   • Prepare the dataset paths and ensure they are accessible.

3. Run Meta-Testing:

   • For full Meta-Dataset:

     python -m torch.distributed.launch --nproc_per_node=8 --use_env test_meta_dataset.py --data-path ../../datasets/meta_dataset --dataset meta_dataset --arch dino_small_patch16 --deploy finetune --output outputs/md_full_dinosmall --resume md_full_128x128_dinosmall_fp16_lr5e-5/best.pth --dist-eval --ada_steps 100 --ada_lr 0.0001
     

   • For ImageNet domain:

     python -m torch.distributed.launch --nproc_per_node=8 --use_env test_meta_dataset.py --data-path ../../datasets/meta_dataset --dataset meta_dataset --arch dino_small_patch16 --deploy finetune --output outputs/md_inet_dinosmall_6gpus --resume pmf_metadataset_dino/md_inet_128x128_dinosmall_fp16_lr5e-5/best.pth --dist-eval --ada_steps 100 --ada_lr 0.0001
     

4. Consider Cloud GPUs: Given the high VRAM requirements, utilizing cloud-based GPUs might be more feasible than local resources.

License

The project files and models are distributed under a license specified in the repository. Users should review the license to ensure compliance with usage terms.