Implement semantic segmentation evaluation procedure

[kutayeroglu]

Following is from ColorMAE:

ADE20k Semantic Segmentation. We employ UperNet [46] as our segmen-
tation model and perform end-to-end fine-tuning on the ADE20k [50] dataset
for 160k iterations with an image resolution of 512 × 512. The evaluation metric
used is the mean Intersection over Union (mIoU)

[kutayeroglu]

Currently working on this locally at /home/kergolu/projects/ijepa

Can run with

conda activate ijepa-seg

# Run from repo root so `src` imports work
export PYTHONPATH="${PWD}:${PYTHONPATH}"

python main_seg_distributed.py \
  --dataset_dir ~/datasets \
  --model_path /path/to/checkpoint.pth.tar \
  --model_name vit_base \
  --patch_size 16 \
  --devices cuda:0 cuda:1 cuda:2 cuda:3