Skip to content

README.md

Latest commit

 

History

History
77 lines (56 loc) · 2 KB

README.md

File metadata and controls

77 lines (56 loc) · 2 KB
Raw

Semantic Segmentation on ADE20K

This directory contains code for semantic segmentation evaluation on ADE20K using mmsegmentation.

Architecture

  • Backbone: MEDiC ViT-Base with FPN (Feature Pyramid Network)
  • Decoder: UPerNet with 150-class output (ADE20K)
  • Resolution: 512x512 with sliding window inference
  • Training: 160K iterations, AdamW with layer decay

Prerequisites

Install mmsegmentation v0.x and mmcv-full:

pip install mmcv-full==1.7.2 -f https://download.openmmlab.com/mmcv/dist/cu121/torch2.4/index.html
pip install mmsegmentation==0.30.0

Adjust the CUDA and PyTorch versions in the mmcv-full URL to match your environment.

Dataset

Download ADE20K and organize as:

data/ADEChallengeData2016/
├── images/
│   ├── training/
│   └── validation/
└── annotations/
    ├── training/
    └── validation/

Update data_root in configs/_base_/datasets/ade20k.py if your data is in a different location.

Training

# Single GPU
cd src/downstream/segmentation
PYTHONPATH=.:$PYTHONPATH python tools/train.py \
    configs/medic/upernet_medic_base_512_160k_ade20k.py \
    --options model.backbone.pretrained='path/to/checkpoint-epoch0290.pth'

# Multi-GPU (4 GPUs)
PYTHONPATH=.:$PYTHONPATH bash tools/dist_train.sh \
    configs/medic/upernet_medic_base_512_160k_ade20k.py 4 \
    --options model.backbone.pretrained='path/to/checkpoint-epoch0290.pth'

Evaluation

PYTHONPATH=.:$PYTHONPATH python tools/test.py \
    configs/medic/upernet_medic_test_ade20k.py \
    path/to/segmentation_checkpoint.pth \
    --eval mIoU

Expected Results

Pretrained Checkpoint mIoU
MEDiC v0 E290 52.7

SLURM Submission

# Training
sbatch scripts/submit_train.sh path/to/pretrain_checkpoint.pth

# Evaluation
sbatch scripts/submit_eval.sh path/to/segmentation_checkpoint.pth