SegLord: Segmentation Lord

Installation

Environment Setup

python3 -m venv .env
source .env/bin/activate
python -m pip install -U pip

Package Installation

bash install.sh

OR

mkdir .cache

TMPDIR=./.cache pip install wheel tqdm wandb
TMPDIR=./.cache pip3 install torch torchvision torchaudio
TMPDIR=./.cache pip install accelerate einops
TMPDIR=./.cache pip install albumentations
TMPDIR=./.cache pip install segmentation_models_pytorch
TMPDIR=./.cache pip install torchmetrics
TMPDIR=./.cache pip install segformer-pytorch

Experiment Conduction

This repo training procedure is built with the support of Accelerator, thus enabling various modes of training. Before training, direct the working folder to path/to/seglord/seglord. There are two main ways for running with Accelerator

accelerate launch {script_name.py} --arg1 --arg2 ...

or 

python -m accelerate.commands.launch --num_processes=2 {script_name.py} {--arg1} {--arg2}

Single GPU

CUDA_VISIBLE_DEVICES={GPU_ID} accelerate launch main.py --ds citynormal --model dl3p --loss dice --wandb 

or

accelerate launch --gpu_ids {GPU_ID} main.py --ds citynormal --model dl3p --loss dice --wandb 

Multi GPUs

To use all available GPUs

accelerate launch --multi_gpu {GPU_ID} main.py --ds citynormal --model dl3p --loss dice --wandb 

Or to specify the number of GPUs in training

accelerate launch --num_processes=2 main.py --ds citynormal --model dl3p --loss dice --wandb

CPU

To use CPU for training

accelerate launch --cpu main.py --ds citynormal --model dl3p --loss dice --wandb

Precision Configuration

accelerate launch --multi_gpu --mixed_precision=fp16 --num_processes=2 main.py --ds citynormal --model dl3p --loss dice --wandb

Run from configs

There are templates for config file at path/to/seglord/seglord/scripts.

Single GPU

CUDA_VISIBLE_DEVICES="0" accelerate launch --config_file ./scripts/single_gpu.yaml main.py --epochs 1 --debug --wandb

Multi GPU

accelerate launch --config_file ./scripts/multi_gpu.yaml main.py --epochs 1 --debug --wandb