LaRS Segmentation Starter Kit (MMSegmentation)

This repository is a fork of MMSegmentation 0.x.

It provides a starting point for running semantic segmentation experiments on the LaRS dataset. - Dataloader for LaRS - Configs for a large number of segmentation methods - Utilities for training and making predictions on LaRS

This document provides the basic information and steps to run simple training and inference tasks. For more complex use case scenarios, please refer to the official MMSegmentation repository.

Installation

Follow the instructions to install this version of MMSegmentation.

Step 1. Clone the repository: shell git clone https://github.com/lojzezust/mmsegmentation-macvi.git cd mmsegmentation-macvi Step 1. Create a conda or virtualenv environment. Install PyTorch following official instructions, e.g.

shell pip3 install torch torchvision

Step 2. Install MMCV using MIM.

shell pip install -U openmim mim install mmcv-full

Step 3. Install MMSegmentation (MaCVi) from source.

```shell pip install -v -e .

"-v" means verbose, or more output

"-e" means installing a project in editable mode,

thus any local modifications made to the code will take effect without reinstallation.

```

Step 4. Install additional requirements.

shell pip install -r requirements.txt

Step 5. Configure paths.

Download the LaRS dataset. Update the path in the dataset config configs/_base_/datasets/lars.py, to point to the location of LaRS dataset.

Getting started

Training methods

Use one of the provided training configs to train a method.

shell export CUDA_VISIBLE_DEVICES=0,1 python tools/train.py configs/fcn/fcn_r50-d8_512x1024_40k_lars.py

By default the configs use a batch size of 4 per GPU. You can change this in the dataset config (configs/_base_/datasets/lars.py).

Running inference

Use the tools/test.py script to run inference on the LaRS test set (for submission to macvi.org).

```shell CONFIG=configs/fcn/fcnr50-d8512x102440klars.py WEIGHTS=workdirs/fcnr50-d8512x102440klars/latest.pth # Weights path OUTDIR=output/fcnr50-d8512x102440klars # Output dir

export CUDAVISIBLEDEVICES=0 python tools/test.py $CONFIG $WEIGHTS --show-dir $OUT_DIR ```

Use the --val flag to run on the validation set instead (for local evaluation).

shell python tools/test.py $CONFIG $WEIGHTS --show-dir $OUT_DIR --val

Configs

The following LaRS configs are included in this repository:

| method | backbone | config | |------------|------------|-------------------------------------------------------------------------| | FCN | ResNet-50 | configs/fcn/fcnr50-d8512x102440klars.py | | FCN | ResNet-101 | configs/fcn/fcnr101-d8512x102440klars.py | | UNet | S5 | configs/unet/fcnunets5-d164x4512x1024160klars.py | | DeepLabv3 | ResNet-101 | configs/deeplabv3/deeplabv3r101-d8512x102440klars.py | | DeepLabv3+ | ResNet-101 | configs/deeplabv3plus/deeplabv3plusr101-d8512x102440klars.py | | BiSeNetv1 | ResNet-50 | configs/bisenetv1/bisenetv1r50-d32in1k-pre4x41024x1024160klars.py | | BiSeNetv2 | - | configs/bisenetv2/bisenetv2fcn4x41024x1024160klars.py | | STDC 1 | - | configs/stdc/stdc1in1k-pre512x102480klars.py | | STDC 2 | - | configs/stdc/stdc2in1k-pre512x102480klars.py | | PointRend | ResNet-101 | configs/pointrend/pointrendr101512x102480klars.py | | SegFormer | MiT-B2 | configs/segformer/segformermit-b28x11024x1024160klars.py | | Segmenter | ViT-B | configs/segmenter/segmentervit-bmask8x1512x512160klars.py | | KNet | Swin-T | configs/knet/knets3upernetswin-t8x2512x512adamw80k_lars.py |