Homework-3 - Tiny Pascal VOC dataset

This model is for instance segmentation for the Tiny Pascal VOC dataset. This repo is based on Detectron2.

Hardware

• Ubuntu 18.04.5 LTS
• Intel® Xeon® Silver 4210 CPU @ 2.20GHz
• NVIDIA GeForce RTX 2080 Ti

Reproduce Submission

To reproduce my submission without training, do the following:

1. Installation
2. Data Preparation
3. Inference

Installation

This repo uses Detectron2 from facebook research team. Please install it first before running this repo. I am using CUDA verion 10.2 and torch version 1.7.0. You can either run the following command or go to Detectron2 and Pytorch to install.

pip install torch torchvision

python -m pip install detectron2 -f https://dl.fbaipublicfiles.com/detectron2/wheels/cu102/torch1.7/index.html

Data Preparation

The data should be placed as follows:

repo
  +- train_images
  |  +- ...
  |
  +- test_images
  |  +- ...
  |
  +- output
  |  +- model_final.pth   (needed for inference)
  +- pascal_train.json
  +- test.json
  +- train.py
  +- infer.py
  +- X-101-32x8d.pkl   (needed for training)
  |  ...

The traing and testing images can be downloaded here.

Training

To train, please download the pretrained weight here and put it beside train.py. Simply run train.py. The weights should be saved in 'output' folder with name 'model_final.pth'. There will be several file saving weights at different training process. The batch_size is set to be 2. It can be changed in line 58, where 'cfg.SOLVER.IMS_PER_BATCH = batch_size'. Make it smaller if needed. The trained model will also be used to infer the test images after training and the prediction will also be made.

Inference

for inference, please download the weights file here and put it in output folder. Simply run infer.py and predictions.json will be created.

Citation

Detectron2