Skip to content

Latest commit

 

History

History
50 lines (30 loc) · 2.46 KB

GETTING_STARTED.md

File metadata and controls

50 lines (30 loc) · 2.46 KB

Getting Started

This document provides tutorials to train and evaluate CircleNet. Before getting started, make sure you have finished installation and dataset setup.

Benchmark evaluation

First, download the models you want to evaluate from our model zoo and put them in CircleNet_ROOT/models/.

MoNuSeg 2018

HG

To evaluate MoNuSeg object detection with HG, run

python main.py circledet exp_id CircleNet_HG_Reproduce --arch hourglass --batch_size 4 --master_batch 4 --lr 2.5e-4 --dataset monuseg --load_model ../models/circledet_monuseg_hg.pth --test --ontestdata --debug 4

This will give an AP of 48.7 if setup correctly.

To evaluate the rotation consistency, run

python main.py circledet --exp_id CircleNet_HG_Reproduce --arch hourglass --batch_size 4 --master_batch 4 --lr 2.5e-4   --dataset monuseg --load_model ../models/circledet_monuseg_hg.pth --test --ontestdata --rotate_reproduce 90

This will give '0.8918255489424661' if setup properly. The qualitative results can be found in 'exp/circledet/CircleNet_HG_Reproduce/debug'

DLA

To test with DLA, run

python main.py circledet --exp_id CircleNet_DLA_Reproduce --arch dla_34 --batch_size 4 --master_batch 4 --lr 2.5e-4  --dataset monuseg --load_model ../models/circledet_monuseg_dla.pth --test --ontestdata --debug 4

This will give an AP of '48.6' if set-up properly.

Rotation consistency can be evaluated by appending '--rotate_reproduce 90'. This should give '0.8855951478392722'.

Training

All the training scripts can be found in experiments folder. The experiment names correspond to the model name in the model zoo.

By default, pytorch evenly splits the total batch size to each GPUs. --master_batch allows using different batchsize for the master GPU, which usually costs more memory than other GPUs.

If the training is terminated before finishing, you can use the same commond with --resume to resume training. It will found the latest model with the same exp_id.

We fine-tune our models using pre-trained COCO models from CenterNet. Our HourglassNet model is fine-tuned from the pretrained COCO CenterNet-HG model. Our DLA model is fine-tuned from the pretrained COCO CenterNet-DLA model