ruyanyinian / cornernetreproduction

复现CornerNet

Pytorch simple CornerNet

This repository is a simple pytorch implementation of CornerNet: Detecting Objects as Paired Keypoints, some of the code is taken from the official implementation. As the name says, this version is simple and easy to read, all the complicated parts (dataloader, hourglass, training loop, etc) are all rewrote in a simpler way.
By the way the support of nn.parallel.DistributedDataParallel is also added, so this implementation trains considerably faster than the official code.

Enjoy!

Requirements:

• python>=3.5
• pytorch==0.4.1 or 1.1.0(DistributedDataParallel training only available using 1.1.0)
• tensorboardX(optional)

Getting Started

1. Disable cudnn batch normalization. Open torch/nn/functional.py and find the line with torch.batch_norm and replace the torch.backends.cudnn.enabled with False.

2. Clone this repo:

   CornerNet_ROOT=/path/to/clone/CornerNet
   git clone https://github.com/zzzxxxttt/pytorch_simple_CornerNet $CornerNet_ROOT
   

3. Install COCOAPI (the cocoapi in this repo is modified to work with python3):

   cd $CornerNet_ROOT/lib/cocoapi/PythonAPI
   make
   python setup.py install --user
   

4. Compile corner pooling. If you are using pytorch 0.4.1, rename $CornerNet_ROOT/lib/cpool_old to $CornerNet_ROOT/lib/cpool, otherwise rename $CornerNet_ROOT/lib/cpool_new to $CornerNet_ROOT/lib/cpool.

   cd $CornerNet_ROOT/lib/cpool
   python setup.py install --user
   

5. Compile NMS.

   cd $CornerNet_ROOT/lib/nms
   make
   

6. For COCO training, Download COCO dataset and put annotations, train2017, val2017, test2017 (or create symlinks) into $CornerNet_ROOT/data/coco

Train

COCO

multi GPU using nn.parallel.DistributedDataParallel

python -m torch.distributed.launch --nproc_per_node NUM_GPUS train.py --dist \
        --log_name coco_hg_511_ddp \
        --dataset coco \
        --arch large_hourglass \
        --lr 5e-4 \
        --lr_step 90,120 \
        --batch_size 48 \
        --num_epochs 200 \
        --num_workers 2

Evaluate

COCO

python test.py --log_name coco_hg_511_dp \
               --dataset coco \
               --arch large_hourglass

# flip test
python test.py --log_name coco_hg_511_dp \
               --dataset coco \
               --arch large_hourglass \
               --test_flip

# multi scale test
python test.py --log_name coco_hg_511_dp \
               --dataset coco \
               --arch large_hourglass \
               --test_flip \
               --test_scales 0.5,0.75,1,1.25,1.5

Results:

COCO:

Model	Training image size	mAP
Hourglass-52 (DDP)	511	35.4/37.1/38.7
Hourglass-104 (DDP)	511	39.4/40.9/42.0