A Pytorch.1.0 Implementation of R-FCN

Introduction

This project is an pytorch implement R-FCN, large part code is forked from princewang1994/RFCN_CoupleNet.pytorch. The R-FCN structure is refer to Caffe R-FCN and Py-R-FCN

What I have done is updating the pytorch version from 0.3 to 1.0. (In theory, it should support pytorch.1.x version)

Preparation

First of all, clone the code

$ git clone https://github.com/RebornL/RFCN-pytorch.1.0.git
$ cd RFCN-pytorch.1.0

prerequisites

• Python 3.6
• Pytorch 1.0.0 (Maybe support higher than 1.0)
• CUDA 10.0 (I use this version)

Compilation

Install all the python dependencies using pip:

$ pip install -r requirements.txt

Compile the cuda dependencies using following simple commands:

$ cd lib
$ python setup.py build develop
$ cd ..

The nms, roi_pool, roi_align, psroi_pool and psroi_align come from below link. Thanks for their open source work.

• NMS, ROIPool, ROIAlign: jwyangfaster-rcnn.pytorch
• PSRoIPool, PSRoIAlign: McDo/PSROIAlign-Multi-Batch-PyTorch

Data Preparation

$ mkdir data
$ cd data

• PASCAL_VOC 07+12: Please follow the instructions in py-faster-rcnn to prepare VOC datasets. Actually, you can refer to any others. After downloading the data, creat softlinks in the folder data/.
• Pretrained ResNet: download resnet50 from here (提取码: fba7) and put it to $RFCN_ROOT/data/pretrained_model/resnet50_rcnn.pth.

Data Folder Presentation

$ tree data

Train

To train a R-FCN model with ResNet101 on pascal_voc, simply run:

$ CUDA_VISIBLE_DEVICES=$GPU_ID python trainval_net.py \
				   --arch rfcn \
                   --dataset pascal_voc --net res101 \
                   --bs $BATCH_SIZE --nw $WORKER_NUMBER \
                   --lr $LEARNING_RATE --lr_decay_step $DECAY_STEP \
                   --cuda

• Set --s to identified differenct experiments.
• Model are saved to $RFCN_ROOT/models

Test

If you want to evlauate the detection performance of a pre-trained model on pascal_voc test set, simply run

$ python test_net.py --dataset pascal_voc --arch rfcn \
				   --net res101 \
                   --checksession $SESSION \
                   --checkepoch $EPOCH \
                   --checkpoint $CHECKPOINT \
                   --cuda

• Specify the specific model session(--s in training phase), chechepoch and checkpoint, e.g., SESSION=1, EPOCH=10, CHECKPOINT=1036.

Problem

1. Train should add --cuda

2. If you have meet this problem below,

File "/home/reborn/Project/RFCN.pytorch.1.0/lib/roi_data_layer/roidb.py", line 9, in <module>
    from datasets.factory import get_imdb
  File "/home/reborn/Project/RFCN.pytorch.1.0/lib/datasets/factory.py", line 15, in <module>
    from datasets.coco import coco
  File "/home/reborn/Project/RFCN.pytorch.1.0/lib/datasets/coco.py", line 23, in <module>
    from pycocotools.coco import COCO
  File "/home/reborn/Project/RFCN.pytorch.1.0/lib/pycocotools/coco.py", line 60, in <module>
    from . import mask
  File "/home/reborn/Project/RFCN.pytorch.1.0/lib/pycocotools/mask.py", line 3, in <module>
    from . import _mask
ImportError: cannot import name '_mask'

You have two way to slove this problem.

• comment the coco dataset part, if you only use the voc datasets.

• install coco api, follow the below step:

  cd data
  git clone https://github.com/pdollar/coco.git 
  cd coco/PythonAPI
  make

Acknowledgement

This project is established by ReornL. Thanks the faster-rcnn.pytorch.1.0's code provider jwyang/faster-rcnn.pytorch , psroialign.pytorch's code provider McDo/PSROIAlign-Multi-Batch-PyTorch and the RFCN_CoupleNet.pytorch' s code provider princewang1994/RFCN_CoupleNet.pytorch. Because of their excellent job, I don't need to start my experiment from scratch.