yingxingde / fasterrcnn-pytorch Goto Github PK

View Code? Open in Web Editor NEW

135.0 7.0 50.0 82 KB

FasterRCNN is implemented in VGG, ResNet and FPN base.

License: MIT License

Python 89.11% Shell 0.16% C++ 0.32% Cuda 5.76% C 4.64%

pytorch faster-rcnn fpn

fasterrcnn-pytorch's Introduction

FasterRCNN-pytorch

FasterRCNN is implemented in VGG, ResNet and FPN base.

reference:

rbg's FasterRCNN code: https://github.com/rbgirshick/py-faster-rcnn

Model Performance

Train on VOC2017 Test on VOC2017

Backbone	mAp
VGG16	0.7061
ResNet101	0.754

Train Your Model

1.Before Run You Need:

cd ./lib

Change gpu_id in make.sh and setup.py.
Detially, you need modify parameter setting in line 5, 12 and 19 in make.sh and line 143 in setup.py where include key words '-arch=' depend on your gpu model.(select appropriate architecture described in table below)

sh make.sh

GPU model Architecture

TitanX (Maxwell/Pascal) sm_52

GTX 960M sm_50

GTX 108 (Ti) sm_61

Grid K520 (AWS g2.2xlarge) sm_30

Tesla K80 (AWS p2.xlarge) sm_37
cd ../

mkdir ./data

mkdir ./data/pretrained_model

download pre-trained weights in ./data/pretrained_model
run train.py

GPU model	Architecture
TitanX (Maxwell/Pascal)	sm_52
GTX 960M	sm_50
GTX 108 (Ti)	sm_61
Grid K520 (AWS g2.2xlarge)	sm_30
Tesla K80 (AWS p2.xlarge)	sm_37

2.How to use?

Note: decentralization in preprocesing is based on BGR channels, so you must guarantee your pre-trained model is trained on the same channel set if you use transfer learning

For example:

VGG: CUDA_VISIBLE_DEVICES=1 python train.py --net='vgg16' --tag=vgg16 --iters=70000 --cfg='./experiments/cfgs/vgg16.yml' --weight='./data/pretrained_model/vgg16_caffe.pth'

CUDA_VISIBLE_DEVICES=2 python test.py --net='vgg16' --tag=vgg16 --model=60000 --cfg='./experiments/cfgs/vgg16.yml' --model_path='voc_2007_trainval/vgg16/vgg16_faster_rcnn' --imdb='voc_2007_test' --comp

ResNet:

CUDA_VISIBLE_DEVICES=2 python train.py --net='res18' --tag=res18 --iters=70000 --cfg='./experiments/cfgs/res18.yml' --weight='./data/pretrained_model/Resnet18_imagenet.pth'

CUDA_VISIBLE_DEVICES=3 python train.py --net='res50' --tag=res50 --iters=70000 --cfg='./experiments/cfgs/res50.yml' --weight='./data/pretrained_model/Resnet50_imagenet.pth'

CUDA_VISIBLE_DEVICES=7 python train.py --net='res101' --tag=res101 --iters=80000 --cfg='./experiments/cfgs/res101.yml' --weight='./data/pretrained_model/resnet101_caffe.pth'

CUDA_VISIBLE_DEVICES=6 python test.py --net='res101' --tag=res101_1 --cfg='./experiments/cfgs/res101.yml' --model=70000 --model_path='voc_2007_trainval/res101_1' --imdb='voc_2007_test' --comp

fasterrcnn-pytorch's People

Contributors

Stargazers

Watchers

fasterrcnn-pytorch's Issues

请问有requirements.txt吗？

请问pytorch、CUDA、setuptool、python版本分别是多少呢？
我一直编译不成功

权重在哪下载

你好！非常感谢你的分享，请问预训练的权重在哪下载呢

assert rpn_keep.numel() == cfg.TRAIN.RPN_BATCHSIZE

Traceback (most recent call last):
File "train.py", line 169, in
pretrained_model=args.weight, max_iters=args.max_iters)
File "/work/tcxia/codes/FasterRCNN-pytorch/lib/model/train_val.py", line 487, in train_net
sw.train_model(resume, max_iters)
File "/work/tcxia/codes/FasterRCNN-pytorch/lib/model/train_val.py", line 130, in train_model
loss_r, image_r = net.train_operation(blobs, None, image_if=True)
File "/work/tcxia/codes/FasterRCNN-pytorch/lib/nets/network.py", line 334, in train_operation
result_cls_prob, result_bbox_pred, result_rois = self(im_data, im_info, gt_boxes)
File "/work/tcxia/anaconda3/envs/py36_th0.4/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in call
result = self.forward(*input, **kwargs)
File "/work/tcxia/codes/FasterRCNN-pytorch/lib/nets/network.py", line 247, in forward
cls_prob, bbox_pred, rois = self._predict(im_data, im_info, gt_boxes)
File "/work/tcxia/codes/FasterRCNN-pytorch/lib/nets/network.py", line 214, in _predict
rois, rpn_scores, features = self._rpn(im_data, im_info, gt_boxes)
File "/work/tcxia/anaconda3/envs/py36_th0.4/lib/python3.6/site-packages/torch/nn/modules/module.py", line 477, in call
result = self.forward(*input, **kwargs)
File "/work/tcxia/codes/FasterRCNN-pytorch/lib/nets/network.py", line 88, in forward
self.cross_entropy, self.loss_box = self._build_loss(rpn_cls_score_reshape, rpn_bbox_score, rpn_data)
File "/work/tcxia/codes/FasterRCNN-pytorch/lib/nets/network.py", line 104, in _build_loss
assert rpn_keep.numel() == cfg.TRAIN.RPN_BATCHSIZE
AssertionError

where is _roi_pooling?

Excuse me，in lib/roi_pooling/_ext/roi_pooling/init.py from ._roi_pooling import lib as _lib, ffi as _ffi .but there don't have module named _roi_pooling. did you miss the upload?

内存溢出

RuntimeError: cuda runtime error (2) : out of memory at /opt/conda/conda-bld/pytorch_1524577177097/work/aten/src/THC/generic/THCStorage.cu:58

question regarding gpu id

can you explain what do you mean by:

change gpu_id in make.sh and setup.py

I could not find gpu_id in make.sh,
Can you please explain a little more

Thanks

About FPN

Hi, have you tested on fpn yet?
tried it with the following command,
python train.py --net='fpn101' --tag=fpn101 --iters=80000 --cfg='./experiments/cfgs/fpn101.yml' --weight='./data/pretrained_model/resnet101_caffe.pth'
but an error occurred.
line 17, in generate_anchors_global ratios=np.array(anchor_ratios), scales=np.array([anchor_scales[index]])) IndexError: list index out of range
It seems like that in the generate_anchor function, although feat_strides were updated to (4, 8, 16, 32, 64), but the anchors were still the same, ie, 8 16 32 .

Also, there seems to be 2 typo in fpn101.yml,
EXP_DIR should be fpn101, SNAPSHOT_PREFIX should be fpn101 I guess?

怎么配置环境

编译错误

make.sh很难编译，建议新用户不要用这个项目了。白浪费我一天时间，pytorch也要降级，setuptools也要降级。醉了