alphajia / pytorch-faster-rcnn Goto Github PK

pytorch based implementation faster rcnn

Python 100.00%

pytorch-faster-rcnn's Introduction

pytorch-faster-rcnn

1. Introduction

Pytorch based implementation of faster rcnn framework.For details about faster R-CNN please refer to the paper Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks by Shaoqing Ren, Kaiming He, Ross Girshick, Jian Sun

This detection framework has the following features:

It can be run as pure python code, and also pure based on pytorch framework, no need to build
It is easily trained by only running a train.py script, just set the data root dir
It has many backbone networks. like vgg, resnet-fpn, mobilenet, high resolution net(HRNet)
It can be a really detection framework. You only need to change super parameters in config file and get different models to compare different model
It's memory-efficient (about 3GB for vgg16)

2. Installation

2.1 Prerequisites

Python 2.7 or 3.5
Pytorch 1.5.1
torchvision 0.6.1
numpy 1.15.4
Pillow 6.1.0
pycocotools 2.0
matplotlib 3.0.2
tensorboardX 2.0

pip install -r requirements.txt

2.2 Code-Preparing

 git clone https://github.com/AlphaJia/pytorch-faster-rcnn.git

3. Data Preparation

COCO

3.1 Download the training, validation, test data and annotations

 wget http://images.cocodataset.org/zips/train2017.zip  
 wget http://images.cocodataset.org/zips/val2017.zip
 wget http://images.cocodataset.org/zips/test2017.zip
 wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip

3.2 Extract all of these tars into one directory named COCODevKit

 tar xvf train2017.zip
 tar xvf val2017.zip
 tar xvf test2017.zip
 tar xvf annotations_trainval2017.zip

3.3 Data dir should like this

   COCODevKit
       |-- train2017
               |-- [xxxxxxxxxxxx].jpg
       |-- val2017
               |-- [xxxxxxxxxxxx].jpg
       |-- test2017
               |-- [xxxxxxxxxxxx].jpg
       |-- annotations
               |-- instances_train2017.json
               |-- instances_val2017.json
               |-- image_info_test2017.json

3.4 modify data_root_dir cfg item in config/train_config.py with /path/COCODevKit/

4. Train

Modify model_save_dir cfg item in config/train_config.py with your own save path and device_name with your own device

Train with mobilenet
Modify backbone cfg item in config/train_config.py with mobilenet, download pretrained weights here, and set backbone_pretrained_weights in config/train_config.py with downloaded path.

 python train.py

Train with resnet-fpn
Modify backbone cfg item in config/train_config.py with resnet50_fpn, download pretrained weights here, and set backbone_pretrained_weights in config/train_config.py with downloaded path

 python train.py

Train with vgg16
Modify backbone cfg item in config/train_config.py with vgg16

 python train.py

Train with HRNet
Modify backbone cfg item in config/train_config.py with HRNe

 python train.py

Weights and tensorboard log will save in your model_save_path dir
you may refer to config/train_config.py for more argument.
Some Key arguments:
--backbone: feature extraction backbone network
--backbone_pretrained_weights: backbone pretrained weights, None or path
--train_horizon_flip_prob: data horizontal flip probability
--num_class: number of classification, including background
--data_root_dir: COCO dataset root dir
--model_save_dir: training weights save path
--device_name: training device
--num_epochs: training epochs

5. Test

Modify model_weights cfg item in config/test_config.py with your trained weights path and gpu_id with your own cuda device ID.
you may refer to config/test_config.py for more argument.
Some Key arguments:
--model_weights: training save path
--image_path: predicted images
--gpu_id: cuda device gpu ID
--num_classes: number of classification, including background
--data_root_dir: COCO dataset root dir

 python test.py

6. Demo

7. Framework Structure

backbone

This module includes backbone feature extraction network

vgg16:vgg16 net network(Very Deep Convolutional Networks for Large-Scale Image Recognition)
fpn101:resnet101 fpn network(Deep Residual Learning for Image Recognition) (Feature Pyramid Networks for Object Detection)
hrnet:high resolution net(Deep High-Resolution Representation Learning for Visual Recognition)
mobile_net:mobile_net v2 network(MobileNetV2: Inverted Residuals and Linear Bottlenecks)

config

This module includes config parameters in training period and testing period

test_config: specify config parameters in testing period like model_file, image_path_dir, save_dir, etc.
train_config: specify config parameters in training period like backbone network, batch_size, image_path_dir, anchor_size, ect.

dataloader

This module inherits pytorch dataloader classes, dataset IO.You can also generate your own dataset dataloader IO and put it in this module

coco_dataset: coco(Common Objects in Context) dataset dataloader IO

test

This module includes the utils function test(common called unit test, also called UT)

anchor_utils_test: some unit testing for utils/anchor_utils.py

utils

This module includes some utilies for image processing, network architectures building, anchor generating, loss function, etc.

anchor_utils: some basic function for building anchors
im_utils: some basic function for image processing

pytorch-faster-rcnn's People

Contributors

Stargazers

Watchers

Forkers

wannie163 hy89173584jiayu danny1879 pxmapple whyisma blackface111 beyondz gitcoder666 tyunh remasterd sevtins harryxqb lunanov2 makroya niltrou zshpro sfumecjf shipscode zengfengzhen tianmei2017 mzywy1 yblbarry maqingya weng-x hiroy101 q-a-a git-hyw fanshanhong boykaa gitai123 hj0921 luoxd1996 ollererr mengchuan87 1046617651 zeroorhero mikec-en f4112cd hxdaze deepaksinghcv j-w-ca wozuidiao mugicon zlx1134558955 dannieldwt wangda11 centos-f renqian805 mhw-nku seohokang pradhunmya iostream11 pushparajamurugan leeiadong gavin-evans shakurslee ccdragon1 wufan55 ixiondbz h-jh20 yiruzzz catofwei dark-knight-02 berserkr thucth-qt dazory shuvo001 aakash-chakraborty1995 nuggy875 tarkers datct00 giter915 hasanirtiza zijing-guo wingjson aminemarnissi selimsavas jibanul loinh1106 indonesiaai eunbinchoi iq-scm ngocvuminh asad-be18 andytom666 liudunxu ishitaathakre

pytorch-faster-rcnn's Issues

Please help me

when I run a command which Train with mobilenet, I got information like this:
Epoch: [0] [ 2950/19715] eta: 1:36:08.972184 loss_rpn_box_reg: 0.8739 (1.0911) lr: 0.005000 loss: 1.6920 (2.0218) loss_classifier: 0.3552 (0.4206) loss_box_reg: 0.1971 (0.2049) loss_objectness: 0.2848 (0.3052) time: 0.3408 data: 0.0152 max mem: 0

I want to know the number 19715 represent what and can you tell me how to reduce it's value, because I want to finish the train process as soon as possible.

thanks for your contribution ,but i still have this issue;can you help you

How does faster rcnn calculate ap per class?

How does faster rcnn calculate ap per class? Thank you very much!

Please help me NameError : name "module" is not defined

Hi , i followed the instructions of README.md, prepared the packages and the dataset, and change some paramenters in train_config.py. All in done.
However,when i started to run train.py, errors saying that "module" is not defined. i don not change the code. I am really wonder how to solve the problem.
Thanks for any help!

Model parameter error

Hi,

I downloaded your model fasterrcnn_resnet50_fpn_coco-258fb6c6.pth, but when i try to load this model using model.load_state_dict(checkpoint['model]), it shows "key error 'model'" so i try to print keys in model dictionaries and i didn't find 'model', Is it possible that you put the wrong model?

add more backbone

Hi， thanks for your contribution， would you like to add more backbone like hrnet, some high resolution backbone network
Thanks

rel_codes can not reshape bug

pred_boxes = self.decode_single(rel_codes.reshape(box_sum, -1), concat_boxes)
=>
rel_codes = rel_codes.view(box_sum, -1)
pred_boxes = self.decode_single(rel_codes, concat_boxes)

########################bug ###########################################
the rel_codes = rel_codes.view(box_sum, -1) #reshape fail on fpn resnet-50

########################correct######################################
rel_codes = rel_codes[:box_sum] #this will be correct

thanks for your contribution ,still lack some code.look for your message

I want to know how much mAP can this method get?

The map is vey low

I have been running the model for a long time, but the map is always 0

Error using resnet50-fpn backbone

Hi
I am getting following error if i use resnet50-fpn backbone. mobilenet just works fine. Could you help me.

Traceback (most recent call last):
File "/Users/Mansoor/PycharmProjects/pythonProject/train.py", line 131, in
main()
File "/Users/Mansoor/PycharmProjects/pythonProject/train.py", line 83, in main
print_freq=50, warmup=False)
File "/Users/Mansoor/PycharmProjects/pythonProject/utils/train_utils.py", line 373, in train_one_epoch
loss_dict = model(images, targets)
File "/Users/Mansoor/opt/anaconda3/envs/pythonProject/lib/python3.6/site-packages/torch/nn/modules/module.py", line 550, in call
result = self.forward(*input, **kwargs)
File "/Users/Mansoor/PycharmProjects/pythonProject/utils/faster_rcnn_utils.py", line 85, in forward
proposals, proposal_losses = self.rpn(images, features, targets)
File "/Users/Mansoor/opt/anaconda3/envs/pythonProject/lib/python3.6/site-packages/torch/nn/modules/module.py", line 550, in call
result = self.forward(*input, **kwargs)
File "/Users/Mansoor/PycharmProjects/pythonProject/utils/rpn_utils.py", line 290, in forward
proposals = self.box_coder.decode(pred_bbox_deltas.detach(), anchors)
File "/Users/Mansoor/PycharmProjects/pythonProject/utils/det_utils.py", line 182, in decode
rel_codes.reshape(box_sum, -1), concat_boxes
RuntimeError: shape '[96768, -1]' is invalid for input of size 392832

RuntimeError: shape '[516096, -1]' is invalid for input of size 2095104

I modified the data_root_dir in train_config.py and also placed the data as mentioned in the README.md
I face this error:

/home/dksingh/anaconda3/envs/mmdet220/lib/python3.7/site-packages/torch/nn/functional.py:3000: UserWarning: The default behavior for interpolate/upsample with float scale
_factor changed in 1.6.0 to align with other frameworks/libraries, and uses scale_factor directly, instead of relying on the computed output size. If you wish to keep the
 old behavior, please set recompute_scale_factor=True. See the documentation of nn.Upsample for details. 
  warnings.warn("The default behavior for interpolate/upsample with float scale_factor changed "
Traceback (most recent call last):
  File "train.py", line 132, in <module>
    main()
  File "train.py", line 84, in main
    print_freq=50, warmup=False)
  File "/home/dksingh/codes/clones/pytorch-faster-rcnn/utils/train_utils.py", line 373, in train_one_epoch
    loss_dict = model(images, targets)
  File "/home/dksingh/anaconda3/envs/mmdet220/lib/python3.7/site-packages/torch/nn/modules/module.py", line 722, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/dksingh/codes/clones/pytorch-faster-rcnn/utils/faster_rcnn_utils.py", line 85, in forward
    proposals, proposal_losses = self.rpn(images, features, targets)
  File "/home/dksingh/anaconda3/envs/mmdet220/lib/python3.7/site-packages/torch/nn/modules/module.py", line 722, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/home/dksingh/codes/clones/pytorch-faster-rcnn/utils/rpn_utils.py", line 290, in forward
    proposals = self.box_coder.decode(pred_bbox_deltas.detach(), anchors)
  File "/home/dksingh/codes/clones/pytorch-faster-rcnn/utils/det_utils.py", line 182, in decode
    rel_codes.reshape(box_sum, -1), concat_boxes
RuntimeError: shape '[516096, -1]' is invalid for input of size 2095104

Default process group has not been initialized, please make sure to call init_process_group

When I run python train.py, it is generating errors regarding to the init_process.

I am using HP Z8-G4 workstation for training.

When I install

ERROR: Invalid requirement: u'torch~=1.5.1+cu101' (from line 1 of requirements.txt)
ERROR: Invalid requirement: u'torchvision~=0.6.1+cu101' (from line 5 of requirements.txt)
What should I do?

Question about the Result

I tried to use VOC dataset, so i have modified the dataloader and the evaluation code with some parts of mmedetction. However, i got really pool result (mAP is about 0.167) after 17 epochs training（about 1 day）. So, i'm wondering how many epochs did your coco dataset cost, and what's your final AP result? I would be very grateful if you could give me the answer.

AttributeError: module 'torch.distributed' has no attribute 'barrier'

Hello, I encountered this problem during verification. I would be very grateful if you could give me the answer.

FPN101 error

is this model combination of ResNet101-FPN?
is so then why FPN codes are different than resnet50-fpn?

Question about CPU use

How could I modify this repository for use on a CPU?

you got my star

just star it~~~thanks full py source in faster-rcnn, easier to learn

i think anchor has error

in anchor generator

self.sizes = ((64,), (128,), (256,))
self.aspect_ratios = ((0.5,), (1,), (2.0,), (0.5,), (1,), (2.0,), (0.5,), (1,), (2.0,))

cell_anchors = [generate_anchors(sizes, aspect_ratios, dtype, device)
                        for sizes, aspect_ratios in zip(self.sizes, self.aspect_ratios)]

and cell_anchors result is

[tensor([[-45., -23.,  45.,  23.]], device='cuda:2'),
 tensor([[-64., -64.,  64.,  64.]], device='cuda:2'),
 tensor([[ -91., -181.,   91.,  181.]], device='cuda:2')]

anchors count should be 9 per pixel but only 3 base anchor..
i think because cell anchors made by zip(size, ratios), not product

please help me how to train my own data

i find that the COCODevkit dataset is too large for me, and i need to finish the training as soon as possible.
i tried to changed dataset i need to COCO format, and i changed the data_root_dir in my train_confi.py, however ,it can not work. i really wonder how to do so.
appreciate for any avaiable help

wrong out features on cls_logits

self.conv = nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1)

'''background/foreground score'''
self.cls_logits = nn.Conv2d(in_channels, num_anchors, kernel_size=1, stride=1) #wrong

'''bbox regression parameters'''
self.bbox_pred = nn.Conv2d(in_channels, num_anchors * 4, kernel_size=1, stride=1)

###################
self.cls_logits = nn.Conv2d(in_channels, num_anchors * 2, kernel_size=1, stride=1) #correct