wannabeog / mask-rcnn Goto Github PK

View Code? Open in Web Editor NEW

990.0 23.0 180.0 70 KB

A PyTorch implementation of the architecture of Mask RCNN

License: Other

Python 84.14% Cuda 5.90% C 9.88% C++ 0.08%

mask-rcnn's Introduction

EDIT (AS OF 4th NOVEMBER 2019):

This implementation has multiple errors and as of the date 4th, November 2019 is insufficient to be utilized as a resource to understanding the architecture of Mask R-CNN. It has been pointed out to me through multiple emails and comments on HackerNews that such a faulty implementation is to the detriment of the research endeavors in the deep learning community. It was a project that I had put together quite early in my academic career and I did not realize the scale of my mistake
I intend to take care of the issues (the issues filed in this repository are representative) and make this code more "readable" and embellish it with better documentation so that it fulfills the purpose for which it was made. Unfortunately, as of right now, I am busy with my academics and cannot attend to this project. I shall start working on bettering this repository by mid-January to early February 2020. Until then, I have provided links to other implementations of Mask R-CNN that I think could help serve your purpose
PR's fixing any one of the issues listed are always welcome and will allow me to get a headstart on this particular task of making this repository more presentable.

Once again I would like to apologize for any inconvenience caused

Mask-RCNN

A PyTorch implementation of the architecture of Mask RCNN

Decription of folders

model.py includes the models of ResNet and FPN which were already implemented by the authors of the papers and reproduced in this implementation
nms and RoiAlign are taken from Robb Girshick's implementation of faster RCNN
Focal loss has been added to this implementtaion on lieu of better results as evidenced by the paper on RetinaNets

Mask-RCNN model:

Features:

The part of the network responsible for bounding box detection derives it's inspiration from the faster RCNN model having a RPN working in tandem with a ConvNet
The pooling layers present in the ConvNet round down or round up to the nearest integer when the stride is not a divisor of the receptive field, which tends to either lose or assume "information" from the image respectively at the non integral points.
ROI align was proposed to deal with this, wherein bilinear interpolation is used to detect the values at the non integral values of the pixels
Using a more complex interpolation scheme( cubic interpolation -> 16 additional features) offers a slightly better result when this model was tested, however not enough to justify the additional complexity
Cross entropy loss when summed over a huge number of proposals tends to take a huge value for proposals that have a high confidence metric thereby dwarfing the contribution from the proposals of interest. Focal Loss was proposed to do away with this problem
However Focal loss gives much better results with single stage networks. This is because a two stage network has some discriminative policy to deal with this class imbalance something which the single stage networks don't enjoy.

If you find any issue in this repsoritory, feel free to fork this repository and submit a PR with the necessary changes

mask-rcnn's People

Contributors

Stargazers

Watchers

Forkers

kywang sdcjimmy hundred06 machinelp hzhang57 trendingtechnology arunkumarramanan kraken000 december-boy ionvision faviovazquez jiafenggang2 wpf535236337 jiaweihe1996 xiaoxunzi kaihuatang jackyspeed lamhocn peterxiaoguo liuguoyou heyecheng cclauss qdet huizhang0110 dreadlord1984 gljhd dw198033 thb1194 xiaoweihappy123 hehuanshu96 bhushan23 mryang23 bruceyang2012 csjfwang zzm422 hbcbh1999 dgreyling tonyle9 johndpope tony32769 meelement eason-ho allensmile ordinaryparanoia himpsky briando2005 zju-plp jianhua2022 venshow-s newcoder0531 xychen9459 hhgxx123 xzf125244170 amandaynzhou ddeeppnneett babylons2018 nmxnql siyuhsu hongbowei eviliclufas asimay re3write clearsky767 yunxinan boozyguo kjain1rms synioe zhdai duckj thomasgui76 danny95333 etmwb vibashan yyf8989 rameezrehman83 fighting-jj shawn5141 xiaobai0417 shiyuan0806 lty1994 toddwyl peggyzlp h11zang jluhuangj a8039974 lianxxx raymondzmc dalei001 manutdzou weiyunchen happyconan kaitodes gongypgit jianglin314 mejihero tclikang linhduongtuan xiesibo wandaoming barongeng

mask-rcnn's Issues

How about accuracy?

Hi,

I want to know accuracy of your implementation.

IndentationError in model.py

Python 3 treats IndentationErrors as syntax errors

flake8 testing of https://github.com/wannabeOG/Mask-RCNN on Python 3.6.3

$ flake8 . --count --select=E901,E999,F821,F822,F823 --show-source --statistics

./model.py:1255:4: E999 IndentationError: unindent does not match any outer indentation level
   .
   ^
1     E999 IndentationError: unindent does not match any outer indentation level
1

What's your pytorch's version?

Undefined names in ./model.py

Undefined names have the potential to raise NameError at runtime.

flake8 testing of https://github.com/wannabeOG/Mask-RCNN on Python 3.7.0

$ flake8 . --count --select=E901,E999,F821,F822,F823 --show-source --statistics

./model.py:828:23: F821 undefined name 'parse_image_meta'
    _, _, window, _ = parse_image_meta(image_meta)
                      ^
./model.py:989:13: F821 undefined name 'one_hot_embedding'
        t = one_hot_embedding(y.data.cpu(), 1+self.num_classes) 
            ^
./model.py:1009:13: F821 undefined name 'one_hot_embedding'
        t = one_hot_embedding(y.data.cpu(), 1+self.num_classes)
            ^
./model.py:1158:22: F821 undefined name 'compute_rpn_class_loss'
    rpn_class_loss = compute_rpn_class_loss(rpn_match, rpn_class_logits)
                     ^
./model.py:1225:18: F821 undefined name 'compose_image_meta'
    image_meta = compose_image_meta(image_id, shape, window, active_class_ids)
                 ^
./model.py:1389:18: F821 undefined name 'mold_image'
        images = mold_image(image.astype(np.float32), self.config)
                 ^
./model.py:1996:28: F821 undefined name 'mold_image'
            molded_image = mold_image(molded_image, self.config)
                           ^
./model.py:1998:26: F821 undefined name 'compose_image_meta'
            image_meta = compose_image_meta(
                         ^
8     F821 undefined name 'parse_image_meta'
8

Line 1048 function name 'forward' is incorrect in file model.py

how to train my own datasets

Did you have trained another datasets?How did you do?

How to download the trained model?

Is ''mask_rcnn_coco.pth'' in demo.py your trained model?
Is there a link to download?

multi-gpu training

Hi, it seems that GPU_COUNT in config.py doesn't work. Is there any way to multi-gpu training? Thanks!

How to train Cityscapes dataset

Hi！How to load and train cityscapes dataset? Cityscapes dataset has no bounding box, so how do you train it for detection? thanks.

which version of the pytorch when you training?

'compose_image_meta' is not defined

How to apply data augmentation?

No module named 'nms._ext'

Could you help me with this problem?

ModuleNotFoundError Traceback (most recent call last)
in ()
17 from config import Config
18 import utils
---> 19 import model as modellib
20 import visualize
21 from model import log

~/cell_seg/Mask-RCNN/codes/model.py in ()
15 import utils
16 import visualize
---> 17 from nms.nms_wrapper import nms
18 from roialign.roi_align.crop_and_resize import CropAndResizeFunction
19

~/cell_seg/Mask-RCNN/codes/nms/nms_wrapper.py in ()
9 from future import print_function
10
---> 11 from nms.pth_nms import pth_nms
12
13

~/cell_seg/Mask-RCNN/codes/nms/pth_nms.py in ()
1 import torch
----> 2 from ._ext import nms
3 import numpy as np
4
5 def pth_nms(dets, thresh):

ModuleNotFoundError: No module named 'nms._ext'

A Bug in model.py

the mistake occurs when detecting nothing! The score of background is the highest!

No module named 'nms._ext'

I run the coco.py, but have this error:
from ._ext import nms
ModuleNotFoundError: No module named 'nms._ext'

can you solve this error, thanks!

IndexError: too many indices for tensor of dimension 1

Hi! when I run coco.py，it appears the error"IndexError: too many indices for tensor of dimension 1
",I tried to output "gt_class_ids "value and found that its value is positive. Thanks.

Only support batchsize=1?

Hi,
in the roialignlayer, Currently only supports batchsize 1?

How to obtain pre-trained imagenet model?

Hi, could you please provide a download link of the pre-trained model: resnet50_imagenet.pth?

Problems about Focal loss addition

Q1.In compute_loss( ) function, rpn_class_loss = utils.compute_rpn_class_loss(rpn_match, rpn_class_logits), but utils.compute_rpn_class_loss() doesn't exist.
Q2.In model.compute_rpn_class_loss(), you use the original cross-entropy instead. The only use of focal loss is in mask.forward(), but it seemed doesn't work.
Looking for your kind reply. Thanks in advance.