BDD100K #264
BDD100K #264
Conversation
Update INSTALL.md
Add ISSUE_TEMPLATE for Github
Fix a small typo
It is simpler and requires less effort (i.e. no auth) compared to SSH
Access repos via https in INSTALL.md
* Add missing __init__.py files * Add packages * Rename logging.py to logger.py Import rules from Python2 makes this a bad idea * Make import_file py2 compatible * list does not have .copy() in py2 * math.log2 does not exist in py2 * Miscellaneous fixes for py2 * Address comments
Fix BoxList constructor call in README.md
Replace baseline with benchmark
Replace BoxList argument size with image_size.
use cuda version torch.randperm to avoid copy from gpu to cpu and a fatal bug in multi-thread cpu version
* quick fix for area * Update bounding_box.py
) ## Bug When I repeatedly invoke build_detection_model function to build the model, I will get the error: `AttributeError: 'ResNet' object has no attribute 'layer1'`. ## To Reproduce ``` model = build_detection_model(cfg) model = build_detection_model(cfg) ``` ## The reason The variable `ResNet50StagesTo4` is a global generator expression, so when we build the model secondly, the code executes to Line 82 `for stage_spec in stage_specs:` and `stage_specs` will return empty leading to do not add any stage. Finally, `self._freeze_backbone` try to freeze the backbone by executing `m = getattr(self, "layer" + str(stage_index))` firstly. At the moment, it will throw the AttributeError `AttributeError: 'ResNet' object has no attribute 'layer1'`. I guess you want to define ResNet50StagesTo4 as the tuple, so I try to fix by add tuple type qualifier. ## The solution Add the tuple type to `ResNet50StagesTo5`, `ResNet50StagesTo4`, `ResNet50FPNStagesTo5`, `ResNet101FPNStagesTo5`. I do not know whether there are similar bug existing, so you need to review my solution. Thank you!
* Inference/evaluation in BDD format * Tensorboard support
|
Thank you for your pull request and welcome to our community. We require contributors to sign our Contributor License Agreement, and we don't seem to have you on file. In order for us to review and merge your code, please sign up at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need the corporate CLA signed. If you have received this in error or have any questions, please contact us at [email protected]. Thanks! |
|
Hi, So, just to clarify: is your code working and giving expected results or not? Also, the branch has several conflicts with the current master which would be good to be addressed if you would like to see this PR merged. Thanks! |
|
Hi! Thank a lot for your reply!
I want to train Mask RCNN model on BDD100K. But now, from your Github project, I do not understand how to start this project, even the model train on COCO or VOC dataset. Could you give me some guidances on how to start?
To be honest, I have experienced Detectron project (based on Caffe2 version). in view of so amazing performance of Mask RCNN Benchmark, I intend to train this model on a new dataset.
I am looking forward for your letters!
Sincerely!
…------------------ 原始邮件 ------------------
发件人: "Francisco Massa"<[email protected]>;
发送时间: 2018年12月12日(星期三) 凌晨0:42
收件人: "facebookresearch/maskrcnn-benchmark"<[email protected]>;
抄送: "林来疯"<[email protected]>; "Author"<[email protected]>;
主题: Re: [facebookresearch/maskrcnn-benchmark] BDD100K (#264)
Hi,
So, just to clarify: is your code working and giving expected results or not?
I haven't myself trained any models on BDD100k dataset, so I might not be able to help on particular aspects of it.
Also, the branch has several conflicts with the current master which would be good to be addressed if you would like to see this PR merged.
Thanks!
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or mute the thread.
|
|
Hi, Did you check the content of this file? I think it will be very helpful as a starting point. |
|
Also, it would be very helpful to know what are the specific questions your have. Creating a new dataset should be quite straightforward, the most involved part is getting it to work correctly with the evaluation code |
|
Thanks a lot.
I have checked up the file you mean, but my question is that whether I need to transform the BDD label to COCO format. In other word, whether I need to prepare a piece of BDD data like COCO format exspecially for label format and content, so that the code could be normallly operated.
Looking for your reply.
Sincerely!
------------------ 原始邮件 ------------------
发件人: "Francisco Massa"<[email protected]>;
发送时间: 2018年12月12日(星期三) 下午5:24
收件人: "facebookresearch/maskrcnn-benchmark"<[email protected]>;
抄送: "林来疯"<[email protected]>; "Author"<[email protected]>;
主题: Re: [facebookresearch/maskrcnn-benchmark] BDD100K (#264)
Also, it would be very helpful to know what are the specific questions your have. Creating a new dataset should be quite straightforward, the most involved part is getting it to work correctly with the evaluation code
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or mute the thread.
|
|
You can convert the BDD ground-truth data to COCO format, and things should run smoothly. |
|
Thanks.
Firstly, I have tried so many bdd2coco script, but it seems to be infeasible. Maybe I found inappropriate scripts. And when I trained mask rcnn on BDD dataset with transformed json file, there always existed some bugs, such as KeyError u'Segmentation or u'Area, etc. Could you help me to make a transformation to generate standard coco-format json?
Furthermore, if I choose another method -- making my own dataset class, I do not know how to code. Aithough there is an example in the mask rcnn benchmark project, I am very confused about how I can go on it with BDD100K dataset. In the coco.py file, if I replace coco-related infromation with BDD information, could I generate my own dataset class correctly and make a successful training process?
Looking forward for your reply! Thanks a lot!
Sincerely!
…------------------ 原始邮件 ------------------
发件人: "Francisco Massa"<[email protected]>;
发送时间: 2018年12月13日(星期四) 晚上6:54
收件人: "facebookresearch/maskrcnn-benchmark"<[email protected]>;
抄送: "林来疯"<[email protected]>; "Author"<[email protected]>;
主题: Re: [facebookresearch/maskrcnn-benchmark] BDD100K (#264)
You can convert the BDD ground-truth data to COCO format, and things should run smoothly.
But this is not strictly necessary, because you can write your own Dataset class for it and the rest should work without problems, apart from the evaluation part, which requires a bit more work.
—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub, or mute the thread.
|
|
Hi, when I ran the command 'python tools/train_net.py --config-file configs/e2e_faster_rcnn_R_50_C4_1x.yaml' to train my BDD100K dataset, I met with this bug. How can I solve it? The dataset class was compiled, and according dataset file has already been prepared. But I still encountered the bug as follows, could you help me or give me some guidance? ImportError: cannot import name '_C' from 'maskrcnn_benchmark' (/mnt/nfsdir/lambda-cloud/data/lcm/github/Detector-KITTI-BDD100k/maskrcnn_benchmark/init.py) Thanks a lot! And I am looking forward for your letters! |
|
@Linstancy it looks like you haven't installed About how to write a Dataset class, you just need to implement a I'd recommend trying to check in a jupyter notebook or in the interpreter if the current implementation that you have is giving expected results. Also, I'm not sure what the format for the segmentation masks in BDD100k, is it a list of polygons or binary masks? |
|
Thanks.
I indeed installed maskrcnn benchmark following the INSTALL.MD, and it seems not to appear bugs during installing process.
After installation, I ran the command "python tools/train.net.py --config-file configs/adas_e2e_mask_rcnn_R_101_FPN_1x.yaml", but the incorrect information presented "maskrcnn_benchmark not found _C". I do not know what lead to it and how to solve it.
I have made my own dataset class according to DatasetCatalog method, and modified cfg file based on BDD100K dataset.
I am very confused and depressed about current status. Could you give me some guidances and help me to solve it?
Looking forward for your letters!
Thanks a lot!
Sincerely!
…------------------ 原始邮件 ------------------
发件人: "Francisco Massa"<[email protected]>;
发送时间: 2018年12月14日(星期五) 晚上6:44
收件人: "facebookresearch/maskrcnn-benchmark"<[email protected]>;
抄送: "林来疯"<[email protected]>; "Mention"<[email protected]>;
主题: Re: [facebookresearch/maskrcnn-benchmark] BDD100K (#264)
@Linstancy it looks like you haven't installed maskrcnn-benchmark following the installation instructions.
About how to write a Dataset class, you just need to implement a __getitem__ method that returns the image, the BoxList object and the image idx.
Creating the BoxList is the only new thing compared to other PyTorch projects, and it should be very straightforward.
I'd recommend trying to check in a jupyter notebook or in the interpreter if the current implementation that you have is giving expected results. Also, I'm not sure what the format for the segmentation masks in BDD100k, is it a list of polygons or binary masks?
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or mute the thread.
|
|
@Linstancy can you run the demos using one of the pre-trained models? I think there is a problem with your installation. Upon a quick look at your code for the dataset it looks reasonable. I'm not sure where / what are the errors you are getting, but I believe you might be almost there to fix it and get it running. |
|
Thanks.
I have checked my dataset class and attacted it in this email. Could you help me to have a look?
I still do not know what problem happens in my installation. Do you meet with the bug that I sent to you in the last email? I indeed done it following the INSTALL.MD file, and there was no abnormal information appeared.
Shall I need to reinstall it?
Sincerely.
…------------------ 原始邮件 ------------------
发件人: "Francisco Massa"<[email protected]>;
发送时间: 2018年12月14日(星期五) 晚上9:25
收件人: "facebookresearch/maskrcnn-benchmark"<[email protected]>;
抄送: "林来疯"<[email protected]>; "Mention"<[email protected]>;
主题: Re: [facebookresearch/maskrcnn-benchmark] BDD100K (#264)
@Linstancy can you run the demos using one of the pre-trained models? I think there is a problem with your installation.
Upon a quick look at your code for the dataset it looks reasonable. I'm not sure where / what are the errors you are getting, but I believe you might be almost there to fix it and get it running.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or mute the thread.
|
|
I think you might be better off trying first to run some working examples, like the demo. |
|
Thanks a lot for your letters during this period, sincerely!
I am trying to run the demo, to test where the problem maybe occur. And I will also consider to reinstall it as you mentioned.
Furthermore, if I reinstall this environment, is it enough for me to follow the requirement of INSTALL.MD file? And shall I need to uninstall this project at first?
Best regards!
…------------------ 原始邮件 ------------------
发件人: "Francisco Massa"<[email protected]>;
发送时间: 2018年12月15日(星期六) 凌晨0:27
收件人: "facebookresearch/maskrcnn-benchmark"<[email protected]>;
抄送: "林来疯"<[email protected]>; "Mention"<[email protected]>;
主题: Re: [facebookresearch/maskrcnn-benchmark] BDD100K (#264)
I think you might be better off trying first to run some working examples, like the demo.
You might probably need to reinstall it, as the current installation doesn't seem to be working.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or mute the thread.
|
|
Oh, I remembered, when I install the environment following the INSTALL.MD, I alo fail to run the 'Webcam and Jupyter notebook demo'. The information seems to be 'No module named maskrcnn.config', which was also a significant proble that I am deeply confused about. Because I have referenced the issue in #22 , but I do not find how to solve thie bug.
Not only that, when I ran the train_net.py, I met with the bug information about 'maskrcn_benchmark not found _C module'. More worse, I have also encountered the trouble 'Segmentation fault (core dumped)', but I do not solve any of these problems as I mention.
I am really a bit depressed, and nearly miss myself in the trouble of so much agnostic bug information.
Anyway, I will go on runing some example to test the code. And I am also looking forward for your introducive advices!
Best regards!
…------------------ 原始邮件 ------------------
发件人: "Francisco Massa"<[email protected]>;
发送时间: 2018年12月15日(星期六) 凌晨0:27
收件人: "facebookresearch/maskrcnn-benchmark"<[email protected]>;
抄送: "林来疯"<[email protected]>; "Mention"<[email protected]>;
主题: Re: [facebookresearch/maskrcnn-benchmark] BDD100K (#264)
I think you might be better off trying first to run some working examples, like the demo.
You might probably need to reinstall it, as the current installation doesn't seem to be working.
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or mute the thread.
|
|
Could you please try uninstalling |
* fix kitti training script * fix
* fix bdd * Update README.md
|
@Linstancy maybe you'll find the converter I wrote useful as a starting point. It's in my fork: |
* bdd100k evaluation * minor fix
* train bdd detection in coco format * Update coco.py * Gradient clipping
|
Has anyone successfully trained Mask-RCNN on BDD100K? P.s. I trained on Cityscapes using the COCO pre-trained weight as well and during the 100,000 iterations, the validation loss decreases quickly and the AP had a large improvement... |
Hi, I want to train BDD100K on the detectron framework. Did you have completed it, or could you please give me some guidance about BDD100K trained on faster rcnn or mask rcnn?
Thanks a lot.
Looking forard for your reply!