stevetsui/ida-det

IDa-Det: An Information Discrepancy-aware Distillation for 1-bit Detectors

Pytorch implementation of our paper "IDa-Det: An Information Discrepancy-aware Distillation for 1-bit Detectors" accepted by ECCV2022.

Tips

Any problem, please contact the first author (Email: [email protected]).

Our code is heavily borrowed from DeFeat (https://github.com/ggjy/DeFeat.pytorch/) and based on MMDetection (https://github.com/open-mmlab/mmdetection).

Environments

• Python 3.7
• MMDetection 2.x
• This repo uses: mmdet-v2.0 mmcv-0.5.6 cuda 10.1

Get Started

• sh script.sh

Update

We simplify and optimize the code. Now IDa-Det is successfully plugged in the original DeFeat project. The training cost is reduced by about 30% compared with the old version.

VOC Results

Pretrained model is here: GoogleDrive

Notes:

• Faster RCNN based model
• Batch: sample_per_gpu x gpu_num

Model	Batch	Lr schd	box AP	Model	Log
R101	4x2	0.01	81.9	GoogleDrive
R101-BiR18	4x1	0.004	76.9	GoogleDrive

If you find this work useful in your research, please consider to cite:

@inproceedings{xu2022ida,
  title={IDa-Det: An Information Discrepancy-Aware Distillation for 1-Bit Detectors},
  author={Xu, Sheng and Li, Yanjing and Zeng, Bohan and Ma, Teli and Zhang, Baochang and Cao, Xianbin and Gao, Peng and L{\"u}, Jinhu},
  booktitle={European Conference on Computer Vision},
  pages={346--361},
  year={2022},
  organization={Springer}
}

ida-det's Issues

教师模型的预训练权重

您好，我目前训练报错：OSError: /tmp/ImageNet-pretrained/faster-rcnn-v2-374.pth is not a checkpoint file
我不知道这个教师模型的预训练权重是哪一个，您可以告诉我吗？

获取正确的计算量

我调用了您代码里的get_flop.py进行计算，似乎不对，请问您是怎么得到论文中的计算量的呢？

关于教师模型的预训练权重

由于没有提供教师模型的预训练权重，我根据您给的代码自己训了一个。但蒸馏时出现以下报错，是我哪里出错了吗？

2023-02-12 12:01:28,423 - mmdet - INFO - workflow: [('train', 1)], max: 12 epochs
kd_decay rate: 1.0
/home/micro/users/zjl/IDa-Det-main/mmdet/core/bbox/coder/delta_xywh_bbox_coder.py:159: UserWarning: This overload of addcmul is deprecated:
addcmul(Tensor input, Number value, Tensor tensor1, Tensor tensor2, *, Tensor out)
Consider using one of the following signatures instead:
addcmul(Tensor input, Tensor tensor1, Tensor tensor2, *, Number value, Tensor out) (Triggered internally at /opt/conda/conda-bld/pytorch_1607370172916/work/torch/csrc/utils/python_arg_parser.cpp:882.)
gx = torch.addcmul(px, 1, pw, dx) # gx = px + pw * dx
dict_keys(['cls_score', 'bbox_pred', 'bbox_targets', 'sampling', 'neck', 'backbone', 'img_metas', 'gt_bboxes', 'gt_labels'])
Traceback (most recent call last):
File "tools/train_entropy.py", line 206, in
main()
File "tools/train_entropy.py", line 194, in main
train_detector_entropy(
File "/home/micro/users/zjl/IDa-Det-main/mmdet/multi_distillation/detector_entropy.py", line 511, in train_detector_entropy
runner.run(data_loaders, cfg.workflow, cfg.total_epochs, kd_cfg=cfg.model.hint_adapt)
File "/home/micro/users/zjl/IDa-Det-main/mmcv-0.5.1/mmcv/runner/runner_kd.py", line 405, in run
epoch_runner(data_loaders[i], **kwargs)
File "/home/micro/users/zjl/IDa-Det-main/mmcv-0.5.1/mmcv/runner/runner_kd.py", line 303, in train
outputs = self.batch_processor(
File "/home/micro/users/zjl/IDa-Det-main/mmdet/multi_distillation/detector_entropy.py", line 211, in batch_processor_entropy
proposal_list_s = head_det_s['proposal_list']
KeyError: 'proposal_list'

关于两个细节问题

博士您好，我有如下疑问，期待您的解答：
1.我注意到您的代码里使用了Shared2FCBBoxHead，这个检测头包含了较大的参数量和计算量，为什么您没有使用参数量、计算量较小的BBoxHead呢？
2.相较于全精度网络，您的二值化网络使用了更小的学习率和batchsize，这是为什么？

计算IDa-Det计算量

如题，如何精确计算GMac？我调用您代码里的get_flop.py发现计算量似乎偏大了，是需要自己写代码定义模块并除以64来计算吗？

AttributeError: module 'mmdet.models.dense_heads' has no attribute 'RPNHead_React_v2'

如题，dense_heads的定义中没有找到v2，是漏传了.py文件吗？

about how to train and test

hello

Thanks for your contributions

I have some questions about how to train or test your code on coco datasets,

Have you considered creating a training or testing instruction document.

with best wishes

多卡训练

如题，论文中提及使用了4张2080Ti，如何启动多卡训练？

为什么使用SGD优化器？

博士您好，非常感谢您共享的代码以及先前的指导，我想请教一下，大多数二值化网络使用Adam优化器，且有论文提出Adam优化器优于SGD。
但我注意到您的代码中的config文件使用了SGD优化器，是实验证明这样做更优吗？期待您的指导。

COCO的训练配置文件

我注意到论文中有在COCO上的训练结果，能否分享一下COCO上的教师、学生网络对应的config文件？

关于idanet更换backbone

作者您好，想请问一下idanet适合除resnet以外的其他主干网络吗？例如ViT、Swin、PVT等基于Transformer的主干网络。

计算模型推理时间或者FPS

mmdetection采用tools/analysis_tools/benchmark.py来测试FPS，而在您的代码里面没有找到，请问该如何测试FPS呢，或者如何计算推理速度得知模型每秒钟在GPU上可以处理多少张图片

安装报错

博士您好，我无法运行程序，希望能得到您的帮助。
我已经完成mmcv和mmdet的setup.py，然而仍然报错，如下所示：

Traceback (most recent call last):
File "tools/train.py", line 176, in
main()
File "tools/train.py", line 71, in main
cfg = Config.fromfile(args.config)
File "/home/micro/users/zjl/DeFeat.pytorch-main/mmcv-0.5.6/mmcv/utils/config.py", line 165, in fromfile
cfg_dict, cfg_text = Config._file2dict(filename)
File "/home/micro/users/zjl/DeFeat.pytorch-main/mmcv-0.5.6/mmcv/utils/config.py", line 84, in _file2dict
filename = osp.abspath(osp.expanduser(filename))
File "/home/micro/anaconda3/envs/ida/lib/python3.8/posixpath.py", line 231, in expanduser
path = os.fspath(path)
TypeError: expected str, bytes or os.PathLike object, not NoneType