haomo-ai / OverlapTransformer

[RAL/IROS 2022] OverlapTransformer: An Efficient and Yaw-Angle-Invariant Transformer Network for LiDAR-Based Place Recognition.
GNU General Public License v3.0
256 stars 32 forks source link

方便提供百度云的预训练模型权重文件吗?(pretrained models in Baidu Netdisk) #3

Closed oym1994 closed 2 years ago

oym1994 commented 2 years ago

谢谢大佬们的大作,膜拜膜拜

    可以提供百度云链接的权重文件(libtorch和pytorch版本均需)吗?国内木得梯子

非常感谢!!

BIT-MJY commented 2 years ago

Hello @oym1994 ,

感谢支持我们的工作,我已经上传了我们的预训练权重至百度云,链接如下:

pretrained_overlap_transformer.pth.tar --- Extraction Code: kftk

overlapTransformer.pt --- Extraction Code: qhmg

oym1994 commented 2 years ago

好的,再次非常感谢

oym1994 commented 2 years ago

您好,请问可以再提供cpu推理的模型吗?电脑还没得gpu, 谢谢啦

BIT-MJY commented 2 years ago

目前我们的预训练模型只有基于gpu训练的,cpu训练将耗费过多时间。

oym1994 commented 2 years ago

抱歉,表述不完整,用于c++推理的cpu版本权重。利用gen_libtorch_model.py生成了, 但这里需要做一些修改 feature_extracter_without_delta_layer-->overlap_transformer

另外, c++的代码里有hardcode,测试发现会造成运行时会出现段错误,如下修改后就可以了

原:gen_range_image(range_image, cloud0, 3, -25, 64, 900, 50);

修改后:gen_range_image(range_image, cloud0, 3, -25, height, width, 50);

BIT-MJY commented 2 years ago

谢谢提醒,gen_libtorch_model.py文件中做了如下修改:(Revision in gen_libtorch_model.py)

from modules.overlap_transformer import featureExtracter
...
...
example = torch.rand(1, 1, 64, 900)    # 64 for KITTI, 32 for Haomo

请注意example的尺寸,此前的文件此处为32

此外,fast_ot.cpp中的几处hardcode也进行了修改。(hard coding in fast_ot.cpp is revised)