Closed AndyYuan96 closed 2 years ago
I use pytorch1.6 and cuda10.1, other mmdetection3d related environments is same with you.
what's more, can you give me some advise for how to train LC.py, I download imgpretrained model from mmdetection3d, and in config file, model's dict, I add pertained = {'img': model.pth}, does it is ok? as I see lots of log say mis match, I am not familiar with mmdetection3d.
Hi, I use 8 v100 to train pillar_L.py, no change for config, kill after epoch15, and comment the gtaug, my final mAP is 0.5326, NDS: 0.6094, do you think is normal? as lower more than 1 point than readme.md result, what is the variance of result after training multi times?
The performance gap seems to be larger than the normal disturbance. Could you please share the detailed results (mATE, mAOE, etc) and your training log?
what's more, can you give me some advise for how to train LC.py, I download imgpretrained model from mmdetection3d, and in config file, model's dict, I add pertained = {'img': model.pth}, does it is ok? as I see lots of log say mis match, I am not familiar with mmdetection3d.
If you use the same mmdet3d version with me, you can use the following script to combine the pretrained TransFusionL and the 2D backbone
img = torch.load('img_backbone.pth', map_location='cpu')
pts = torch.load('transfusionL.pth', map_location='cpu')
new_model = {"state_dict": pts["state_dict"]}
for k,v in img["state_dict"].items():
if 'backbone' in k or 'neck' in k:
new_model["state_dict"]['img_'+k] = v
torch.save(new_model, "fusion_model.pth")
Then you can set the load_from
of your config file to the path of fusion_model.pth
Hi, I use 8 v100 to train pillar_L.py, no change for config, kill after epoch15, and comment the gtaug, my final mAP is 0.5326, NDS: 0.6094, do you think is normal? as lower more than 1 point than readme.md result, what is the variance of result after training multi times?
The performance gap seems to be larger than the normal disturbance. Could you please share the detailed results (mATE, mAOE, etc) and your training log?
what's more, can you give me some advise for how to train LC.py, I download imgpretrained model from mmdetection3d, and in config file, model's dict, I add pertained = {'img': model.pth}, does it is ok? as I see lots of log say mis match, I am not familiar with mmdetection3d.
If you use the same mmdet3d version with me, you can use the following script to combine the pretrained TransFusionL and the 2D backbone
img = torch.load('img_backbone.pth', map_location='cpu') pts = torch.load('transfusionL.pth', map_location='cpu') new_model = {"state_dict": pts["state_dict"]} for k,v in img["state_dict"].items(): if 'backbone' in k or 'neck' in k: new_model["state_dict"]['img_'+k] = v torch.save(new_model, "fusion_model.pth")
Then you can set the
load_from
of your config file to the path offusion_model.pth
can I send the picture of the log to your email? as the file in in my company's computer.
Sure, you can send it to xbaiad@connet.ust.hk
@AndyYuan96 I haven't received your email.
@AndyYuan96 I haven't received your email.
Sorry,I sent to you quickly。
@AndyYuan96 I haven't received your email.
I try to use my school email and gmail email, both can't send email to xbaiad@connet.ust.hk, the reason is that system can't find connet.ust.hk.
@AndyYuan96 I haven't received your email.
Hi,xuyang,now I can reproduce the result of pillar backbone, it’s gtaug problem, with batch_size=4, I can get map 55.58, nds 63.01, thanks for sharing the great work.
Hi, thanks for your feedback, what do you mean by it's gt aug problem
? could you explain a little bit as someone else might also meet this problem.
@AndyYuan96 I haven't received your email.
Hi,xuyang,now I can reproduce the result of pillar backbone, it’s gtaug problem, with batch_size=4, I can get map 55.58, nds 63.01, thanks for sharing the great work.
@AndyYuan96 Hi, what's the gt aug problem? I maybe meet the same problem with you. thanks for your reply
Hi, I use 8 v100 to train pillar_L.py, no change for config, kill after epoch15, and comment the gtaug, my final mAP is 0.5326, NDS: 0.6094, do you think is normal? as lower more than 1 point than readme.md result, what is the variance of result after training multi times?