Closed ZWXCV closed 6 months ago
From the configs, try replacing the structure from hitnet
to serial
. And please tell me the results you got
Do you mean to change the STRUCTURE: "hitnet" to STRUCTURE: "serial"? However, an error occurs when the training is performed after modification:
Traceback (most recent call last):
File "train_net.py", line 255, in
The result shows an incorrect number of parameters
The result before modification was:
VideoAP_ 0.5: 84.80
VideoAP_ 0.2: 86.40
Yes, hitnet
to serial
as you did.
Thanks for reporting the numbers. Others have pointed out disparities of results for different runs (though I used seed). I will fix the serial
bug and get back to you.
By the way, what is the frame mAP result for that run?
frame mAP : 81.02
Hello author, I found that the weights of some parameters were not loaded during the training process, will this affect the final training results and mAP calculation? If so, what is the cause? Is the weight file I loaded incomplete? Could the author give me some suggestions, thank you very much. Here are some log messages from the training process: 2023-11-12 10:39:16,803 hit.utils.model_serialization INFO: backbone.slow.res_nl4.res_2.btnk.conv3.bn.weight loaded from backbone.slow.res_nl4.res_2.btnk.conv3.bn.weight of shape (2048,) 2023-11-12 10:39:16,803 hit.utils.model_serialization INFO: backbone.slow.res_nl4.res_2.btnk.conv3.conv.weight loaded from backbone.slow.res_nl4.res_2.btnk.conv3.conv.weight of shape (2048, 512, 1, 1, 1) 2023-11-12 10:39:16,803 hit.utils.model_serialization INFO: roi_heads.action.feature_extractor.fc1.bias will not be loaded. 2023-11-12 10:39:16,803 hit.utils.model_serialization INFO: roi_heads.action.feature_extractor.fc1.weight will not be loaded.
In your case, weights not being loaded might be because you are training from an existing checkpoint.
I uploaded the pretrained model. Possibly due to the dataset being very small, different runs might give different results. I also tested different checkpoints (for JHMDB the model converges very fast then start overfitting).
Thank you for the author's reply, but why is the value of video mAP that I received significantly different from the value given in your paper, about 3% or 5%? Should I make modifications to certain parameters to achieve the results shown in your paper?I hope the author can provide some suggestions. Thank you very much