xuyu0010 / atcon Goto Github PK

Thanks for your great work and fast response to issues.
It said that more specifications on benchmark details and network implementation are provided in the Appendix.
As far as I know, the Appendix does not exist(or not opened yet?).
So, I want to implement your method with datasets except for UCF,HMDB and would you let me know the experiment configuration?

I want to know where to get the pretrain model "pretrained/tsn2d_rgb_r50.pth". And the code in the file train_source_model.py, there is two parameter “end_epoch” in the function train_source_model.

Hi Authors. Thanks for the excellent work. Could you please share your source-only models that are trained with source datasets. Thanks a lot.

Thanks for your great work and fast response to issues.
In the paper, it is written that all methods adopt the TRN as the backbone pre-trained on 'ImageNet'. As this might imply, it will be reasonable that the repored UCF-HMDB full
results and reimplemented results differ not so many.
But, all the configuration is all the same with your file but the result is quite higher than table in the paper and the training procedure saturates so fast due to pre-trained weight.
So, I wonder whether the backbone is pre-trained on 'ImageNet' or not (Kinetics400?) and think that you might mistake which is used to pre-training backbone.

I'm sorry to bother you again, I also can't find the pretrain model tsn2d_kinetics_rgb_r50.pth'.

The coding here(

Thanks for your great work.
I've been downloading DA datasets for implementing your work.
There is no problem except for downloading and extracting the Sports1M dataset.
Downloaded from Google drive, it must be some parted zip data.
So, do other parts of the dataset exists?

I would like to know if i want to get the result of Daily-DA（H51->A11), is it true to use hmdb51_msda_train.txt for source training and arid_msda_train.txt for target unlable training and testing