Closed SixGodGG closed 3 years ago
maybe try to disable memory pinning in the data loader
By changing line 62 in run_training.py
and following from this:
dataloader = DataLoader(train_data, batch_size=batch_size, drop_last=True,
shuffle=True, num_workers=workers, collate_fn=cut_paste_collate_fn,
persistent_workers=True, pin_memory=True, prefetch_factor=5)
to this:
dataloader = DataLoader(train_data, batch_size=batch_size, drop_last=True,
shuffle=True, num_workers=workers, collate_fn=cut_paste_collate_fn,
persistent_workers=True, pin_memory=False, prefetch_factor=5)
For more information about memory pinning see the PyTorch docs here.Turning it off just results in a performance hit but the code should still work.
I think this issue is related: https://discuss.pytorch.org/t/persistent-workers-true-for-the-dataloader-causes-errors/106669
I'm using PytTorch version 1.8.1 maybe try to upgrade your PyTorch version
maybe try to disable memory pinning in the data loader By changing line 62 in
run_training.py
and following from this:dataloader = DataLoader(train_data, batch_size=batch_size, drop_last=True, shuffle=True, num_workers=workers, collate_fn=cut_paste_collate_fn, persistent_workers=True, pin_memory=True, prefetch_factor=5)
to this:
dataloader = DataLoader(train_data, batch_size=batch_size, drop_last=True, shuffle=True, num_workers=workers, collate_fn=cut_paste_collate_fn, persistent_workers=True, pin_memory=False, prefetch_factor=5)
For more information about memory pinning see the PyTorch docs here.Turning it off just results in a performance hit but the code should still work.
i set pin_memory=False and solved this issue
I think this issue is related: https://discuss.pytorch.org/t/persistent-workers-true-for-the-dataloader-causes-errors/106669
I'm using PytTorch version 1.8.1 maybe try to upgrade your PyTorch version
did u tried to run this code at windows10? and if u have tried use GPU?
Yes it's running on my Machine with windows 10 and a GPU.
Yes it's running on my Machine with windows 10 and a GPU.
nice,and how many work_num u set? i set --workers = 1. how many yours? and where are u come from
I currently do not have my Windows 10 machine at hand but I think I'm using the default of 8.
Windows multithreading is in my experience sometimes not that well supported by deep learning frameworks.
You can also disable the loading of the data in another thread by setting workers=0
.
As described in the PyTorch documentation here.
But this results in longer training times.
I'm from Germany. What about you?
I currently do not have my Windows 10 machine at hand but I think I'm using the default of 8. Windows multithreading is in my experience sometimes not that well supported by deep learning frameworks. You can also disable the loading of the data in another thread by setting
workers=0
. As described in the PyTorch documentation here. But this results in longer training times.I'm from Germany. What about you?
asia, and thanks for your response, i will try to read code and paper again. now training have succeed.
now training have succeed.
Nice.
I also added my results to the readme
now training have succeed.
Nice.
I also added my results to the readme
i have seen your result,and i got some same result too, when finish the (implement gradCam) work and (Patch Heatmap) work, its very useful
when finish the (implement gradCam) work and (Patch Heatmap) work, its very useful
I'm currently not planning to implement gradCam or the patch heatmap.
when finish the (implement gradCam) work and (Patch Heatmap) work, its very useful
I'm currently not planning to implement gradCam or the patch heatmap.
hi, how to understand your hyperparameters --head_layer ,i try to set it 1 or 2 and get different results, some type better some became bad, i read whole paper, and cant find the details about --head_layer, does it come from level-7 feature? could u help me to explan it? how to amend this hyperparameters , it affect the time ?
when finish the (implement gradCam) work and (Patch Heatmap) work, its very useful
I'm currently not planning to implement gradCam or the patch heatmap.
hi, do u have any other communication tool? i want to talk with u,sir. please, my ins: triple_night_sss have some question to talk,thank you and can be pay pp if necessary
anyone seen this problem?
window10 gtx1650 parser.add_argument('--workers', default=1, type=int, help="number of workers to use for data loading (default:8)")
i have set --workers = 1