Checkpoint handler save files with wrong permissions

[Nic-Ma]

🐛 Bug description

Hi @vfdev-5 ,

We always run programs in dockers, usually it's root user, and found an issue that: pure PyTorch program saves model(torch.save) and has the read permission for non-root users, but ignite Checkpoint handler saves model and only has read permission for root users.
You can reproduce this issue by below 2 MONAI examples:

1. https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/torch/unet_training_dict.py
   The saved model will be:

-rw-r--r-- 1 root root 19253862 Mar 26 02:08 best_metric_model_segmentation3d_dict.pth

2. https://github.com/Project-MONAI/tutorials/blob/master/3d_segmentation/ignite/unet_training_dict.py
   The saved model will be:

-rw------- 1 root root 57738580 Mar 26 01:54 net_checkpoint_50.pt

ignite version is 0.4.4, could you please help take a look at this issue?
As we are close to MONAI and Clara next release, it's an urgent bug for us.

Thanks in advance.

[vfdev-5]

@Nic-Ma sure! I'll take a look at this issue today.

[vfdev-5]

@Nic-Ma workaround now is to set atomic=False in ModelCheckpoint or DiskSaver.

I'll fix the issue for atomic=True in a PR.

Thanks for reporting !

[Nic-Ma]

Thanks for your quick response.
Let me submit a PR to set False in MONAI first.

Thanks.

[Nic-Ma]

Hi @vfdev-5 ,

I tried atomic=False locally, seems it works, so I put a PR in MONAI, could you please help review it?
Project-MONAI/MONAI#1863

Thanks in advance.

[vfdev-5]

@Nic-Ma sure! Our #1880 should fix the issue and you could update MONAI code along with our new release in future.