Unify visualisation

[rijobro]

Unify visualisation classes.

Description

Beforehand, occlusion sensitivity was a function in metrics, whereas gradcam functionality was in classes inside the visualize module. Now, both can be called with similar calls:

from monai.networks.nets import densenet121
from monai.visualise import OcclusionSensitivity
from monai.visualise import CAM
model_2d = densenet121(spatial_dims=2, in_channels=1, out_channels=3)

cam = CAM(nn_module=model_2d, target_layers="class_layers.relu", fc_layers="class_layers.out")
result_cam = cam(x=torch.rand((1, 1, 48, 64)))

occ_sens = OcclusionSensitivity(nn_module=model_2d)
result_occ_sens = occ_sens(x=torch.rand((1, 1, 48, 64)), class_idx=None, b_box=[-1, -1, 2, 40, 1, 62])

We also had mixed American and British spelling, so changed all instances of visualize to visualise. Maybe I should have gone the other way, as it made lots of changes...

Status

Ready

Types of changes

• Non-breaking change (fix or new feature that would not break existing functionality).
• New tests added to cover the changes.
• Integration tests passed locally by running ./runtests.sh --codeformat --coverage.
• Quick tests passed locally by running ./runtests.sh --quick.
• In-line docstrings updated.
• Documentation updated, tested make html command in the docs/ folder.

[wyli]

I feel we should go for visualize... to be consistent with pytorch and numpy
thanks for raising this!

[rijobro]

ok, updated.

[wyli]

thanks for the clean up! but I think it's too early to define a proper NetVisualizer API, we should keep the module simple for now, and it needs more discussions along with #528

[rijobro]

but I think it's too early to define a proper NetVisualizer API

I can remove it if you'd like but it's only a skeleton at the moment, with a constructor and __call__ method. Just so that when future visualisations get added, they are more likely to conform (avoid the difference we saw with OccSens and GradCAM).

Up to you, I'll delete the base class if you'd like, but I'm in favour of keeping it.

[wyli]

but I think it's too early to define a proper NetVisualizer API

I can remove it if you'd like but it's only a skeleton at the moment, with a constructor and __call__ method. Just so that when future visualisations get added, they are more likely to conform (avoid the difference we saw with OccSens and GradCAM).

Up to you, I'll delete the base class if you'd like, but I'm in favour of keeping it.

I feel the benefit of having a NetVisualizer class is not very clear at the moment (and the interface is also not well documented), all the other changes look very nice. perhaps @Nic-Ma @ericspod could have more comments.

[rijobro]

Ok, I can remove it.

I also created a train_mode context manager (opposite of eval_mode) here. We can use this when calculating the gradient for GradCam, which should avoid the problem you saw here: Project-MONAI/tutorials#83 (comment).

[wyli]

thanks!