Padding frame to a multiple of an integer n

[thibo73800]

The method SpatialAugmentedTensor.pad() could accept a multiple argument. This argument would result in padding the tensor spatial dimensions to a multiple of multiple.

Some flag(s) should indicate if we should pad the labels or not. We could have one boolean flag pad_label, or make it possible to specify which labels should be padded.

Unpadding
Padding to a multiple is often used before performing inference with a model. Given the type of outputs of the model, it is sometimes necessary to unpad the outputs. We should provide a method to unpad too.

[tflahaul]

@thibo73800 Can you give me more explanations on how the multiple argument should be used ? I guess its different than a scale factor ?

[thibo73800]

Ok, this is the case in RAFT inference. Padding is currently achieve using an external transformation because in the model input should be a multiple of 8.

aloception-oss/alonet/raft/eval_on_sintel.py

Line 46 in fd8e8c9

padder = Padder() # pads inputs to multiple of 8

Once the inference is done, the padder will unpad the the prediction.

aloception-oss/alonet/raft/eval_on_sintel.py

Line 62 in fd8e8c9

flow_pred = padder.unpad(flow_pred)[0]

Note that this PR could be hard to implement completely if you want to properly handle all scenearios

• The attached labels
• Padding multiple time an already padded tensor
• Unpadding or not the labels...

Proper unit test must be done to complete it. You can start by exploring padding unit test implemented for the boxes here
https://github.com/Visual-Behavior/aloception-oss/blob/master/unittest/test_boxes.py

[tflahaul]

• padding to the next multiple of arg multiple
• unit tests for the tensor
• unit tests for the children
• unpadding to the original shape (using a _pad_parameters ATen stack attribute?)
• unpadding unit tests