Training stops in epoch 10 with Assertion Error

[vardeep-sandhu]

Hey,

So I was training Stratified Transformer on my own dataset. Till epoch 10 everything worked fine and nothing was going wrong. But during epoch 10 this error is thrown below. It is related to the relative_position_index here in the position encoding part, where it checks if all the indices are positive. I am just thinking why would any index be negative in the first place. A little more information, I am using global coordinates of points and am also doing some custom augmentations which flip and rotate the point cloud. During the first 10 epochs, the validation results are better than with simple augs, so the augs are working well.

As a solution, I thought if I made sure that the relative position here is always positive by taking absolute values of this subtraction, how would that affect the training.

Let me know what you think

[X-Lai]

Here we use contextual relative position encoding (cRPE). For more detail, please refer to our paper.