Reference to a runtime error: One of the variables needed to calculate the gradient has been changed using the built-in operation: [torch.cuda.FloatTensor [3, 3]], which is the result of 0 for ReluBackward0, has version 1; version 0 was expected instead

1

I’m trying to train a model using this code GitHub - toandaominh1997/EfficientDet.Pytorch: Implementation EfficientDet: Scalable and Efficient Object Detection in PyTorch but such a mistake comes out, what should I do?

2

This generally indicates that the backward pass is failing because the way the gradient flow is taking place is breaking down because one of the components is breaking the pass.

I imagine the model itself seems faulty.