The clamp of DAQActivations might has a bit problem.

[Neilhy]

https://github.com/mike9251/DAQ/blob/main/daq.py#L49

x_clamped[:, c, ...] = torch.clamp(x[:, c, ...], -self.alpha + self.beta[0, c, 0, 0].item(),
                                                     self.alpha + self.beta[0, c, 0, 0].item())

might be

x_clamped[:, c, ...] = torch.clamp(x[:, c, ...],  -self.alpha + self.beta[0, c, 0, 0].item() - self.step_size, 
                                                    self.alpha + self.beta[0, c, 0, 0].item())

according to paper.

[mike9251]

Hi @Neilhy,

I don't remember details of the paper. Can you tell me where exactly to look in the paper for this information?

Did you notice any problems without the proposed modification?