You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
In the paper, the stochastic quantization was done by rounding up with probability p=clip(0.5x, 0, 1), and rounding down with probability 1-p. However, in the code it's done by adding random uniform noise before quantization: noise = output.new(output.shape).uniform_(-0.5, 0.5) output.add_(noise)
This noise does not depend on the magnitude of x. I wonder what is the reasoning behind this discrepancy?
The text was updated successfully, but these errors were encountered:
In the paper, the stochastic quantization was done by rounding up with probability p=clip(0.5x, 0, 1), and rounding down with probability 1-p. However, in the code it's done by adding random uniform noise before quantization:
noise = output.new(output.shape).uniform_(-0.5, 0.5)
output.add_(noise)
This noise does not depend on the magnitude of x. I wonder what is the reasoning behind this discrepancy?
The text was updated successfully, but these errors were encountered: