Question about the prune rate

Hi,

Recently I tried the code and found that there seems to have inconsistency of the definition of prune rate throughout the code.

In builder.py, when the scale_fan is True, the fan is scaled by a factor of (1-prune_rate), where it denotes the top k% of weights as indicated in the paper. So here prune_rate means the ratio of weights to be removed. This meaning is consistent with main.py line 318, which calculates the number of parameters in the subnetwork.

But in conv_type.py lines 20-25, when generating the subnet mask, (1-prune_rate) of the weights are removed:
`j = int((1 - k) * scores.numel())
flat_out = out.flatten()
flat_out[idx[:j]] = 0
flat_out[idx[j:]] = 1`
which seems incorrect. So I think this is a bug and line 20 should be changed to
`j = int(k * scores.numel())`

Could you please clarify a bit what exactly does the prune_rate means in your code?

Thanks