why did you put them in brackets?

reverted (it was a legacy change from when I imported everything from prune here)

while I'm here, I think the whole scheme of conv.weight having a pruned weight, conv.weight_mask holding a mask and conv.weight_orig holding the original weight might not work well for the optimizer scheme.
I'll dig in to it as I review the PR further.

yea, this is doing the weight_norm trick.
This is going to break in silent and bad ways, unless users prune the weights of the network before initializing the optimizers. Actually even then it wont work, as optimizer will report the parameter() of a Conv2d to still be .weight and .bias, which wont be a leaf node (and that's not valid)

this works because you explicitly call m.parameters() after masking has happened (i guess?)

*optimizers

numpy hard-dependency is a no-go in pytorch core

np comment as above

np comment as above

numpy as a hard-dependency is a no-go in pytorch core

while all of these methods were exposed for taking in a module and name, what about if we want to use these methods functionally, directly on a weight Tensor?

For example:

weight = torch.randn(20, 30)
weight_pruned = prune.random_structured(weight, amount=0.4, axis=1)

as of now, these three lines will do it:

weight = torch.randn(20, 30)
rsp = prune.RandomStructuredPruningMethod(amount=0.4, dim=1)
pruned_weight = weight * rsp.compute_mask(weight, default_mask=torch.ones_like(weight))

honestly, doesn't look great.
Can we check if Module or Tensor and pack the Tensor if needed?

I asked the question because I suspect people will want to do it more often than you think.

hmmk, so what's your suggested api here? what does "pack the Tensor" mean?

first, I think prune.RandomStructuredPruningMethod is sooo much more verbosity than prune.random_structured. Can we have them align better. For example the class be called: prune.RandomStructured.

Second, I was wondering if something like: pruned_weight = rsp.prune(weight) that implicitly will do weight * rsp.compute_mask(weight, default_mask=torch.ones_like(weight))

I've changed them all to align with the function names and added a .prune method to the base pruning class. I added a quick test too.