File size: 1,363 Bytes
d6682b6 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 |
import torch
def magnitude(tensor: torch.Tensor, density: float) -> torch.Tensor:
"""Masks out the smallest values, retaining a proportion of `density`."""
if density >= 1:
return tensor
k = int(density * tensor.view(-1).shape[0])
assert k > 0, "not gonna zero out the whole tensor buddy"
mask = torch.zeros_like(tensor)
w = tensor.abs().view(-1)
if w.device.type == "cpu":
w = w.float()
topk = torch.topk(w, k=k, largest=True)
mask.view(-1)[topk.indices] = 1
return tensor * mask
def bernoulli(
tensor: torch.Tensor, density: float, rescale: bool = True
) -> torch.Tensor:
if density >= 1:
return tensor
if (tensor.device.type != "cpu") or tensor.dtype == torch.bfloat16:
work_dtype = tensor.dtype
else:
# torch.bernoulli not implemented for float16 on CPU, upcast to float32
work_dtype = torch.float32
mask = torch.bernoulli(
torch.full_like(input=tensor, fill_value=density, dtype=work_dtype)
)
res = tensor.to(work_dtype) * mask
if rescale:
res /= density
return res.to(tensor.dtype)
def rescaled_random(tensor: torch.Tensor, density: float):
return bernoulli(tensor, density, rescale=True)
def random_wo_rescaled(tensor: torch.Tensor, density: float):
return bernoulli(tensor, density, rescale=False) |