| import torch.nn as nn | |
| from itertools import repeat | |
| from typing import Iterable | |
| def _ntuple(n): | |
| """Copy item to be a tuple with n length (Implemented as timm) | |
| """ | |
| def parse(x): | |
| if isinstance(x, Iterable): | |
| return x | |
| else: | |
| return tuple(repeat(x, n)) | |
| return parse | |
| to_2tuple = _ntuple(2) | |
| to_ntuple = _ntuple | |
| class DropPath(nn.Module): | |
| """Stochasticly zero channels of data.(Implemented as timm) | |
| """ | |
| def __init__(self, drop=0.5, scale=True): | |
| super().__init__() | |
| self.drop = drop | |
| self.scale = scale | |
| def forward(self, x): | |
| return self.drop_path(x, self.drop, self.training, self.scale) | |
| def drop_path(self, x, drop=0.5, training=True, scale=True): | |
| if drop == 0. or not training: | |
| return x | |
| drop_p = 1 - drop | |
| shape = (x.shape[0],) + (1,) * (x.ndim - 1) | |
| random_tensor = x.new_empty(shape).bernoulli_(drop_p) | |
| if drop_p > 0. and scale: | |
| random_tensor.div_(drop_p) | |
| return x * random_tensor |