DropBlock1D

[kayuksel]

Can you please also include DropBlock1D implementation to use it for time-series? Thank you very much.

[kayuksel]

Would this be a right implementation?

class DropBlock1D(nn.Module):
    def __init__(self, drop_prob = 0.2, block_size = 3):
        super(DropBlock1D, self).__init__()
        self.p = drop_prob
        self.bs = block_size

    def forward(self, x):
        if not self.training: return x
        gamma = self.p / (self.bs ** 2)
        mask = (torch.rand(x.shape[0], *x.shape[2:]) < gamma)
        bm = self._compute_block_mask(mask.float().to(x.device))
        return x * bm[:, None, :] * bm.numel() / bm.sum()

    def _compute_block_mask(self, mask):
        block_mask = F.max_pool1d(input=mask[:, None, :],
        kernel_size=self.bs, stride=1, padding=self.bs // 2)
        if self.bs % 2 == 0: block_mask = block_mask[:, :, :-1]
        return (1 - block_mask.squeeze(1))

[miguelvr]

looks fine to me. you can make a PR with that code and add a few unit tests (check the current ones). I'll happily review it for you.

[kayuksel]

@miguelvr Thanks for your response. I now checked many of the existing unit tests. Also, using it already in two projects where in both it contributed. I am not very familiar with pull requests, etc. Therefore, please feel free to use the above block within the project if you would like including it.