On Fri, Mar 23, 2018 at 14:59 Francisco Massa ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In torch/nn/functional.py <#5968 (comment)>: > + raise ValueError('Expected number of channels in input to be divisible ' + 'by num_groups, but got {} input and num_groups={}' + .format(input_shape, num_groups)) + + if weight is not None and (weight.dim() != 1 or weight.numel() != c): + raise ValueError('Expected weight to be a vector of size equal to the ' + 'number of channels in input, but got {} weight and {} ' + 'input'.format(weight.size(), input_shape)) + + if bias is not None and (bias.dim() != 1 or bias.numel() != c): + raise ValueError('Expected bias to be a vector of size equal to the ' + 'number of channels in input, but got {} bias and {} ' + 'input'.format(bias.size(), input_shape)) + + # Apply group norm + input_reshaped = input.contiguous().view(1, b * g, -1) nit: this can be simplified with .reshape(1, b * g, -1), which was added in #5575 <#5575> and consists roughly as .contiguous().view() — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#5968 (review)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AFaWZfssR6F0bmL77IJKzDYzMTR_9a_jks5thUYLgaJpZM4S5Bef> .

On Fri, Mar 23, 2018 at 16:30 Edward Z. Yang ***@***.***> wrote: ***@***.**** commented on this pull request. ------------------------------ In torch/nn/functional.py <#5968 (comment)>: > @@ -1308,6 +1318,53 @@ def layer_norm(input, normalized_shape, running_mean, running_var, return out +def group_norm(input, num_groups, weight=None, bias=None, eps=1e-5): Why not in ATen? — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#5968 (review)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AFaWZZaRyyWs7ZFKViheK4m8a7lpP171ks5thVtMgaJpZM4S5Bef> .