-
Notifications
You must be signed in to change notification settings - Fork 87
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
2 changed files
with
127 additions
and
107 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,110 @@ | ||
| submodule(nf_layernorm_layer) nf_layernorm_layer_submodule | ||
| implicit none | ||
| contains | ||
| module function layernorm_layer_cons() & | ||
| result(res) | ||
| type(layernorm_layer) :: res | ||
|
|
||
| res % eps = 1e-5 | ||
| end function layernorm_layer_cons | ||
|
|
||
| pure module subroutine forward(self, input) | ||
| class(layernorm_layer), intent(in out) :: self | ||
| real, intent(in) :: input(:, :) | ||
| real, allocatable :: normalized(:, :) | ||
| integer :: i | ||
|
|
||
| allocate(normalized(self % sequence_length, self % model_dimension)) | ||
|
|
||
| ! mu = x - MEAN_last_dim(x) | ||
| do concurrent(i = 1: self % model_dimension) | ||
| self % mu(:, i) = input(:, i) - (sum(input, dim=2) / self % model_dimension) | ||
| end do | ||
|
|
||
| ! square root of variance shifted be eps | ||
| self % sigma = sqrt((sum(self % mu ** 2, dim=2) / self % model_dimension) + self % eps) | ||
|
|
||
| ! normalize mu by variance by first axis | ||
| do concurrent(i = 1: self % model_dimension) | ||
| normalized(:, i) = self % mu(:, i) / self % sigma | ||
| end do | ||
|
|
||
| ! forward through trainable params gamma and beta | ||
| do concurrent(i = 1: self % sequence_length) | ||
| self % output(i, :) = normalized(i, :) * self % gamma + self % beta | ||
| end do | ||
|
|
||
| deallocate(normalized) | ||
| end subroutine forward | ||
|
|
||
| pure module subroutine backward(self, input, gradient) | ||
| class(layernorm_layer), intent(in out) :: self | ||
| real, intent(in) :: input(:, :) | ||
| real, intent(in) :: gradient(:, :) | ||
| real, allocatable :: one_over_sigma(:, :) | ||
| real, allocatable :: gradient_by_gamma_over_sigma(:, :) | ||
|
|
||
| allocate(one_over_sigma(self % sequence_length, self % model_dimension)) | ||
| allocate(gradient_by_gamma_over_sigma(self % sequence_length, self % model_dimension)) | ||
|
|
||
| one_over_sigma = (1 / spread(self % sigma, dim=2, ncopies=self % model_dimension)) | ||
| gradient_by_gamma_over_sigma = & | ||
| gradient & | ||
| * spread(self % gamma, dim=1, ncopies=self % sequence_length) & | ||
| * one_over_sigma | ||
|
|
||
| ! d_output/d_gamma = sum(d_output/d_y * mu/sigma) | ||
| self % d_gamma = sum(gradient * self % mu * one_over_sigma, dim=1) | ||
|
|
||
| ! d_output/d_beta = sum(d_output/d_y) * 1 | ||
| self % d_beta = sum(gradient, dim=1) | ||
|
|
||
| ! From this article: | ||
| ! https://robotchinwag.com/posts/layer-normalization-deriving-the-gradient-for-the-backward-pass/ | ||
| ! d_output/d_x = d_output/d_y * gamma/sigma | ||
| ! - d_output/d_y | ||
| ! - sum(d_output/d_y * gamma/sigma) / len | ||
| ! - mu * sum(d_output/d_y * gamma * mu * sigma^(03)) / len | ||
| self % gradient = & | ||
| gradient_by_gamma_over_sigma & | ||
| - spread(& | ||
| sum(gradient_by_gamma_over_sigma, dim=2),& | ||
| dim=2,& | ||
| ncopies=self % model_dimension& | ||
| ) / self % model_dimension & | ||
| - self % mu * spread(& | ||
| sum(gradient_by_gamma_over_sigma * self % mu * (one_over_sigma ** 2), dim=2),& | ||
| dim=2,& | ||
| ncopies=self % model_dimension& | ||
| ) / self % model_dimension | ||
|
|
||
| deallocate(one_over_sigma) | ||
| deallocate(gradient_by_gamma_over_sigma) | ||
| end subroutine backward | ||
|
|
||
| module subroutine init(self, input_shape) | ||
| class(layernorm_layer), intent(in out) :: self | ||
| integer, intent(in) :: input_shape(:) | ||
|
|
||
| if (size(input_shape) /= 2) then | ||
| error stop "LayerNorm Layer accepts 2D input" | ||
| end if | ||
| self % sequence_length = input_shape(1) | ||
| self % model_dimension = input_shape(2) | ||
|
|
||
| ! default initialization from PyTorch | ||
| allocate(self % gamma(self % model_dimension)) | ||
| self % gamma = 1. | ||
| allocate(self % beta(self % model_dimension)) | ||
| self % beta = 0. | ||
|
|
||
| allocate(self % d_gamma(self % model_dimension)) | ||
| allocate(self % d_beta(self % model_dimension)) | ||
| allocate(self % gradient(self % sequence_length, self % model_dimension)) | ||
|
|
||
| allocate(self % mu(self % sequence_length, self % model_dimension)) | ||
| allocate(self % sigma(self % sequence_length)) | ||
|
|
||
| allocate(self % output(self % sequence_length, self % model_dimension)) | ||
| end subroutine init | ||
| end submodule nf_layernorm_layer_submodule |