dtw_loss_functions.soft_dtw_implementations.soft_dtw_cuda_ron.cuda.launcher module

dtw_loss_functions.soft_dtw_implementations.soft_dtw_cuda_ron.cuda.launcher.softdtw_backward_cpu(D: Tensor, R: Tensor, gamma: float, bandwidth: float)

dtw_loss_functions.soft_dtw_implementations.soft_dtw_cuda_ron.cuda.launcher.softdtw_backward_cuda_fused_sqeuclid(X: Tensor, Y: Tensor, R: Tensor, gamma: float, bandwidth: float)

Fused SoftDTW backward (log-space) for squared-euclidean distance that does NOT materialize D_pad.

Inputs:

X: (B,N,D) CUDA Y: (B,M,D) CUDA R: (B,N+2,M+2) CUDA (from forward)

Returns:

(B,N,M) CUDA (E = d SoftDTW / d D in linear space, via exp(logE))

Return type:

E

dtw_loss_functions.soft_dtw_implementations.soft_dtw_cuda_ron.cuda.launcher.softdtw_backward_cuda_log(D: Tensor, R: Tensor, gamma: float, bandwidth: float)

dtw_loss_functions.soft_dtw_implementations.soft_dtw_cuda_ron.cuda.launcher.softdtw_forward_cpu(D: Tensor, gamma: float, bandwidth: float)

dtw_loss_functions.soft_dtw_implementations.soft_dtw_cuda_ron.cuda.launcher.softdtw_forward_cuda(D: Tensor, gamma: float, bandwidth: float)

dtw_loss_functions.soft_dtw_implementations.soft_dtw_cuda_ron.cuda.launcher.softdtw_forward_cuda_fused_sqeuclid(X: Tensor, Y: Tensor, gamma: float, bandwidth: float)

Fused SoftDTW forward for squared-euclidean distance that does NOT materialize D (B,N,M).

X: (B,N,D), Y: (B,M,D) CUDA tensors Returns: (out: (B,), R: (B,N+2,M+2))