PyTorchGAF: PyTorch-accelerated implementation of the Gramian Angular Field (GAF)

The provided GAF implementation uses pure PyTorch. It provides the following benefits and features:

• Supports batched data.
• Supports multivariate time series data.
• Handles necessary scaling before the GAF transformation.
• Allows to use the GAF transform directly as a neural network layer.
• Allows directly transforming the data on the target device (CPU or GPU).
• Uses efficient vectorized Einstein Summation Notation to compute the outer products to avoid using loops.

Optimizations

Our implementation is based on the GAF transform from pyts. However, as Figure 1 shows, the GAF transform significantly benefits from GPU acceleration, achieving speedups of almost 90x for larger batch sizes.

Figure 1: Execution time comparison between CPU-based pyts implementation and our implementation on CPU and GPU.

Usage/Examples

import torch

gaf = GAFTransform(method="summation")
inputs = torch.randn(32, 8, 40, device="cuda")  # (N, C, L)
output = gaf(inputs)  # (N, C, L, L)

Acknowledgements

• pyts GitHub
• pyts GAF