Skip to content

augs.md

Latest commit

 

History

History
128 lines (92 loc) · 3.61 KB

augs.md

File metadata and controls

128 lines (92 loc) · 3.61 KB
Raw

Time-Series Augmentations

FedProC provides a dedicated, high-performance, GPU-native time-series augmentation library under the augs package. Built entirely in pure PyTorch, these augmentations eliminate CPU-GPU round-trip latency bottlenecks, enabling ultra-fast, learnable differentiable data augmentation directly on GPU memory.

These augmentations are fully compatible with batch inputs of shape [B, T, D] (Batch size, Time length, Dimensions/Features) and adapt automatically to the active device of the input tensor.

Augmentations API Reference

All augmentations are implemented as callable classes.

1. Cutout

Stochastically zeroes out a single contiguous temporal segment within the time-series.

  • Class: cutout(perc=0.1)
  • Parameters:
    • perc (float): The percentage of the sequence length to zero out.
from augs import cutout

aug = cutout(perc=0.15)
augmented_x = aug(x)  # x is [B, T, D]

2. Jitter

Adds random zero-mean Gaussian noise directly to the signal.

  • Class: jitter(sigma=0.3)
  • Parameters:
    • sigma (float): Standard deviation of the Gaussian noise.
from augs import jitter

aug = jitter(sigma=0.2)
augmented_x = aug(x)

3. Scaling

Scales each feature's magnitude along the entire sequence length stochastically using random Gaussian multiplier.

  • Class: scaling(sigma=0.5)
  • Parameters:
    • sigma (float): Standard deviation of the scaling factor distribution centered at 1.0.
from augs import scaling

aug = scaling(sigma=0.3)
augmented_x = aug(x)

4. Time Warp

Implements highly optimized, GPU-native time-warping using PyTorch linear interpolation over a random monotonic speed-change anchor grid.

  • Class: time_warp(n_speed_change=4, max_speed_ratio=2.0)
  • Parameters:
    • n_speed_change (int): Number of speed change segments.
    • max_speed_ratio (float): Maximum speed scaling ratio.
from augs import time_warp

aug = time_warp(n_speed_change=3, max_speed_ratio=1.5)
augmented_x = aug(x)

5. Magnitude Warp

Applies the monotonic warping function along the channel/feature magnitude dimension instead of the temporal dimension.

  • Class: magnitude_warp(n_speed_change=4, max_speed_ratio=2.0)
  • Parameters: Same as time_warp.
from augs import magnitude_warp

aug = magnitude_warp(n_speed_change=3, max_speed_ratio=1.5)
augmented_x = aug(x)

6. Window Slice

Stochastically crops a random window of the time series and linear interpolates it back to its original length T.

  • Class: window_slice(reduce_ratio=0.5)
  • Parameters:
    • reduce_ratio (float): The fraction of the original sequence length to keep.
from augs import window_slice

aug = window_slice(reduce_ratio=0.7)
augmented_x = aug(x)

7. Window Warp

Extracts a random window of the time series, compresses or expands it stochastically by a factor selected from scales, and then linear interpolates the concatenated sequences back to the original sequence length T.

  • Class: window_warp(window_ratio=0.3, scales=[0.5, 2.0])
  • Parameters:
    • window_ratio (float): The percentage length of the window to warp.
    • scales (list): Candidates for window length scale ratio.
from augs import window_warp

aug = window_warp(window_ratio=0.3, scales=[0.6, 1.8])
augmented_x = aug(x)

8. Subsequence

Stochastically crops a random subsequence, leaving it active, and zero-masks all elements outside the crop boundaries.

  • Class: subsequence()
from augs import subsequence

aug = subsequence()
augmented_x = aug(x)