| from __future__ import annotations |
|
|
| import numpy as np |
| import torch |
|
|
|
|
| def multiscale_batch(batch: int, length: int, seed: int) -> torch.Tensor: |
| rng = np.random.default_rng(seed) |
| time = np.arange(length, dtype=np.float32)[None] |
| values = np.zeros((batch, length), dtype=np.float32) |
| for low, high, scale in [(6, 12, 0.45), (20, 40, 0.35), (70, 110, 0.55)]: |
| periods = rng.uniform(low, high, size=(batch, 1)) |
| phases = rng.uniform(0, 2 * np.pi, size=(batch, 1)) |
| amplitudes = rng.uniform(0.5, 1.0, size=(batch, 1)) * scale |
| values += amplitudes * np.sin(2 * np.pi * time / periods + phases) |
| values += rng.normal(0, 0.015, size=values.shape) |
| return torch.from_numpy(values).float().unsqueeze(2) |
|
|
|
|
