Implemented Perlin as underlying epsilon dist

priorConditionedAnnealing/noise.py

@@ -1,6 +1,7 @@
 import numpy as np
 import torch as th
 import colorednoise as cn
+from perlin_noise import PerlinNoise
 from torch.distributions import Normal
 
 
@@ -78,3 +79,23 @@ class SDE_Noise():
         # (batch_size, 1, n_actions)
         noise = th.bmm(th.bmm(latent_sde, self.exploration_matrices), chol)
         return noise.squeeze(dim=1)
+
+
+class Perlin_Noise():
+    def __init__(self, known_shape=None, scale=0.1, octaves=1):
+        self.known_shape = known_shape
+        self.scale = scale
+        self.octaves = octaves
+        self.magic = 3.14159  # Axis offset
+        # We want to genrate samples, that approx ~N(0,1)
+        self.normal_factor = 0.0471
+        self.reset()
+
+    def __call__(self, shape):
+        self.index += 1
+        return [self.noise([self.index*self.scale, self.magic*a]) / self.normal_factor
+                for a in range(self.shape[-1])]
+
+    def reset(self):
+        self.index = 0
+        self.noise = PerlinNoise(octaves=self.octaves)