evolutionary-policy-optimization 0.0.55__tar.gz → 0.0.56__tar.gz

{evolutionary_policy_optimization-0.0.55 → evolutionary_policy_optimization-0.0.56}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.0.55
+Version: 0.0.56
 Summary: EPO - Pytorch
 Project-URL: Homepage, https://pypi.org/project/evolutionary-policy-optimization/
 Project-URL: Repository, https://github.com/lucidrains/evolutionary-policy-optimization

{evolutionary_policy_optimization-0.0.55 → evolutionary_policy_optimization-0.0.56}/evolutionary_policy_optimization/epo.py

@@ -374,13 +374,16 @@ class LatentGenePool(Module):
         should_run_genetic_algorithm: Module | None = None, # eq (3) in paper
         default_should_run_ga_gamma = 1.5,
         migrate_every = 100,                 # how many steps before a migration between islands
-        apply_genetic_algorithm_every = 2    # how many steps before crossover + mutation happens for genes
+        apply_genetic_algorithm_every = 2,   # how many steps before crossover + mutation happens for genes
+        init_latent_fn: Callable = None
     ):
         super().__init__()
 
         maybe_l2norm = l2norm if l2norm_latent else identity
 
-        latents = torch.randn(num_latents, dim_latent)
+        init_fn = default(init_latent_fn, torch.randn)
+
+        latents = init_fn((num_latents, dim_latent))
 
         if l2norm_latent:
             latents = maybe_l2norm(latents, dim = -1)

{evolutionary_policy_optimization-0.0.55 → evolutionary_policy_optimization-0.0.56}/evolutionary_policy_optimization/experimental.py

@@ -39,9 +39,34 @@ def crossover_weights(w1, w2, transpose = False):
 
     return out
 
+def mutate_weight(
+    w,
+    transpose = False,
+    mutation_strength = 1.
+):
+
+    if transpose:
+        w = w.transpose(-1, -2)
+
+    rank = min(w2.shape[1:])
+    assert rank >= 2
+
+    u, s, v = torch.svd(w)
+    u = u + torch.randn_like(u) * mutation_strength
+    v = v + torch.randn_like(v) * mutation_strength
+
+    out = u @ torch.diag_embed(s) @ v.mT
+
+    if transpose:
+        out = out.transpose(-1, -2)
+
+    return out
+
 if __name__ == '__main__':
     w1 = torch.randn(32, 16)
     w2 = torch.randn(32, 16)
-    child = crossover_weights(w2, w2)
+
+    child = crossover_weights(w1, w2)
+    mutated_w1 = mutate_weight(w1)
 
     assert child.shape == w2.shape

{evolutionary_policy_optimization-0.0.55 → evolutionary_policy_optimization-0.0.56}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "evolutionary-policy-optimization"
-version = "0.0.55"
+version = "0.0.56"
 description = "EPO - Pytorch"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }