evolutionary-policy-optimization 0.0.54__py3-none-any.whl → 0.0.56__py3-none-any.whl

evolutionary_policy_optimization/distributed.py

@@ -25,6 +25,12 @@ def pad_dim_to(t, length, dim = 0):
 def is_distributed():
     return dist.is_initialized() and dist.get_world_size() > 1
 
+def get_world_and_rank():
+    if not is_distributed():
+        return 1, 0
+
+    return dist.get_world_size(), dist.get_rank()
+
 def maybe_sync_seed(device, max_size = int(1e6)):
     rand_int = torch.randint(0, max_size, (), device = device)
 

evolutionary_policy_optimization/epo.py

@@ -1,7 +1,8 @@
 from __future__ import annotations
 
-from functools import partial, wraps
 from pathlib import Path
+from math import ceil
+from functools import partial, wraps
 from collections import namedtuple
 from random import randrange
 
@@ -19,6 +20,7 @@ from einops.layers.torch import Rearrange
 
 from evolutionary_policy_optimization.distributed import (
     is_distributed,
+    get_world_and_rank,
     maybe_sync_seed,
     all_gather_variable_dim,
     maybe_barrier
@@ -372,13 +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)
@@ -1061,22 +1066,20 @@ class EPO(Module):
     def latents_for_machine(self):
         num_latents = self.num_latents
 
-        if not is_distributed():
-            return list(range(self.num_latents))
+        world_size, rank = get_world_and_rank()
 
-        world_size, rank = dist.get_world_size(), dist.get_rank()
         assert num_latents >= world_size, 'number of latents must be greater than world size for now'
         assert rank < world_size
 
-        pad_id = -1
-        num_latents_rounded_up = ceil(num_latents / world_size) * world_size
-        latent_ids = torch.arange(num_latents_rounded_up)
-        latent_ids[latent_ids >= num_latents] = pad_id
+        num_latents_per_machine = ceil(num_latents / world_size)
+
+        for i in range(num_latents_per_machine):
+            latent_id = rank * num_latents_per_machine + i
 
-        latent_ids = rearrange(latent_ids, '(world latents) -> world latents', world = world_size)
-        out = latent_ids[rank]
+            if latent_id >= num_latents:
+                continue
 
-        return out[out != pad_id].tolist()
+            yield i
 
     @torch.no_grad()
     def forward(
@@ -1093,7 +1096,7 @@ class EPO(Module):
 
         cumulative_rewards = torch.zeros((self.num_latents))
 
-        latent_ids = self.latents_for_machine()
+        latent_ids_gen = self.latents_for_machine()
 
         for episode_id in tqdm(range(self.episodes_per_latent), desc = 'episode'):
 
@@ -1109,7 +1112,7 @@ class EPO(Module):
 
             # for each latent (on a single machine for now)
 
-            for latent_id in tqdm(latent_ids, desc = 'latent'):
+            for latent_id in tqdm(latent_ids_gen, desc = 'latent'):
                 time = 0
 
                 # initial state

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-0.0.54.dist-info → evolutionary_policy_optimization-0.0.56.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.0.54
+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-0.0.56.dist-info/RECORD

@@ -0,0 +1,9 @@
+evolutionary_policy_optimization/__init__.py,sha256=0q0aBuFgWi06MLMD8FiHzBYQ3_W4LYWrwmCtF3u5H2A,201
+evolutionary_policy_optimization/distributed.py,sha256=7KgZdeS_wxBHo_du9XZFB1Cu318J-Bp66Xdr6Log_20,2423
+evolutionary_policy_optimization/epo.py,sha256=N7xmO3CRXeaJAy-2rysZg-DBvkZCZB2ySJT7Iq__r6w,35217
+evolutionary_policy_optimization/experimental.py,sha256=-IgqjJ_Wk_CMB1y9YYWpoYqTG9GZHAS6kbRdTluVevg,1563
+evolutionary_policy_optimization/mock_env.py,sha256=202KJ5g57wQvOzhGYzgHfBa7Y2do5uuDvl5kFg5o73g,934
+evolutionary_policy_optimization-0.0.56.dist-info/METADATA,sha256=o2-1eCh8MuQVd0SH0GiUBBIAcqdK7cceuiu093cuEA4,6213
+evolutionary_policy_optimization-0.0.56.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+evolutionary_policy_optimization-0.0.56.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+evolutionary_policy_optimization-0.0.56.dist-info/RECORD,,

evolutionary_policy_optimization-0.0.54.dist-info/RECORD

@@ -1,9 +0,0 @@
-evolutionary_policy_optimization/__init__.py,sha256=0q0aBuFgWi06MLMD8FiHzBYQ3_W4LYWrwmCtF3u5H2A,201
-evolutionary_policy_optimization/distributed.py,sha256=lSSf_vB04NgVJFBh2n36cGuKZWgOpp8PnPpLDmHT6nU,2296
-evolutionary_policy_optimization/epo.py,sha256=5QJj_l4pihbSdRk1aZnE2dUyWlaqb_VjIKo6Azzksgs,35292
-evolutionary_policy_optimization/experimental.py,sha256=9FrJGviLESlYysHI3i83efT9g2ZB9ha4u3K9HXN98_w,1100
-evolutionary_policy_optimization/mock_env.py,sha256=202KJ5g57wQvOzhGYzgHfBa7Y2do5uuDvl5kFg5o73g,934
-evolutionary_policy_optimization-0.0.54.dist-info/METADATA,sha256=phQq8QaMT7TQQG2Sqz1BW4E1dln1HU10DMExwRvGGkg,6213
-evolutionary_policy_optimization-0.0.54.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-evolutionary_policy_optimization-0.0.54.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-evolutionary_policy_optimization-0.0.54.dist-info/RECORD,,