evolutionary-policy-optimization 0.1.9__py3-none-any.whl → 0.1.12__py3-none-any.whl

evolutionary_policy_optimization/epo.py

@@ -146,6 +146,24 @@ def temp_batch_dim(fn):
 
     return inner
 
+# plasticity related
+
+def shrink_and_perturb_(
+    module,
+    shrink_factor = 0.5,
+    perturb_factor = 0.01
+):
+    # Shrink & Perturb
+    # Ash et al. https://arxiv.org/abs/1910.08475
+
+    assert 0. <= shrink_factor <= 1.
+
+    for p in module.parameters():
+        noise = torch.randn_like(p.data)
+        p.data.mul_(1. - shrink_factor).add_(noise * perturb_factor)
+
+    return module
+
 # fitness related
 
 def get_fitness_scores(
@@ -431,6 +449,8 @@ class Critic(Module):
 
         value = self.maybe_bins_to_value(logits)
 
+        loss_fn = partial(self.loss_fn, reduction = 'none')
+
         if use_improved:
             clipped_target = target.clamp(-eps_clip, eps_clip)
 
@@ -439,8 +459,8 @@ class Critic(Module):
 
             is_between = lambda lo, hi: (lo < value) & (value < hi)
 
-            clipped_loss = self.loss_fn(logits, clipped_target, reduction = 'none')
-            loss = self.loss_fn(logits, target, reduction = 'none')
+            clipped_loss = loss_fn(logits, clipped_target)
+            loss = loss_fn(logits, target)
 
             value_loss = torch.where(
                 is_between(target, old_values_lo) | is_between(old_values_hi, target),
@@ -448,10 +468,10 @@ class Critic(Module):
                 torch.min(loss, clipped_loss)
             )
         else:
-            clipped_value = old_values + (value - old_values).clamp(1. - eps_clip, 1. + eps_clip)
+            clipped_value = old_values + (value - old_values).clamp(-eps_clip, eps_clip)
 
-            loss = self.loss_fn(logits, target, reduction = 'none')
-            clipped_loss = self.loss_fn(clipped_value, target, reduction = 'none')
+            loss = loss_fn(logits, target)
+            clipped_loss = loss_fn(clipped_value, target)
 
             value_loss = torch.max(loss, clipped_loss)
 

{evolutionary_policy_optimization-0.1.9.dist-info → evolutionary_policy_optimization-0.1.12.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.1.9
+Version: 0.1.12
 Summary: EPO - Pytorch
 Project-URL: Homepage, https://pypi.org/project/evolutionary-policy-optimization/
 Project-URL: Repository, https://github.com/lucidrains/evolutionary-policy-optimization
@@ -215,4 +215,15 @@ agent.load('./agent.pt')
 }
 ```
 
+```bibtex
+@article{Ash2019OnTD,
+    title   = {On the Difficulty of Warm-Starting Neural Network Training},
+    author  = {Jordan T. Ash and Ryan P. Adams},
+    journal = {ArXiv},
+    year    = {2019},
+    volume  = {abs/1910.08475},
+    url     = {https://api.semanticscholar.org/CorpusID:204788802}
+}
+```
+
 *Evolution is cleverer than you are.* - Leslie Orgel

{evolutionary_policy_optimization-0.1.9.dist-info → evolutionary_policy_optimization-0.1.12.dist-info}/RECORD

@@ -1,10 +1,10 @@
 evolutionary_policy_optimization/__init__.py,sha256=NyiYDYU7DlpmOTM7xiBQET3r1WwX0ebrgMCBLSQrW3c,288
 evolutionary_policy_optimization/distributed.py,sha256=7KgZdeS_wxBHo_du9XZFB1Cu318J-Bp66Xdr6Log_20,2423
 evolutionary_policy_optimization/env_wrappers.py,sha256=bDL06o9_b1iW6k3fw2xifnOnYlzs643tdW6Yv2gsIdw,803
-evolutionary_policy_optimization/epo.py,sha256=9GfSvOz6SwjAuZyhyvsLHPY8b2svMQlM3BRjilwsQ-g,45717
+evolutionary_policy_optimization/epo.py,sha256=aOltJBkZVi2FxXao51zdfYaLynIi8T8v3qO1ex2HElg,46058
 evolutionary_policy_optimization/experimental.py,sha256=-IgqjJ_Wk_CMB1y9YYWpoYqTG9GZHAS6kbRdTluVevg,1563
 evolutionary_policy_optimization/mock_env.py,sha256=TLyyRm6tOD0Kdn9QqJJQriaSnsR-YmNQHo4OohmZFG4,1410
-evolutionary_policy_optimization-0.1.9.dist-info/METADATA,sha256=y5w_NwtKNQ07HeYa5r6hcPn7RsqDpehMmt5vj6mTESQ,7316
-evolutionary_policy_optimization-0.1.9.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-evolutionary_policy_optimization-0.1.9.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-evolutionary_policy_optimization-0.1.9.dist-info/RECORD,,
+evolutionary_policy_optimization-0.1.12.dist-info/METADATA,sha256=B_eK4c8-dp4rr4K0HxRiNJqY6fK10XmwBbsPm-PE0_k,7625
+evolutionary_policy_optimization-0.1.12.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+evolutionary_policy_optimization-0.1.12.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+evolutionary_policy_optimization-0.1.12.dist-info/RECORD,,