evolutionary-policy-optimization 0.0.63__py3-none-any.whl → 0.0.64__py3-none-any.whl

evolutionary_policy_optimization/epo.py

@@ -1068,7 +1068,8 @@ class EPO(Module):
         agent: Agent,
         episodes_per_latent,
         max_episode_length,
-        action_sample_temperature = 1.
+        action_sample_temperature = 1.,
+        fix_environ_across_latents = True
     ):
         super().__init__()
         self.agent = agent
@@ -1077,6 +1078,7 @@ class EPO(Module):
         self.num_latents = agent.latent_gene_pool.num_latents
         self.episodes_per_latent = episodes_per_latent
         self.max_episode_length = max_episode_length
+        self.fix_environ_across_latents = fix_environ_across_latents
 
         self.register_buffer('dummy', tensor(0))
 
@@ -1133,9 +1135,11 @@ class EPO(Module):
     def forward(
         self,
         env,
-        fix_environ_across_latents = True
+        fix_environ_across_latents = None
     ) -> MemoriesAndCumulativeRewards:
 
+        fix_environ_across_latents = default(fix_environ_across_latents, self.fix_environ_across_latents)
+
         self.agent.eval()
 
         invalid_episode = tensor(-1) # will use `episode_id` value of `-1` for the `next_value`, needed for not discarding last reward for generalized advantage estimate
@@ -1202,16 +1206,18 @@ class EPO(Module):
 
                 time += 1
 
-            # need the final next value for GAE, iiuc
+            if not done:
+                # add bootstrap value if truncated
 
-            next_value = temp_batch_dim(self.agent.get_critic_values)(state, latent = latent)
+                next_value = temp_batch_dim(self.agent.get_critic_values)(state, latent = latent)
 
-            memory_for_gae = memory._replace(
-                episode_id = invalid_episode,
-                value = next_value
-            )
+                memory_for_gae = memory._replace(
+                    episode_id = invalid_episode,
+                    value = next_value,
+                    done = tensor(True)
+                )
 
-            memories.append(memory_for_gae)
+                memories.append(memory_for_gae)
 
         return MemoriesAndCumulativeRewards(
             memories = memories,

{evolutionary_policy_optimization-0.0.63.dist-info → evolutionary_policy_optimization-0.0.64.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: evolutionary-policy-optimization
-Version: 0.0.63
+Version: 0.0.64
 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-0.0.63.dist-info → evolutionary_policy_optimization-0.0.64.dist-info}/RECORD

@@ -1,9 +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=DSG2fYWLk0cyHhfoiwqmSzh2TBOWhz25sD1oWIM5p1k,36695
+evolutionary_policy_optimization/epo.py,sha256=0_jC9Tbl6FiscLHklvTKtuQTwZL8egqFKW-4JUxxwvw,37001
 evolutionary_policy_optimization/experimental.py,sha256=-IgqjJ_Wk_CMB1y9YYWpoYqTG9GZHAS6kbRdTluVevg,1563
 evolutionary_policy_optimization/mock_env.py,sha256=gvATGA51Ym5sf3jiR2VmlpjiCcT7KCDDY_SrR-MEwsU,941
-evolutionary_policy_optimization-0.0.63.dist-info/METADATA,sha256=X2FKT8WJ9T1t0ydEdtxrJsJGXY1ubfvydQSykv2G03M,6220
-evolutionary_policy_optimization-0.0.63.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-evolutionary_policy_optimization-0.0.63.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-evolutionary_policy_optimization-0.0.63.dist-info/RECORD,,
+evolutionary_policy_optimization-0.0.64.dist-info/METADATA,sha256=vWdnTe2a86wTenEh29TNJlYEjD8A5CPtsyylxh4XsE0,6220
+evolutionary_policy_optimization-0.0.64.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+evolutionary_policy_optimization-0.0.64.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+evolutionary_policy_optimization-0.0.64.dist-info/RECORD,,