added save and load for optimizers

dreamer.py

@@ -319,8 +319,10 @@ def main(config):
         train_dataset,
     ).to(config.device)
     agent.requires_grad_(requires_grad=False)
-    if (logdir / "latest_model.pt").exists():
+    if (logdir / "latest.pt").exists():
-        agent.load_state_dict(torch.load(logdir / "latest_model.pt"))
+        checkpoint = torch.load(logdir / "latest.pt")
+        agent.load_state_dict(checkpoint["agent_state_dict"])
+        tools.recursively_load_optim_state_dict(agent, checkpoint["optims_state_dict"])
         agent._should_pretrain._once = False
 
     # make sure eval will be executed once after config.steps
@@ -352,7 +354,11 @@ def main(config):
             steps=config.eval_every,
             state=state,
         )
-        torch.save(agent.state_dict(), logdir / "latest_model.pt")
+        items_to_save = {
+            "agent_state_dict": agent.state_dict(),
+            "optims_state_dict": tools.recursively_collect_optim_state_dict(agent),
+        }
+        torch.save(items_to_save, logdir / "latest.pt")
     for env in train_envs + eval_envs:
         try:
             env.close()

exploration.py

@@ -70,7 +70,7 @@ class Plan2Explore(nn.Module):
             [networks.MLP(**kw) for _ in range(config.disag_models)]
         )
         kw = dict(wd=config.weight_decay, opt=config.opt, use_amp=self._use_amp)
-        self._model_opt = tools.Optimizer(
+        self._expl_opt = tools.Optimizer(
             "explorer",
             self.parameters(),
             config.model_lr,
@@ -129,5 +129,5 @@ class Plan2Explore(nn.Module):
                 [torch.mean(pred.log_prob(targets))[None] for pred in preds], 0
             )
             loss = -torch.mean(likes)
-        metrics = self._model_opt(loss, self.parameters())
+        metrics = self._expl_opt(loss, self.parameters())
         return metrics

tools.py

@@ -970,3 +970,35 @@ def enable_deterministic_run():
     os.environ["CUBLAS_WORKSPACE_CONFIG"] = ":4096:8"
     torch.backends.cudnn.benchmark = False
     torch.use_deterministic_algorithms(True)
+
+
+def recursively_collect_optim_state_dict(obj, path="", optimizers_state_dicts=None):
+    if optimizers_state_dicts is None:
+        optimizers_state_dicts = {}
+    attrs = obj.__dict__
+    if isinstance(obj, torch.nn.Module):
+        attrs.update(
+            {k: attr for k, attr in obj.named_modules() if "." not in k and obj != attr}
+        )
+    for name, attr in attrs.items():
+        new_path = path + "." + name if path else name
+        if isinstance(attr, torch.optim.Optimizer):
+            optimizers_state_dicts[new_path] = attr.state_dict()
+        elif hasattr(attr, "__dict__"):
+            optimizers_state_dicts.update(
+                recursively_collect_optim_state_dict(
+                    attr, new_path, optimizers_state_dicts
+                )
+            )
+    return optimizers_state_dicts
+
+
+def recursively_load_optim_state_dict(obj, optimizers_state_dicts):
+    print(optimizers_state_dicts)
+    for path, state_dict in optimizers_state_dicts.items():
+        keys = path.split(".")
+        obj_now = obj
+        for key in keys:
+            obj_now = getattr(obj_now, key)
+        print(keys)
+        obj_now.load_state_dict(state_dict)