metacontroller-pytorch 0.0.36__tar.gz → 0.0.37__tar.gz

{metacontroller_pytorch-0.0.36 → metacontroller_pytorch-0.0.37}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metacontroller-pytorch
-Version: 0.0.36
+Version: 0.0.37
 Summary: Transformer Metacontroller
 Project-URL: Homepage, https://pypi.org/project/metacontroller/
 Project-URL: Repository, https://github.com/lucidrains/metacontroller

{metacontroller_pytorch-0.0.36 → metacontroller_pytorch-0.0.37}/metacontroller/metacontroller.py

@@ -408,6 +408,7 @@ class Transformer(Module):
         meta_controller: Module | None = None,
         cache: TransformerOutput | None = None,
         discovery_phase = False,
+        force_behavior_cloning = False,
         meta_controller_temperature = 1.,
         return_raw_action_dist = False,
         return_latents = False,
@@ -420,11 +421,15 @@ class Transformer(Module):
 
         meta_controller = default(meta_controller, self.meta_controller)
 
+        if force_behavior_cloning:
+            assert not discovery_phase, 'discovery phase cannot be set to True if force behavioral cloning is set to True'
+            meta_controller = None
+
         has_meta_controller = exists(meta_controller)
 
         assert not (discovery_phase and not has_meta_controller), 'meta controller must be made available during discovery phase'
 
-        behavioral_cloning = not has_meta_controller and not return_raw_action_dist
+        behavioral_cloning = force_behavior_cloning or (not has_meta_controller and not return_raw_action_dist)
 
         # by default, if meta controller is passed in, transformer is no grad
 
@@ -472,7 +477,7 @@ class Transformer(Module):
 
         with meta_controller_context():
 
-            if exists(meta_controller):
+            if exists(meta_controller) and not behavioral_cloning:
                 control_signal, next_meta_hiddens = meta_controller(residual_stream, cache = meta_hiddens, discovery_phase = discovery_phase, temperature = meta_controller_temperature, episode_lens = episode_lens)
             else:
                 control_signal, next_meta_hiddens = self.zero, None

{metacontroller_pytorch-0.0.36 → metacontroller_pytorch-0.0.37}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "metacontroller-pytorch"
-version = "0.0.36"
+version = "0.0.37"
 description = "Transformer Metacontroller"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{metacontroller_pytorch-0.0.36 → metacontroller_pytorch-0.0.37}/train_behavior_clone_babyai.py

@@ -130,8 +130,8 @@ def train(
     for epoch in range(cloning_epochs + discovery_epochs):
 
         model.train()
-        total_state_loss = 0.
-        total_action_loss = 0.
+        from collections import defaultdict
+        total_losses = defaultdict(float)
 
         progress_bar = tqdm(dataloader, desc = f"Epoch {epoch}", disable = not accelerator.is_local_main_process)
 
@@ -200,9 +200,9 @@ def train(
                 optim.zero_grad()
 
             # log
-
-            total_state_loss += state_loss.item()
-            total_action_loss += action_loss.item()
+            
+            for key, value in log.items():
+                total_losses[key] += value
 
             accelerator.log({
                 **log,
@@ -210,15 +210,11 @@ def train(
                 "grad_norm": grad_norm.item()
             })
 
-            progress_bar.set_postfix(
-                state_loss = state_loss.item(),
-                action_loss = action_loss.item()
-            )
-
-        avg_state_loss = total_state_loss / len(dataloader)
-        avg_action_loss = total_action_loss / len(dataloader)
+            progress_bar.set_postfix(**log)
 
-        accelerator.print(f"Epoch {epoch}: state_loss={avg_state_loss:.4f}, action_loss={avg_action_loss:.4f}")
+        avg_losses = {k: v / len(dataloader) for k, v in total_losses.items()}
+        avg_losses_str = ", ".join([f"{k}={v:.4f}" for k, v in avg_losses.items()])
+        accelerator.print(f"Epoch {epoch}: {avg_losses_str}")
 
     # save weights
 
@@ -231,7 +227,7 @@ def train(
         unwrapped_meta_controller = accelerator.unwrap_model(meta_controller)
         unwrapped_meta_controller.save(meta_controller_checkpoint_path)
 
-        accelerator.print(f"Model saved to {checkpoint_path}, MetaControler to {meta_controller_checkpoint_path}")
+        accelerator.print(f"Model saved to {checkpoint_path}, MetaController to {meta_controller_checkpoint_path}")
 
     accelerator.end_training()