hippoformer 0.0.4__tar.gz → 0.0.5__tar.gz

{hippoformer-0.0.4 → hippoformer-0.0.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hippoformer
-Version: 0.0.4
+Version: 0.0.5
 Summary: hippoformer
 Project-URL: Homepage, https://pypi.org/project/hippoformer/
 Project-URL: Repository, https://github.com/lucidrains/hippoformer

{hippoformer-0.0.4 → hippoformer-0.0.5}/hippoformer/hippoformer.py

@@ -158,6 +158,14 @@ class mmTEM(Module):
             activation = nn.ReLU()
         )
 
+        def forward_with_mse_loss(params, keys, values):
+            pred = functional_call(self.meta_memory_mlp, params, keys)
+            return F.mse_loss(pred, values)
+
+        grad_fn = grad(forward_with_mse_loss)
+
+        self.per_sample_grad_fn = vmap(vmap(grad_fn, in_dims = (None, 0, 0)), in_dims = (None, 0, 0))
+
         # mlp decoder (from meta mlp output to joint)
 
         self.memory_output_decoder = create_mlp(
@@ -206,7 +214,8 @@ class mmTEM(Module):
     def forward(
         self,
         sensory,
-        actions
+        actions,
+        return_losses = False
     ):
         structural_codes = self.path_integrator(actions)
 
@@ -244,18 +253,27 @@ class mmTEM(Module):
 
         pred_variance = self.structure_variance_pred_mlp(cat((corrected_structural_code, decoded_gen_structure, sensory_sse), dim = -1))
 
-        inf_structural_code = decoded_gen_structure + (corrected_structural_code - decoded_gen_structure) * self.integration_ratio * pred_variance
+        inf_structural_code = decoded_gen_structure + (corrected_structural_code - decoded_gen_structure) * self.integration_ratio.sigmoid() * pred_variance
 
         consistency_loss = F.mse_loss(decoded_gen_structure, inf_structural_code)
 
         # 4. final inference loss
 
-        _, inf_encoded_sensory = self.retrieve(inf_structural_code, zeros_like(encoded_sensory))
+        final_structural_code, inf_encoded_sensory = self.retrieve(inf_structural_code, zeros_like(encoded_sensory))
 
         decoded_inf_sensory = self.sensory_decoder(inf_encoded_sensory)
 
         inference_pred_loss = F.mse_loss(sensory, decoded_inf_sensory)
 
+        # 5. store the final structural code from step 4 + encoded sensory
+
+        joint_code_to_store = cat((final_structural_code, encoded_sensory), dim = -1)
+
+        keys = self.to_keys(joint_code_to_store)
+        values = self.to_values(joint_code_to_store)
+
+        grads = self.per_sample_grad_fn(dict(self.meta_memory_mlp.named_parameters()), keys, values)
+
         # losses
 
         total_loss = (
@@ -272,4 +290,7 @@ class mmTEM(Module):
             inference_pred_loss
         )
 
+        if not return_losses:
+            return total_loss
+
         return total_loss, losses

{hippoformer-0.0.4 → hippoformer-0.0.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hippoformer"
-version = "0.0.4"
+version = "0.0.5"
 description = "hippoformer"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{hippoformer-0.0.4 → hippoformer-0.0.5}/tests/test_hippoformer.py

@@ -33,5 +33,5 @@ def test_mm_tem():
     actions = torch.randn(2, 16, 7)
     sensory = torch.randn(2, 16, 11)
 
-    loss, losses = model(sensory, actions)
+    loss = model(sensory, actions)
     loss.backward()