hippoformer 0.0.1__tar.gz → 0.0.2__tar.gz

{hippoformer-0.0.1 → hippoformer-0.0.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hippoformer
-Version: 0.0.1
+Version: 0.0.2
 Summary: hippoformer
 Project-URL: Homepage, https://pypi.org/project/hippoformer/
 Project-URL: Repository, https://github.com/lucidrains/hippoformer

hippoformer-0.0.2/hippoformer/__init__.py

@@ -0,0 +1,4 @@
+from hippoformer.hippoformer import (
+    PathIntegration,
+    mmTEM
+)

{hippoformer-0.0.1 → hippoformer-0.0.2}/hippoformer/hippoformer.py

@@ -104,13 +104,73 @@ class PathIntegration(Module):
 class mmTEM(Module):
     def __init__(
         self,
-        dim
+        dim,
+        *,
+        sensory_encoder: Module,
+        sensory_decoder: Module,
+        dim_sensory,
+        dim_action,
+        dim_encoded_sensory,
+        dim_structure,
+        meta_mlp_depth = 2,
+        decoder_mlp_depth = 2,
+        structure_variance_pred_mlp_depth = 2,
+        path_integrate_kwargs: dict = dict(),
+        loss_weight_generative = 1.,
+        loss_weight_inference = 1.,
+        loss_weight_consistency = 1.,
+        loss_weight_relational = 1.,
     ):
         super().__init__()
 
+        dim_joint_rep = dim_encoded_sensory + dim_structure
+
+        # path integrator
+
+        self.path_integrator = PathIntegration(
+            dim_action = dim_action,
+            dim_structure = dim_structure,
+            **path_integrate_kwargs
+        )
+
+        # meta mlp related
+
+        self.to_queries = nn.Linear(dim_joint_rep, dim, bias = False)
+        self.to_keys = nn.Linear(dim_joint_rep, dim, bias = False)
+        self.to_values = nn.Linear(dim_joint_rep, dim, bias = False)
+
+        self.meta_mlp = create_mlp(
+            dim = dim * 2,
+            depth = meta_mlp_depth,
+            dim_in = dim,
+            dim_out = dim,
+            activation = nn.ReLU()
+        )
+
+        # mlp decoder (from meta mlp output to joint)
+
+        self.meta_mlp_output_decoder = create_mlp(
+            dim = dim * 2,
+            dim_in = dim,
+            dim_out = dim_joint_rep,
+            depth = decoder_mlp_depth,
+            activation = nn.ReLU()
+        )
+
+        # the mlp that predicts the variance for the structural code
+        # for correcting the generated structural code modeling the feedback from HC to MEC
+
+        self.structure_variance_pred_mlp_depth = create_mlp(
+            dim = dim_structure * 2,
+            dim_in = dim_structure * 2 + 1,
+            dim_out = dim_structure,
+            depth = structure_variance_pred_mlp_depth
+        )
 
     def forward(
         self,
-        data
+        sensory,
+        actions
     ):
-        raise NotImplementedError
+        structural_codes = self.path_integrator(actions)
+        return structural_codes.sum()

{hippoformer-0.0.1 → hippoformer-0.0.2}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hippoformer"
-version = "0.0.1"
+version = "0.0.2"
 description = "hippoformer"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

hippoformer-0.0.2/tests/test_hippoformer.py

@@ -0,0 +1,37 @@
+import pytest
+
+import torch
+
+def test_path_integrate():
+    from hippoformer.hippoformer import PathIntegration
+
+    path_integrator = PathIntegration(32, 64)
+
+    actions = torch.randn(2, 16, 32)
+
+    structure_codes = path_integrator(actions)
+    structure_codes = path_integrator(actions, structure_codes) # pass in previous structure codes, it will auto use the last one as hidden and pass it to the RNN
+
+    assert structure_codes.shape == (2, 16, 64)
+
+def test_mm_tem():
+    import torch
+    from hippoformer.hippoformer import mmTEM
+
+    from torch.nn import Linear
+
+    model = mmTEM(
+        dim = 32,
+        sensory_encoder = Linear(11, 32),
+        sensory_decoder = Linear(32, 11),
+        dim_sensory = 11,
+        dim_action = 7,
+        dim_structure = 32,
+        dim_encoded_sensory = 32
+    )
+
+    actions = torch.randn(2, 16, 7)
+    sensory = torch.randn(2, 16, 11)
+
+    loss = model(sensory, actions)
+    loss.backward()

hippoformer-0.0.1/tests/test_hippoformer.py

@@ -1,15 +0,0 @@
-import pytest
-
-import torch
-
-def test_path_integrate():
-    from hippoformer.hippoformer import PathIntegration
-
-    path_integrator = PathIntegration(32, 64)
-
-    actions = torch.randn(2, 16, 32)
-
-    structure_codes = path_integrator(actions)
-    structure_codes = path_integrator(actions, structure_codes) # pass in previous structure codes, it will auto use the last one as hidden and pass it to the RNN
-
-    assert structure_codes.shape == (2, 16, 64)