hippoformer 0.0.3__tar.gz → 0.0.5__tar.gz

{hippoformer-0.0.3 → hippoformer-0.0.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hippoformer
-Version: 0.0.3
+Version: 0.0.5
 Summary: hippoformer
 Project-URL: Homepage, https://pypi.org/project/hippoformer/
 Project-URL: Repository, https://github.com/lucidrains/hippoformer

{hippoformer-0.0.3 → hippoformer-0.0.5}/hippoformer/hippoformer.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
 import torch
-from torch import nn, Tensor, stack, einsum, tensor
+from torch import nn, Tensor, cat, stack, zeros_like, einsum, tensor
 import torch.nn.functional as F
 from torch.nn import Module
 from torch.jit import ScriptModule, script_method
@@ -121,6 +121,7 @@ class mmTEM(Module):
         loss_weight_inference = 1.,
         loss_weight_consistency = 1.,
         loss_weight_relational = 1.,
+        integration_ratio_learned = True
     ):
         super().__init__()
 
@@ -157,6 +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(
@@ -170,7 +179,7 @@ class mmTEM(Module):
         # 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(
+        self.structure_variance_pred_mlp = create_mlp(
             dim = dim_structure * 2,
             dim_in = dim_structure * 2 + 1,
             dim_out = dim_structure,
@@ -185,35 +194,103 @@ class mmTEM(Module):
         self.loss_weight_consistency = loss_weight_consistency
         self.register_buffer('zero', tensor(0.), persistent = False)
 
+        # there is an integration ratio for error correction, but unclear what value this is fixed to or whether it is learned
+
+        self.integration_ratio = nn.Parameter(tensor(0.), requires_grad = integration_ratio_learned)
+
+    def retrieve(
+        self,
+        structural_codes,
+        encoded_sensory
+    ):
+        joint = cat((structural_codes, encoded_sensory), dim = -1)
+
+        queries = self.to_queries(joint)
+
+        retrieved = self.meta_memory_mlp(queries)
+
+        return self.memory_output_decoder(retrieved).split(self.joint_dims, dim = -1)
+
     def forward(
         self,
         sensory,
-        actions
+        actions,
+        return_losses = False
     ):
         structural_codes = self.path_integrator(actions)
 
-        # first have the structure code be able to fetch from the meta memory mlp
-
-        structure_codes_with_zero_sensory = F.pad(structural_codes, (0, self.dim_encoded_sensory))
+        encoded_sensory = self.sensory_encoder(sensory)
 
-        queries = self.to_queries(structure_codes_with_zero_sensory)
-
-        retrieved = self.meta_memory_mlp(queries)
+        # 1. first have the structure code be able to fetch from the meta memory mlp
 
-        decoded_structure, decoded_encoded_sensory = self.memory_output_decoder(retrieved).split(self.joint_dims, dim = -1)
+        decoded_gen_structure, decoded_encoded_sensory = self.retrieve(structural_codes, zeros_like(encoded_sensory))
 
         decoded_sensory = self.sensory_decoder(decoded_encoded_sensory)
 
         generative_pred_loss = F.mse_loss(sensory, decoded_sensory)
 
+        # 2. relational
+
+        # 2a. structure from content
+
+        decoded_structure, decoded_encoded_sensory = self.retrieve(zeros_like(structural_codes), encoded_sensory)
+
+        structure_from_content_loss = F.mse_loss(decoded_structure, structural_codes)
+
+        # 2b. structure from structure
+
+        decoded_structure, decoded_encoded_sensory = self.retrieve(zeros_like(structural_codes), encoded_sensory)
+
+        structure_from_structure_loss = F.mse_loss(decoded_structure, structural_codes)
+
+        relational_loss = structure_from_content_loss + structure_from_structure_loss
+
+        # 3. consistency - modeling a feedback system from hippocampus to path integration
+
+        corrected_structural_code, corrected_encoded_sensory = self.retrieve(decoded_gen_structure, encoded_sensory)
+
+        sensory_sse = (corrected_encoded_sensory - encoded_sensory).norm(dim = -1, keepdim = True).pow(2)
+
+        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.sigmoid() * pred_variance
+
+        consistency_loss = F.mse_loss(decoded_gen_structure, inf_structural_code)
+
+        # 4. final inference loss
+
+        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 = (
-            generative_pred_loss * self.loss_weight_generative
+            generative_pred_loss * self.loss_weight_generative +
+            relational_loss * self.loss_weight_relational +
+            consistency_loss * self.loss_weight_consistency +
+            inference_pred_loss * self.loss_weight_inference
         )
 
         losses = (
             generative_pred_loss,
+            relational_loss,
+            consistency_loss,
+            inference_pred_loss
         )
 
+        if not return_losses:
+            return total_loss
+
         return total_loss, losses

{hippoformer-0.0.3 → hippoformer-0.0.5}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hippoformer"
-version = "0.0.3"
+version = "0.0.5"
 description = "hippoformer"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{hippoformer-0.0.3 → 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()