hippoformer 0.0.2__tar.gz → 0.0.4__tar.gz

{hippoformer-0.0.2 → hippoformer-0.0.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hippoformer
-Version: 0.0.2
+Version: 0.0.4
 Summary: hippoformer
 Project-URL: Homepage, https://pypi.org/project/hippoformer/
 Project-URL: Repository, https://github.com/lucidrains/hippoformer

{hippoformer-0.0.2 → hippoformer-0.0.4}/hippoformer/hippoformer.py

@@ -1,10 +1,11 @@
 from __future__ import annotations
 
 import torch
-from torch import nn, Tensor, stack, einsum
+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
+from torch.func import vmap, grad, functional_call
 
 from einops import repeat, rearrange
 from einops.layers.torch import Rearrange
@@ -120,11 +121,21 @@ class mmTEM(Module):
         loss_weight_inference = 1.,
         loss_weight_consistency = 1.,
         loss_weight_relational = 1.,
+        integration_ratio_learned = True
     ):
         super().__init__()
 
+        # sensory
+
+        self.sensory_encoder = sensory_encoder
+        self.sensory_decoder = sensory_decoder
+
         dim_joint_rep = dim_encoded_sensory + dim_structure
 
+        self.dim_encoded_sensory = dim_encoded_sensory
+        self.dim_structure = dim_structure
+        self.joint_dims = (dim_structure, dim_encoded_sensory)
+
         # path integrator
 
         self.path_integrator = PathIntegration(
@@ -139,7 +150,7 @@ class mmTEM(Module):
         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(
+        self.meta_memory_mlp = create_mlp(
             dim = dim * 2,
             depth = meta_mlp_depth,
             dim_in = dim,
@@ -149,7 +160,7 @@ class mmTEM(Module):
 
         # mlp decoder (from meta mlp output to joint)
 
-        self.meta_mlp_output_decoder = create_mlp(
+        self.memory_output_decoder = create_mlp(
             dim = dim * 2,
             dim_in = dim,
             dim_out = dim_joint_rep,
@@ -160,17 +171,105 @@ 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,
             depth = structure_variance_pred_mlp_depth
         )
 
+        # loss related
+
+        self.loss_weight_generative = loss_weight_generative
+        self.loss_weight_inference = loss_weight_inference
+        self.loss_weight_relational = loss_weight_relational
+        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
     ):
         structural_codes = self.path_integrator(actions)
-        return structural_codes.sum()
+
+        encoded_sensory = self.sensory_encoder(sensory)
+
+        # 1. first have the structure code be able to fetch from the meta memory mlp
+
+        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 * 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))
+
+        decoded_inf_sensory = self.sensory_decoder(inf_encoded_sensory)
+
+        inference_pred_loss = F.mse_loss(sensory, decoded_inf_sensory)
+
+        # losses
+
+        total_loss = (
+            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
+        )
+
+        return total_loss, losses

{hippoformer-0.0.2 → hippoformer-0.0.4}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hippoformer"
-version = "0.0.2"
+version = "0.0.4"
 description = "hippoformer"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }

{hippoformer-0.0.2 → hippoformer-0.0.4}/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 = model(sensory, actions)
+    loss, losses = model(sensory, actions)
     loss.backward()