hippoformer 0.0.1__py3-none-any.whl → 0.0.3__py3-none-any.whl

hippoformer/__init__.py

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

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, stack, 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
@@ -104,13 +105,115 @@ 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__()
 
+        # 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(
+            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_memory_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.memory_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
+        )
+
+        # 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)
 
     def forward(
         self,
-        data
+        sensory,
+        actions
     ):
-        raise NotImplementedError
+        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))
+
+        queries = self.to_queries(structure_codes_with_zero_sensory)
+
+        retrieved = self.meta_memory_mlp(queries)
+
+        decoded_structure, decoded_encoded_sensory = self.memory_output_decoder(retrieved).split(self.joint_dims, dim = -1)
+
+        decoded_sensory = self.sensory_decoder(decoded_encoded_sensory)
+
+        generative_pred_loss = F.mse_loss(sensory, decoded_sensory)
+
+        # losses
+
+        total_loss = (
+            generative_pred_loss * self.loss_weight_generative
+        )
+
+        losses = (
+            generative_pred_loss,
+        )
+
+        return total_loss, losses

{hippoformer-0.0.1.dist-info → hippoformer-0.0.3.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hippoformer
-Version: 0.0.1
+Version: 0.0.3
 Summary: hippoformer
 Project-URL: Homepage, https://pypi.org/project/hippoformer/
 Project-URL: Repository, https://github.com/lucidrains/hippoformer

hippoformer-0.0.3.dist-info/RECORD

@@ -0,0 +1,6 @@
+hippoformer/__init__.py,sha256=A7N8GsRAZH4yP-L5hb7IVDnNjnhfjNyolg5MZ6vnGyE,71
+hippoformer/hippoformer.py,sha256=XnLDm12BHuJSOcb1Y8EKCSqkvdw20anuGT2O4kFsWHg,6041
+hippoformer-0.0.3.dist-info/METADATA,sha256=L1SJ76ffExU4DQen0ppCv3gzt7waXXaE6znPuon60_g,2773
+hippoformer-0.0.3.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+hippoformer-0.0.3.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+hippoformer-0.0.3.dist-info/RECORD,,

hippoformer-0.0.1.dist-info/RECORD

@@ -1,6 +0,0 @@
-hippoformer/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-hippoformer/hippoformer.py,sha256=6tA4ZWYKbzclpeTUhJtr2OguVOyyAGFxuLf9bfnfO_M,2682
-hippoformer-0.0.1.dist-info/METADATA,sha256=4hnfh1oIIlcGsIQ7qD7fZHWfM5ltnHhATAPcN-4vkxQ,2773
-hippoformer-0.0.1.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-hippoformer-0.0.1.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-hippoformer-0.0.1.dist-info/RECORD,,