hippoformer 0.0.4__tar.gz → 0.0.6__tar.gz

{hippoformer-0.0.4 → hippoformer-0.0.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hippoformer
-Version: 0.0.4
+Version: 0.0.6
 Summary: hippoformer
 Project-URL: Homepage, https://pypi.org/project/hippoformer/
 Project-URL: Repository, https://github.com/lucidrains/hippoformer
@@ -36,6 +36,7 @@ Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.9
 Requires-Dist: assoc-scan
 Requires-Dist: einops>=0.8.1
+Requires-Dist: einx>=0.3.0
 Requires-Dist: torch>=2.4
 Requires-Dist: x-mlps-pytorch
 Provides-Extra: examples

{hippoformer-0.0.4 → hippoformer-0.0.6}/hippoformer/hippoformer.py

@@ -7,7 +7,8 @@ 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 einx import multiply
+from einops import repeat, rearrange, pack, unpack
 from einops.layers.torch import Rearrange
 
 from x_mlps_pytorch import create_mlp
@@ -22,6 +23,16 @@ def exists(v):
 def default(v, d):
     return v if exists(v) else d
 
+def pack_with_inverse(t, pattern):
+    packed, packed_shape = pack([t], pattern)
+
+    def inverse(out, inv_pattern = None):
+        inv_pattern = default(inv_pattern, pattern)
+        unpacked, = unpack(out, packed_shape, inv_pattern)
+        return unpacked
+
+    return packed, inverse
+
 def l2norm(t):
     return F.normalize(t, dim = -1)
 
@@ -121,7 +132,8 @@ class mmTEM(Module):
         loss_weight_inference = 1.,
         loss_weight_consistency = 1.,
         loss_weight_relational = 1.,
-        integration_ratio_learned = True
+        integration_ratio_learned = True,
+        assoc_scan_kwargs: dict = dict()
     ):
         super().__init__()
 
@@ -150,6 +162,9 @@ 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.to_learned_optim_hparams = nn.Linear(dim_joint_rep, 3, bias = False) # for learning rate, forget gate, and momentum
+        self.assoc_scan = AssocScan(*assoc_scan_kwargs)
+
         self.meta_memory_mlp = create_mlp(
             dim = dim * 2,
             depth = meta_mlp_depth,
@@ -158,6 +173,14 @@ 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(
@@ -206,8 +229,11 @@ class mmTEM(Module):
     def forward(
         self,
         sensory,
-        actions
+        actions,
+        return_losses = False
     ):
+        batch = actions.shape[0]
+
         structural_codes = self.path_integrator(actions)
 
         encoded_sensory = self.sensory_encoder(sensory)
@@ -244,18 +270,63 @@ 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)
+
+        lr, forget, beta = self.to_learned_optim_hparams(joint_code_to_store).unbind(dim = -1)
+
+        params = dict(self.meta_memory_mlp.named_parameters())
+        grads = self.per_sample_grad_fn(params, keys, values)
+
+        # update the meta mlp parameters
+
+        init_momentums = {k: zeros_like(v) for k, v in params.items()}
+        next_params = dict()
+
+        for (
+            (key, param),
+            (_, grad),
+            (_, init_momentum)
+        ) in zip(
+            params.items(),
+            grads.items(),
+            init_momentums.items()
+        ):
+
+            grad, inverse_pack = pack_with_inverse(grad, 'b t *')
+
+            grad = multiply('b t ..., b t', grad, lr)
+
+            expanded_beta = repeat(beta, 'b t -> b t w', w = grad.shape[-1])
+
+            init_momentum = repeat(init_momentum, '... -> b ...', b = batch)
+
+            update = self.assoc_scan(grad, expanded_beta.sigmoid(), init_momentum)
+
+            expanded_forget = repeat(forget, 'b t -> b t w', w = grad.shape[-1])
+
+            acc_update = self.assoc_scan(update, expanded_forget.sigmoid())
+
+            acc_update = inverse_pack(acc_update)
+
+            next_params[key] = param - acc_update[:, -1]
+
         # losses
 
         total_loss = (
@@ -272,4 +343,7 @@ class mmTEM(Module):
             inference_pred_loss
         )
 
+        if not return_losses:
+            return total_loss
+
         return total_loss, losses

{hippoformer-0.0.4 → hippoformer-0.0.6}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "hippoformer"
-version = "0.0.4"
+version = "0.0.6"
 description = "hippoformer"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -25,6 +25,7 @@ classifiers=[
 
 dependencies = [
     "assoc-scan",
+    "einx>=0.3.0",
     "einops>=0.8.1",
     "torch>=2.4",
     "x-mlps-pytorch",

{hippoformer-0.0.4 → hippoformer-0.0.6}/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()