hippoformer 0.0.5__py3-none-any.whl → 0.0.7__py3-none-any.whl

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,
@@ -164,7 +179,7 @@ class mmTEM(Module):
 
         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))
+        self.per_sample_grad_fn = vmap(vmap(grad_fn, in_dims = (None, 0, 0)), in_dims = (0, 0, 0))
 
         # mlp decoder (from meta mlp output to joint)
 
@@ -198,6 +213,19 @@ class mmTEM(Module):
 
         self.integration_ratio = nn.Parameter(tensor(0.), requires_grad = integration_ratio_learned)
 
+    def init_params_and_momentum(
+        self,
+        batch_size
+    ):
+
+        params_dict = dict(self.meta_memory_mlp.named_parameters())
+
+        params = {name: repeat(param, '... -> b ...', b = batch_size) for name, param in params_dict.items()}
+
+        momentums = {name: zeros_like(param) for name, param in params.items()}
+
+        return params, momentums
+
     def retrieve(
         self,
         structural_codes,
@@ -215,8 +243,12 @@ class mmTEM(Module):
         self,
         sensory,
         actions,
-        return_losses = False
+        memory_mlp_params = None,
+        return_losses = False,
+        return_memory_mlp_params = False
     ):
+        batch = actions.shape[0]
+
         structural_codes = self.path_integrator(actions)
 
         encoded_sensory = self.sensory_encoder(sensory)
@@ -272,7 +304,50 @@ class mmTEM(Module):
         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)
+        lr, forget, beta = self.to_learned_optim_hparams(joint_code_to_store).unbind(dim = -1)
+
+        if exists(memory_mlp_params):
+            params, momentums = memory_mlp_params
+        else:
+            params, momentums = self.init_params_and_momentum(batch)
+
+        # store by getting gradients of mse loss of keys and values
+
+        grads = self.per_sample_grad_fn(params, keys, values)
+
+        # update the meta mlp parameters and momentums
+
+        next_params = dict()
+        next_momentum = dict()
+
+        for (
+            (key, param),
+            (_, grad),
+            (_, momentum)
+        ) in zip(
+            params.items(),
+            grads.items(),
+            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])
+
+            update = self.assoc_scan(grad, expanded_beta.sigmoid(), 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)
+
+            # set the next params and momentum, which can be passed back in
+
+            next_params[key] = param - acc_update[:, -1]
+            next_momentum[key] = update[:, -1]
 
         # losses
 
@@ -290,6 +365,9 @@ class mmTEM(Module):
             inference_pred_loss
         )
 
+        if return_memory_mlp_params:
+            return next_params, next_momentum
+
         if not return_losses:
             return total_loss
 

{hippoformer-0.0.5.dist-info → hippoformer-0.0.7.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hippoformer
-Version: 0.0.5
+Version: 0.0.7
 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-0.0.7.dist-info/RECORD

@@ -0,0 +1,6 @@
+hippoformer/__init__.py,sha256=A7N8GsRAZH4yP-L5hb7IVDnNjnhfjNyolg5MZ6vnGyE,71
+hippoformer/hippoformer.py,sha256=yYoJ5XO0YVAyp3LcRxpunU-0HA97mpCBeQFyi-NSkF0,11549
+hippoformer-0.0.7.dist-info/METADATA,sha256=Xg6NZ6VAQGmuiOo8mMwIAM39Gf6TpVOpyn7o4PMq7JE,2800
+hippoformer-0.0.7.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
+hippoformer-0.0.7.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
+hippoformer-0.0.7.dist-info/RECORD,,

hippoformer-0.0.5.dist-info/RECORD

@@ -1,6 +0,0 @@
-hippoformer/__init__.py,sha256=A7N8GsRAZH4yP-L5hb7IVDnNjnhfjNyolg5MZ6vnGyE,71
-hippoformer/hippoformer.py,sha256=PP2KmTygOP6MyYuhmr_8iEBbywIaTW4TpoIycYRugMo,9142
-hippoformer-0.0.5.dist-info/METADATA,sha256=83iG4F_6ibQy6XSCWht-aF2ZVYmiEq-KSF4XR9YaBtY,2773
-hippoformer-0.0.5.dist-info/WHEEL,sha256=qtCwoSJWgHk21S1Kb4ihdzI2rlJ1ZKaIurTj_ngOhyQ,87
-hippoformer-0.0.5.dist-info/licenses/LICENSE,sha256=1yCiA9b5nhslTavxPjsQAO-wpOnwJR9-l8LTVi7GJuk,1066
-hippoformer-0.0.5.dist-info/RECORD,,