metacontroller-pytorch 0.0.22__tar.gz → 0.0.24__tar.gz

{metacontroller_pytorch-0.0.22 → metacontroller_pytorch-0.0.24}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: metacontroller-pytorch
-Version: 0.0.22
+Version: 0.0.24
 Summary: Transformer Metacontroller
 Project-URL: Homepage, https://pypi.org/project/metacontroller/
 Project-URL: Repository, https://github.com/lucidrains/metacontroller
@@ -39,7 +39,8 @@ Requires-Dist: discrete-continuous-embed-readout>=0.1.12
 Requires-Dist: einops>=0.8.1
 Requires-Dist: einx>=0.3.0
 Requires-Dist: loguru
-Requires-Dist: memmap-replay-buffer>=0.0.1
+Requires-Dist: memmap-replay-buffer>=0.0.23
+Requires-Dist: torch-einops-utils>=0.0.16
 Requires-Dist: torch>=2.5
 Requires-Dist: x-evolution>=0.1.23
 Requires-Dist: x-mlps-pytorch

{metacontroller_pytorch-0.0.22 → metacontroller_pytorch-0.0.24}/metacontroller/metacontroller.py

@@ -26,6 +26,9 @@ from discrete_continuous_embed_readout import Embed, Readout, EmbedAndReadout
 
 from assoc_scan import AssocScan
 
+from torch_einops_utils import pad_at_dim
+from torch_einops_utils.save_load import save_load
+
 # constants
 
 LinearNoBias = partial(Linear, bias = False)
@@ -46,14 +49,6 @@ def default(*args):
             return arg
     return None
 
-def pad_at_dim(t, pad: tuple[int, int], dim = -1, value = 0.):
-    if pad == (0, 0):
-        return t
-
-    dims_from_right = (- dim - 1) if dim < 0 else (t.ndim - dim - 1)
-    zeros = ((0, 0) * dims_from_right)
-    return F.pad(t, (*zeros, *pad), value = value)
-
 # tensor helpers
 
 def straight_through(src, tgt):
@@ -69,6 +64,7 @@ MetaControllerOutput = namedtuple('MetaControllerOutput', (
     'switch_loss'
 ))
 
+@save_load()
 class MetaController(Module):
     def __init__(
         self,
@@ -278,6 +274,7 @@ TransformerOutput = namedtuple('TransformerOutput', (
     'prev_hiddens'
 ))
 
+@save_load()
 class Transformer(Module):
     def __init__(
         self,

{metacontroller_pytorch-0.0.22 → metacontroller_pytorch-0.0.24}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "metacontroller-pytorch"
-version = "0.0.22"
+version = "0.0.24"
 description = "Transformer Metacontroller"
 authors = [
     { name = "Phil Wang", email = "lucidrains@gmail.com" }
@@ -29,8 +29,9 @@ dependencies = [
     "einx>=0.3.0",
     "einops>=0.8.1",
     "loguru",
-    "memmap-replay-buffer>=0.0.1",
+    "memmap-replay-buffer>=0.0.23",
     "torch>=2.5",
+    "torch-einops-utils>=0.0.16",
     "x-evolution>=0.1.23",
     "x-mlps-pytorch",
     "x-transformers"

{metacontroller_pytorch-0.0.22 → metacontroller_pytorch-0.0.24}/tests/test_metacontroller.py

@@ -1,6 +1,8 @@
 import pytest
 param = pytest.mark.parametrize
 
+from pathlib import Path
+
 import torch
 from metacontroller.metacontroller import Transformer, MetaController
 
@@ -68,3 +70,16 @@ def test_metacontroller(
 
     model.meta_controller = meta_controller
     model.evolve(1, lambda _: 1., noise_population_size = 2)
+
+    # saving and loading
+
+    meta_controller.save('./meta_controller.pt')
+
+    rehydrated_meta_controller = MetaController.init_and_load('./meta_controller.pt')
+
+    model.save('./trained.pt')
+
+    rehydrated_model = Transformer.init_and_load('./trained.pt', strict = False)
+
+    Path('./meta_controller.pt').unlink()
+    Path('./trained.pt').unlink()

{metacontroller_pytorch-0.0.22 → metacontroller_pytorch-0.0.24}/train.py

@@ -3,7 +3,7 @@
 #   "fire",
 #   "gymnasium",
 #   "gymnasium[other]",
-#   "memmap-replay-buffer>=0.0.10",
+#   "memmap-replay-buffer>=0.0.12",
 #   "metacontroller-pytorch",
 #   "minigrid",
 #   "tqdm"