workbench 0.8.415__tar.gz → 0.8.416__tar.gz

{workbench-0.8.415 → workbench-0.8.416}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: workbench
-Version: 0.8.415
+Version: 0.8.416
 Summary: Workbench: A Dashboard and Python API for creating and deploying AWS SageMaker Model Pipelines
 Author-email: SuperCowPowers LLC <support@supercowpowers.com>
 License: MIT License

{workbench-0.8.415 → workbench-0.8.416}/src/workbench/training/chemprop_hpo.py

@@ -27,13 +27,18 @@ from workbench.training.hpo_harness import Choice, FloatRange, IntRange
 # split_strategy, criterion, seed — stays fixed at its configured value.
 _SEARCH_GROUPS = {
     "basic": {
-        # depth / hidden_dim / dropout / ffn_num_layers ranges are chemprop's canonical
-        # search space verbatim. ffn_hidden_dim additionally offers tapered heads (lists),
-        # which chemprop's own space can't express — ffn_num_layers is ignored when a
-        # tapered list is chosen (the list length sets the depth).
+        # depth / hidden_dim / ffn_num_layers ranges are chemprop's canonical search space
+        # verbatim. ffn_hidden_dim additionally offers tapered heads (lists), which
+        # chemprop's own space can't express — ffn_num_layers is ignored when a tapered
+        # list is chosen (the list length sets the depth).
+        #
+        # `dropout` is deliberately absent even though chemprop searches it: a trial trains
+        # a *single* model, while the published model is an `n_folds` ensemble. Dropout
+        # chosen without ensembling over-regularizes once the ensemble is stacked on top
+        # (DEFAULT_HYPERPARAMETERS keeps it low precisely because the ensemble already
+        # regularizes), so it stays at that ensemble-tuned default.
         "depth": IntRange(2, 6, 1),
         "hidden_dim": IntRange(300, 2400, 100),
-        "dropout": FloatRange(0.0, 0.4, 0.05),
         "ffn_num_layers": IntRange(1, 3, 1),
         "ffn_hidden_dim": Choice([300, 600, 1200, 1800, 2400, [1024, 256, 64], [512, 128]]),
     },
@@ -48,6 +53,11 @@ _SEARCH_GROUPS = {
 
 # Knobs outside the default space — the working list for when this is revisited:
 #
+#   * `dropout` — chemprop searches it ({0.0,0.05,…,0.4}); we can't transfer a value found
+#     single-model to an ensemble (see the `basic` group note). Searching it honestly needs
+#     trials scored in the *deployment* regime — e.g. shortlist cheaply single-fold, then
+#     re-rank the top-K configs by training each as a real `n_folds` ensemble and selecting
+#     on that. Same treatment would make the capacity knobs transfer better too.
 #   * learning rate (`max_lr`, `warmup_epochs`) — already available as the opt-in "lr"
 #     group. Chemprop's maintainers describe their recommended search as focusing on
 #     learning rate and batch size, which makes this the first candidate to promote into

{workbench-0.8.415 → workbench-0.8.416}/src/workbench.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: workbench
-Version: 0.8.415
+Version: 0.8.416
 Summary: Workbench: A Dashboard and Python API for creating and deploying AWS SageMaker Model Pipelines
 Author-email: SuperCowPowers LLC <support@supercowpowers.com>
 License: MIT License

{workbench-0.8.415 → workbench-0.8.416}/src/workbench.egg-info/scm_version.json

@@ -1,7 +1,7 @@
 {
-  "tag": "0.8.415",
+  "tag": "0.8.416",
   "distance": 0,
-  "node": "gd26599192a9c2b2e08473a6852cca6e837bff4ae",
+  "node": "g4a56f50f91f9c0dc52aafebbc2ac306a95b8e38d",
   "dirty": false,
   "branch": "HEAD",
   "node_date": "2026-07-23"

{workbench-0.8.415 → workbench-0.8.416}/tests/training/chemprop_hpo_test.py

@@ -6,17 +6,19 @@ imports are deferred in the search entry point, not at module top).
 
 # Workbench Imports
 from workbench.training.chemprop_hpo import chemprop_search_space, merge_best_config, resolve_search_space
-from workbench.training.hpo_harness import Choice, FloatRange, IntRange
+from workbench.training.hpo_harness import Choice, IntRange
 
 
 def test_basic_group_is_default():
     """The default space is the `basic` group, matching chemprop's canonical ranges."""
     space = chemprop_search_space()
-    assert set(space) == {"depth", "hidden_dim", "dropout", "ffn_num_layers", "ffn_hidden_dim"}
+    assert set(space) == {"depth", "hidden_dim", "ffn_num_layers", "ffn_hidden_dim"}
     assert space["depth"] == IntRange(2, 6, 1)  # chemprop {2,3,4,5,6}
     assert space["hidden_dim"] == IntRange(300, 2400, 100)  # chemprop {300,400,...,2400}
-    assert space["dropout"] == FloatRange(0.0, 0.4, 0.05)  # chemprop {0.0,0.05,...,0.4}
     assert space["ffn_num_layers"] == IntRange(1, 3, 1)  # chemprop {1,2,3}
+    # dropout is NOT searched: single-model trials can't transfer it to the published
+    # n_folds ensemble, so it stays at the ensemble-tuned default.
+    assert "dropout" not in space
     assert isinstance(space["ffn_hidden_dim"], Choice)
     assert [1024, 256, 64] in space["ffn_hidden_dim"].options  # tapered head is a choice