SGD - fix extra model refresh when sample comes before data.

Author: cboulay

pyproject.toml

@@ -10,7 +10,7 @@ requires-python = ">=3.10.15"
 dynamic = ["version"]
 dependencies = [
     "ezmsg-baseproc>=1.0.2",
-    "ezmsg-sigproc",
+    "ezmsg-sigproc>=2.10.0",
     "river>=0.22.0",
     "scikit-learn>=1.6.0",
     "torch>=2.6.0",

src/ezmsg/learn/process/sgd.py

@@ -62,7 +62,6 @@ def _refreshed_model(self):
 
     def _reset_state(self, message: AxisArray) -> None:
         self._state.model = self._refreshed_model()
-        self._state.b_first_train = True
 
     def _process(self, message: AxisArray) -> ClassifierMessage | None:
         if self._state.model is None or not message.data.size:
@@ -89,10 +88,9 @@ def _process(self, message: AxisArray) -> ClassifierMessage | None:
         )
 
     def partial_fit(self, message: SampleMessage) -> None:
-        if self._state.model is None:
-            # Initialize model on first training sample
-            self._state.model = self._refreshed_model()
-            self._state.b_first_train = True
+        if self._hash != 0:
+            self._reset_state(message.sample)
+            self._hash = 0
 
         if np.any(np.isnan(message.sample.data)):
             return

tests/unit/test_sgd.py

@@ -53,7 +53,7 @@ def test_sgd():
 
     # Now let's try training on all samples
     for sample in samples:
-        decoder(sample)
+        decoder(sample)  # Should invoke partial_fit
     # Then doing inference on all multi-wins
     probas = [decoder(win) for win in windows]
 
@@ -71,7 +71,7 @@ def test_sgd():
     )
     probas = []
     for samp_ix, samp in enumerate(samples):
-        decoder(samp)
+        decoder(samp)  # Should invoke partial_fit
         probas.append(decoder(windows[samp_ix * 2]))
         probas.append(decoder(windows[samp_ix * 2 + 1]))
     class_ids = []