Results form a short training run

Author: liebharc

Training.md

@@ -77,6 +77,16 @@ This validation provides a **more representative indication of overall system pe
 
 Implementation: `rate_validation_result.py`
 
+## Run 410 - at epoch 7
+
+Commit: 23c9006d8f23b8f178b5c7ee64bb9cb430c4d7bc
+Day: 03 July 2026
+Transformer Smoke Test: 12%
+System Level: 10.6 diffs, SER: 6.4%
+Polish scores with musicdiff: OMR-NED 47.3%
+
+Quick training run after updating data augmentation.
+
 ## Run 407 - at epoch 6
 
 Commit: fabab4b0f8480be20d13686edf5f91d913ab00fd

training/omr_datasets/convert_lieder.py

@@ -322,11 +322,18 @@ def _position_to_staff_no(self, symbol: EncodedSymbol) -> int:
             return 1
         return 0
 
-    def extract_measures(self, count: int) -> list[EncodedSymbol]:
+    def extract_measures(
+        self, count: int, always_include_time: bool = False
+    ) -> list[EncodedSymbol]:
         clefs = self.clefs.copy()
         key = self.key
         time = self.time
-        has_time = False
+        # Lieder pages are crops of one continuously rendered score, so a courtesy time
+        # signature is only visible where the source XML actually redeclares it. pdmx and
+        # musetrainer windows are each re-rendered standalone (see generate_xml), and that
+        # renderer always draws a time signature on a fresh score - so those callers pass
+        # always_include_time=True to keep the label in sync with the image.
+        has_time = always_include_time
         result: list[EncodedSymbol] = []
         for i in range(count):
             selected_measure = self.voice.pop(0)
@@ -457,31 +464,32 @@ def _split_file_into_staffs(
     return result
 
 
+def _leading_clefs(measure: Measure) -> list[EncodedSymbol]:
+    # The clefs of a measure always precede its notes/rests, but other preamble symbols
+    # (e.g. repeatStart on the very first measure) can come before them, so scan rather
+    # than assume fixed indices.
+    clefs = []
+    for symbol in measure:
+        if symbol.rhythm.startswith(("note", "rest")):
+            break
+        if symbol.rhythm.startswith("clef"):
+            clefs.append(symbol)
+    return clefs
+
+
 def _count_staffs(voice: list[Measure]) -> int:
     if len(voice) == 0:
         return 0
-    first_measure = voice[0]
-    if len(first_measure) == 0:
+    clefs = _leading_clefs(voice[0])
+    if len(clefs) == 0:
         return 0
-    if len(first_measure) < 3:
-        return 0
-    third_symbol = first_measure[2]
-    if third_symbol.rhythm.startswith("clef"):
-        return 2
-    return 1
+    return 2 if len(clefs) >= 2 else 1
 
 
 def is_grandstaff(voice: list[Measure]) -> bool:
     if len(voice) == 0:
         return False
-    first_measure = voice[0]
-    if len(first_measure) < 3:
-        return False
-    return (
-        first_measure[0].rhythm.startswith("clef")
-        and first_measure[1].rhythm == "chord"
-        and first_measure[2].rhythm.startswith("clef")
-    )
+    return len(_leading_clefs(voice[0])) >= 2
 
 
 def get_svg_voice_count(voice: list[Measure]) -> int:

training/omr_datasets/convert_musetrainer.py

@@ -268,7 +268,7 @@ def _convert_file_impl(path: Path) -> list[str]:
             cutter.key = key
             cutter.time = time_sym
 
-            tokens = cutter.extract_measures(len(window_measures))
+            tokens = cutter.extract_measures(len(window_measures), always_include_time=True)
 
             if calc_ratio_of_tuplets(tokens) <= 0.2 and contains_only_supported_clefs(tokens):
                 tokens = strip_naturals(tokens)

training/omr_datasets/convert_pdmx.py

@@ -106,7 +106,7 @@ def _convert_file_impl(mxl_path: Path) -> list[str]:
             cutter.key = key
             cutter.time = time_sym
 
-            tokens = cutter.extract_measures(len(window_measures))
+            tokens = cutter.extract_measures(len(window_measures), always_include_time=True)
 
             if calc_ratio_of_tuplets(tokens) <= 0.2 and contains_only_supported_clefs(tokens):
                 tokens = strip_naturals(tokens)