v0.2.3 - eval using onnx models

benchmarks/README.md

@@ -23,6 +23,13 @@ python benchmarks/benchmark.py \
   --device mps \
   --split train
 
+# Benchmark hybrid model with ONNX fallback
+python benchmarks/benchmark.py \
+  --model hybrid \
+  --data-dir data/test \
+  --model-path models/artifacts/layoutlmv3_invoice_ner.onnx \
+  --run-name "layoutlmv3-lora-heuristics-ONNX"
+
 # Benchmark Finetuned LayoutLMv3 only
 python benchmarks/benchmark.py \
   --model layoutlmv3 \

benchmarks/models/hybrid_model.py

@@ -61,10 +61,21 @@ def load(self) -> None:
         Note: Heuristics don't require loading, they're pure pattern matching.
         """
         if self.fallback_model is None:
-            # Default to LayoutLMv3
-            from benchmarks.models.layoutlmv3_model import LayoutLMv3Model
+            # Check for ONNX model path
+            model_path = ""
+            if self.model_config:
+                model_path = self.model_config.get("model_path", "")
 
-            self.fallback_model = LayoutLMv3Model(self.model_config)
+            if str(model_path).endswith(".onnx"):
+                logger.info("Detected ONNX model path, using OnnxModel fallback")
+                from benchmarks.models.onnx_model import OnnxModel
+
+                self.fallback_model = OnnxModel(self.model_config)
+            else:
+                # Default to LayoutLMv3
+                from benchmarks.models.layoutlmv3_model import LayoutLMv3Model
+
+                self.fallback_model = LayoutLMv3Model(self.model_config)
 
         logger.info("Loading hybrid model...")
         logger.info("✓ Heuristics ready (no loading required)")
@@ -142,6 +153,16 @@ def predict(
 
             result.metadata["fallback_used"] = True
             result.metadata["extraction_stage"] = "model_fallback"
+
+            # Get model name from config
+            fb_config = self.fallback_model.get_config()
+            model_path = (
+                fb_config.get("model_path")
+                or fb_config.get("checkpoint_path")
+                or "unknown"
+            )
+            result.metadata["model_name"] = os.path.basename(str(model_path))
+
             result.method = "model_fallback"
 
             return result
@@ -167,8 +188,7 @@ def get_config(self) -> Dict[str, Any]:
             fallback_config = self.fallback_model.get_config()
 
         return {
-            "model_name": "HybridModel (Heuristics + ML)",
-            "model_version": "v1.0",
+            "model_name": "HybridModel (Heuristics + LM)",
             "architecture": "Heuristics → Model Fallback",
             "heuristic_patterns": 14,  # From heuristics.py
             "fallback_model": fallback_config,

pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "invoice"
-version = "0.2.2"
+version = "0.2.3"
 description = "Finetuned LayoutLMv3 model for invoice number extraction"
 authors = [{ name = "Ryan Z. Nie", email = "ryanznie@gatech.edu" }]
 readme = "README.md"
@@ -67,4 +67,4 @@ required-environments = ["sys_platform == 'darwin'"]
 [tool.coverage.run]
 relative_files = true
 branch = true
-source = ["."]
+source = ["."]