From 12be9834872f511f61a2c17b836ce2a8988d2898 Mon Sep 17 00:00:00 2001
From: Robert Sachunsky <sachunsky@informatik.uni-leipzig.de>
Date: Sun, 12 Jul 2026 03:49:41 +0200
Subject: [PATCH] =?UTF-8?q?cnn-rnn-ocr=20inference:=20switch=20back=20to?=
 =?UTF-8?q?=20beam=20search,=20only=20run=20in=20TF=E2=80=A6?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

- training.models.CTCDecoder: prefer beam search over greedy
  (because it is more accurate)
- training reload makefile: skip ONNX and TF-Serving conversion
  for cnn-rnn-ocr models (because these would not work)
- training reload makefile: default to onnx and tf conversions
  for all models (tf for training and onnx for inference)
  instead of tf-serving export
---
 src/eynollah/training/models.py             |  4 +++-
 src/eynollah/training/reload-models-v0.8.mk | 10 ++++++++--
 2 files changed, 11 insertions(+), 3 deletions(-)

diff --git a/src/eynollah/training/models.py b/src/eynollah/training/models.py
index eda6b0f..9076f0a 100644
--- a/src/eynollah/training/models.py
+++ b/src/eynollah/training/models.py
@@ -85,9 +85,11 @@ class CTCDecoder(Layer):
         # tf.compat.v1.nn.ctc_beam_search_decoder() also needs merge_repeated=False
         # tf.nn.ctc_beam_search_decoder() is not supported by ONNX, yet
         # tf.nn.ctc_greedy_decoder() is not as precise, though:
-        decoded, logits = tf.nn.ctc_greedy_decoder(
+        decoded, logits = tf.nn.ctc_beam_search_decoder(
             inputs,
             lengths,
+            beam_width=10,
+            top_paths=1
         )
         # get top path for all sequences in batch
         decoded = decoded[0]
diff --git a/src/eynollah/training/reload-models-v0.8.mk b/src/eynollah/training/reload-models-v0.8.mk
index 342d211..eab5ffb 100644
--- a/src/eynollah/training/reload-models-v0.8.mk
+++ b/src/eynollah/training/reload-models-v0.8.mk
@@ -23,7 +23,9 @@ CURRENT_MODELS += eynollah-binarization_20210425
 CURRENT_MODELS += eynollah-column-classifier_20210425
 CURRENT_MODELS += eynollah-enhancement_20210425
 
-all: tf-serving
+# tf (SavedModel format) for training
+# onnx conversion for fast inference
+all: tf onnx
 
 tf-serving: $(CURRENT_MODELS:%=$(MODELS_DST)/%)
 tf: $(CURRENT_MODELS:%=$(MODELS_DST)/%)
@@ -31,8 +33,11 @@ keras: $(CURRENT_MODELS:%=$(MODELS_DST)/%.keras)
 hdf5: $(CURRENT_MODELS:%=$(MODELS_DST)/%.h5)
 onnx: $(CURRENT_MODELS:%=$(MODELS_DST)/%.onnx)
 
-$(MODELS_DST)/%: FORMAT = $(or $(filter tf,$(MAKECMDGOALS)), tf-serving)
+# distinguish tf from tf-serving: target pattern is the same,
+# so check if either is current goal, otherwise assumg tf
+$(MODELS_DST)/%: FORMAT = $(or $(filter tf-serving,$(MAKECMDGOALS)), tf)
 $(MODELS_DST)/%: $(MODELS_SRC)/%
+	$(if $(and $(filter tf-serving,$(FORMAT)),$(findstring _ocr,$@)),$(warning skipping $@: OCR CTC decoder fails in TF-Serving) : )\
 	eynollah-training convert \
 		$(and $(wildcard $</config.json),--rebuild) \
 		--in $< \
@@ -57,6 +62,7 @@ $(MODELS_DST)/%.h5: $(MODELS_SRC)/%
 	> $(notdir $<).hdf5.log 2>&1 || { cat $(notdir $<).hdf5.log; false; }
 
 $(MODELS_DST)/%.onnx: $(MODELS_SRC)/%
+	$(if $(findstring _ocr,$@),$(warning skipping $@: OCR CTC decoder is buggy in ONNX) : )\
 	eynollah-training convert \
 		$(and $(wildcard $</config.json),--rebuild) \
 		--in $< \