fix latexocr bug (#4178)

paddlex/inference/models/formula_recognition/predictor.py

@@ -136,7 +136,7 @@ class FormulaRecPredictor(BasePredictor):
         }
 
     @register("DecodeImage")
-    def build_readimg(self, channel_first, img_mode):
+    def build_readimg(self, channel_first, img_mode="RGB"):
         assert channel_first == False
         return "Read", ReadImage(format=img_mode)
 

paddlex/inference/models/formula_recognition/processors.py

@@ -365,7 +365,7 @@ class LaTeXOCRDecode(object):
         dec = [self.tokenizer.decode(tok) for tok in tokens]
         dec_str_list = [
             "".join(detok.split(" "))
-            .replace("臓", " ")
+            .replace("Ġ", " ")
             .replace("[EOS]", "")
             .replace("[BOS]", "")
             .replace("[PAD]", "")