Eliminated dependency on paddleocr

Change to trocr

requirements.txt

@@ -9,9 +9,10 @@ nltk
 python-multipart
 
 augraphy
-onnxruntime-gpu
 
 streamlit==1.30
 streamlit-paste-button
 
-paddleocr
+shapely
+pyclipper
+onnxruntime-gpu

src/models/ocr_model/model/TexTeller.py

@@ -4,29 +4,26 @@ from ...globals import (
     VOCAB_SIZE,
     FIXED_IMG_SIZE,
     IMG_CHANNELS,
+    MAX_TOKEN_SIZE
 )
 
 from transformers import (
-    ViTConfig,
-    ViTModel,
-    TrOCRConfig,
-    TrOCRForCausalLM,
     RobertaTokenizerFast,
     VisionEncoderDecoderModel,
+    VisionEncoderDecoderConfig
 )
 
 
 class TexTeller(VisionEncoderDecoderModel):
     REPO_NAME = 'OleehyO/TexTeller'
-    def __init__(self, decoder_path=None, tokenizer_path=None):
+    def __init__(self):
-        encoder = ViTModel(ViTConfig(
+        config = VisionEncoderDecoderConfig.from_pretrained('/home/lhy/code/TexTeller/src/models/ocr_model/model/trocr-small')
-            image_size=FIXED_IMG_SIZE,
+        config.encoder.image_size = FIXED_IMG_SIZE
-            num_channels=IMG_CHANNELS
+        config.encoder.num_channels = IMG_CHANNELS
-        ))
+        config.decoder.vocab_size=VOCAB_SIZE
-        decoder = TrOCRForCausalLM(TrOCRConfig(
+        config.decoder.max_position_embeddings=MAX_TOKEN_SIZE
-            vocab_size=VOCAB_SIZE,
+
-        ))
+        super().__init__(config=config)
-        super().__init__(encoder=encoder, decoder=decoder)
     
     @classmethod
     def from_pretrained(cls, model_path: str = None):