Update inference.py

inference.py

@@ -4,6 +4,9 @@ from PIL import Image
 from transformers import DonutProcessor, VisionEncoderDecoderModel
 import io
 import json
+import logging
+
+logging.basicConfig(level=logging.INFO)
 
 def model_fn(model_dir, context=None):
     device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
@@ -12,37 +15,51 @@ def model_fn(model_dir, context=None):
     model.to(device)
     return model, processor, device
 
-def transform_fn(model, request_body, request_content_type, response_content_type, context=None):
+def input_fn(input_data, content_type, context=None):
+    """Deserialize the input data."""
+    logging.info("Entering input_fn")
+    if content_type == 'application/json':
+        image = Image.open(io.BytesIO(input_data))
+        return image
+    else:
+        raise ValueError(f"Unsupported content type: {content_type}")
+
+def predict_fn(data, model, context=None):
+    """Apply the model to the input data."""
+    logging.info("Entering predict_fn")
     model, processor, device = model
-    if request_content_type == 'application/json':
-        image = Image.open(io.BytesIO(request_body))
-        
-        # Preprocess the image
-        pixel_values = processor(image, return_tensors="pt").pixel_values.to(device)
-        
-        # Run inference
-        model.eval()
-        with torch.no_grad():
-            task_prompt = "<s_receipt>"
-            decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids.to(device)
-            generated_outputs = model.generate(
-                pixel_values,
-                decoder_input_ids=decoder_input_ids,
-                max_length=model.config.decoder.max_position_embeddings, 
-                pad_token_id=processor.tokenizer.pad_token_id,
-                eos_token_id=processor.tokenizer.eos_token_id,
-                early_stopping=True,
-                bad_words_ids=[[processor.tokenizer.unk_token_id]],
-                return_dict_in_generate=True
-            )
-        
-        # Decode the output
-        decoded_text = processor.batch_decode(generated_outputs.sequences)[0]
-        decoded_text = decoded_text.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
-        decoded_text = re.sub(r"<.*?>", "", decoded_text, count=1).strip()
-        
-        # Prepare the response
-        prediction = {'result': decoded_text}
-        return json.dumps(prediction), response_content_type
+
+    # Preprocess the image
+    pixel_values = processor(data, return_tensors="pt").pixel_values.to(device)
+    
+    # Run inference
+    model.eval()
+    with torch.no_grad():
+        task_prompt = "<s_receipt>"
+        decoder_input_ids = processor.tokenizer(task_prompt, add_special_tokens=False, return_tensors="pt").input_ids.to(device)
+        generated_outputs = model.generate(
+            pixel_values,
+            decoder_input_ids=decoder_input_ids,
+            max_length=model.config.decoder.max_position_embeddings, 
+            pad_token_id=processor.tokenizer.pad_token_id,
+            eos_token_id=processor.tokenizer.eos_token_id,
+            early_stopping=True,
+            bad_words_ids=[[processor.tokenizer.unk_token_id]],
+            return_dict_in_generate=True
+        )
+    
+    # Decode the output
+    decoded_text = processor.batch_decode(generated_outputs.sequences)[0]
+    decoded_text = decoded_text.replace(processor.tokenizer.eos_token, "").replace(processor.tokenizer.pad_token, "")
+    decoded_text = re.sub(r"<.*?>", "", decoded_text, count=1).strip()
+
+    prediction = {'result': decoded_text}
+    return prediction
+
+def output_fn(prediction, accept):
+    """Serialize the prediction output."""
+    logging.info("Entering output_fn")
+    if accept == 'application/json':
+        return json.dumps(prediction)
     else:
-        raise ValueError(f"Unsupported content type: {request_content_type}")
+        raise ValueError(f"Unsupported response content type: {accept}")