Update app.py

app.py

@@ -1,7 +1,11 @@
 import gradio as gr
 from transformers import TrOCRProcessor, VisionEncoderDecoderModel
 import torch
-import spaces
+import subprocess
+import json
+from PIL import Image, ImageDraw
+import os
+import tempfile
 
 # Dictionary of model names and their corresponding HuggingFace model IDs
 MODEL_OPTIONS = {
@@ -32,93 +36,78 @@ def load_model(model_name):
         current_model = VisionEncoderDecoderModel.from_pretrained(model_id)
         current_model_name = model_name
         
-        # Move model to GPU
-        current_model = current_model.to('cuda')
+        # Move model to GPU if available
+        device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+        current_model = current_model.to(device)
     
     return current_processor, current_model
 
-@spaces.GPU
 def process_image(image, model_name):
+    # Save the uploaded image to a temporary file
+    with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as temp_img:
+        image.save(temp_img, format="JPEG")
+        temp_img_path = temp_img.name
+
+    # Run Kraken for line detection
+    lines_json_path = "lines.json"
+    kraken_command = f"kraken -i {temp_img_path} {lines_json_path} binarize segment -bl"
+    subprocess.run(kraken_command, shell=True, check=True)
+
+    # Load the lines from the JSON file
+    with open(lines_json_path, 'r') as f:
+        lines_data = json.load(f)
+
     processor, model = load_model(model_name)
-    
-    # Prepare image
-    pixel_values = processor(image, return_tensors="pt").pixel_values
-    
-    # Move input to GPU
-    pixel_values = pixel_values.to('cuda')
-    
-    # Generate (no beam search)
-    with torch.no_grad():
-        generated_ids = model.generate(pixel_values)
-    
-    # Decode
-    generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
-    return generated_text
-
-# Base URL for the images
-base_url = "https://huggingface.co/medieval-data/trocr-medieval-base/resolve/main/images/"
-
-# List of example images and their corresponding models
-examples = [
-    [f"{base_url}caroline-1.png", "Medieval Latin Caroline"],
-    [f"{base_url}caroline-2.png", "Medieval Latin Caroline"],
-    [f"{base_url}cursiva-1.png", "Medieval Cursiva"],
-    [f"{base_url}cursiva-2.png", "Medieval Cursiva"],
-    [f"{base_url}cursiva-3.png", "Medieval Cursiva"],
-    [f"{base_url}humanistica-1.png", "Medieval Humanistica"],
-    [f"{base_url}humanistica-2.png", "Medieval Humanistica"],
-    [f"{base_url}humanistica-3.png", "Medieval Humanistica"],
-    [f"{base_url}hybrida-1.png", "Medieval Castilian Hybrida"],
-    [f"{base_url}hybrida-2.png", "Medieval Castilian Hybrida"],
-    [f"{base_url}hybrida-3.png", "Medieval Castilian Hybrida"],
-    [f"{base_url}praegothica-1.png", "Medieval Praegothica"],
-    [f"{base_url}praegothica-2.png", "Medieval Praegothica"],
-    [f"{base_url}praegothica-3.png", "Medieval Praegothica"],
-    [f"{base_url}print-1.png", "Medieval Print"],
-    [f"{base_url}print-2.png", "Medieval Print"],
-    [f"{base_url}print-3.png", "Medieval Print"],
-    [f"{base_url}semihybrida-1.png", "Medieval Semihybrida"],
-    [f"{base_url}semihybrida-2.png", "Medieval Semihybrida"],
-    [f"{base_url}semihybrida-3.png", "Medieval Semihybrida"],
-    [f"{base_url}semitextualis-1.png", "Medieval Semitextualis"],
-    [f"{base_url}semitextualis-2.png", "Medieval Semitextualis"],
-    [f"{base_url}semitextualis-3.png", "Medieval Semitextualis"],
-    [f"{base_url}textualis-1.png", "Medieval Textualis"],
-    [f"{base_url}textualis-2.png", "Medieval Textualis"],
-    [f"{base_url}textualis-3.png", "Medieval Textualis"],
-]
-
-# Custom CSS to make the image wider
-custom_css = """
-#image_upload {
-    max-width: 100% !important;
-    width: 100% !important;
-    height: auto !important;
-}
-#image_upload > div:first-child {
-    width: 100% !important;
-}
-#image_upload img {
-    max-width: 100% !important;
-    width: 100% !important;
-    height: auto !important;
-}
-"""
+
+    # Process each line
+    transcriptions = []
+    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+    for line in lines_data['lines']:
+        # Extract line coordinates
+        x1, y1 = line['baseline'][0]
+        x2, y2 = line['baseline'][-1]
+        
+        # Crop the line from the original image
+        line_image = image.crop((x1, y1, x2, y2))
+
+        # Prepare image for TrOCR
+        pixel_values = processor(line_image, return_tensors="pt").pixel_values
+        pixel_values = pixel_values.to(device)
+
+        # Generate (no beam search)
+        with torch.no_grad():
+            generated_ids = model.generate(pixel_values)
+
+        # Decode
+        generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+        transcriptions.append(generated_text)
+
+    # Clean up temporary files
+    os.unlink(temp_img_path)
+    os.unlink(lines_json_path)
+
+    # Create an image with bounding boxes
+    draw = ImageDraw.Draw(image)
+    for line in lines_data['lines']:
+        coords = line['baseline']
+        draw.line(coords, fill="red", width=2)
+
+    return image, "\n".join(transcriptions)
 
 # Gradio interface
-with gr.Blocks(css=custom_css) as iface:
-    gr.Markdown("# Medieval TrOCR Model Switcher")
-    gr.Markdown("Upload an image of medieval text and select a model to transcribe it. Note: This tool is designed to work on a single line of text at a time for optimal results.")
+with gr.Blocks() as iface:
+    gr.Markdown("# Medieval Document Transcription")
+    gr.Markdown("Upload an image of a medieval document and select a model to transcribe it. The tool will detect lines and transcribe each line separately.")
     
     with gr.Row():
-        input_image = gr.Image(type="pil", label="Input Image", elem_id="image_upload")
+        input_image = gr.Image(type="pil", label="Input Image")
         model_dropdown = gr.Dropdown(choices=list(MODEL_OPTIONS.keys()), label="Select Model", value="Medieval Base")
     
-    transcription_output = gr.Textbox(label="Transcription")
+    with gr.Row():
+        output_image = gr.Image(type="pil", label="Detected Lines")
+        transcription_output = gr.Textbox(label="Transcription", lines=10)
     
     submit_button = gr.Button("Transcribe")
-    submit_button.click(fn=process_image, inputs=[input_image, model_dropdown], outputs=transcription_output)
-    
-    gr.Examples(examples, inputs=[input_image, model_dropdown], outputs=transcription_output)
+    submit_button.click(fn=process_image, inputs=[input_image, model_dropdown], outputs=[output_image, transcription_output])
 
 iface.launch()