Add screenshot text location feature

Also add option to try more detailed captioning.

app.py

@@ -1,9 +1,8 @@
 import gradio as gr
+import re
 import torch
-from transformers import FuyuForCausalLM, AutoTokenizer
-from transformers.models.fuyu.processing_fuyu import FuyuProcessor
-from transformers.models.fuyu.image_processing_fuyu import FuyuImageProcessor
 from PIL import Image
+from transformers import AutoTokenizer, FuyuForCausalLM, FuyuImageProcessor, FuyuProcessor
 
 model_id = "adept/fuyu-8b"
 dtype = torch.bfloat16
@@ -13,36 +12,89 @@ tokenizer = AutoTokenizer.from_pretrained(model_id)
 model = FuyuForCausalLM.from_pretrained(model_id, device_map="auto", torch_dtype=dtype)
 processor = FuyuProcessor(image_processor=FuyuImageProcessor(), tokenizer=tokenizer)
 
-caption_prompt = "Generate a coco-style caption.\\n"
-
-def resize_to_max(image, max_width=1080, max_height=1080):
-    width, height = image.size
-    if width <= max_width and height <= max_height:
-        return image
-
-    scale = min(max_width/width, max_height/height)
-    width = int(width*scale)
-    height = int(height*scale)
-
-    return image.resize((width, height), Image.LANCZOS)
+CAPTION_PROMPT = "Generate a coco-style caption.\n"
+DETAILED_CAPTION_PROMPT = "What is happening in this image?"
 
 def predict(image, prompt):
     # image = image.convert('RGB')
-    image = resize_to_max(image)
-
     model_inputs = processor(text=prompt, images=[image])
     model_inputs = {k: v.to(dtype=dtype if torch.is_floating_point(v) else v.dtype, device=device) for k,v in model_inputs.items()}
 
-    generation_output = model.generate(**model_inputs, max_new_tokens=40)
+    generation_output = model.generate(**model_inputs, max_new_tokens=50)
     prompt_len = model_inputs["input_ids"].shape[-1]
     return tokenizer.decode(generation_output[0][prompt_len:], skip_special_tokens=True)
 
-def caption(image):
-    return predict(image, caption_prompt)
+def caption(image, detailed_captioning):
+    if detailed_captioning:
+        caption_prompt = DETAILED_CAPTION_PROMPT
+    else:
+        caption_prompt = CAPTION_PROMPT
+    return predict(image, caption_prompt).lstrip()
 
 def set_example_image(example: list) -> dict:
     return gr.Image.update(value=example[0])
 
+def scale_factor_to_fit(original_size, target_size=(1920, 1080)):
+    width, height = original_size
+    max_width, max_height = target_size
+    if width <= max_width and height <= max_height:
+        return 1.0
+    return min(max_width/width, max_height/height)
+    
+def tokens_to_box(tokens, original_size):
+    bbox_start = tokenizer.convert_tokens_to_ids("<0x00>")
+    bbox_end = tokenizer.convert_tokens_to_ids("<0x01>")
+    try:
+        # Assumes a single box
+        bbox_start_pos = (tokens == bbox_start).nonzero(as_tuple=True)[0].item()
+        bbox_end_pos = (tokens == bbox_end).nonzero(as_tuple=True)[0].item()
+        
+        if bbox_end_pos != bbox_start_pos + 5:
+            return tokens
+
+        # Retrieve transformed coordinates from tokens
+        coords = tokenizer.convert_ids_to_tokens(tokens[bbox_start_pos+1:bbox_end_pos])
+
+        # Scale back to original image size and multiply by 2
+        scale = scale_factor_to_fit(original_size)
+        top, left, bottom, right = [2 * int(float(c)/scale) for c in coords]
+        
+        # Replace the IDs so they get detokenized right
+        replacement = f" <box>{top}, {left}, {bottom}, {right}</box>"
+        replacement = tokenizer.tokenize(replacement)[1:]
+        replacement = tokenizer.convert_tokens_to_ids(replacement)
+        replacement = torch.tensor(replacement).to(tokens)
+
+        tokens = torch.cat([tokens[:bbox_start_pos], replacement, tokens[bbox_end_pos+1:]], 0)
+        return tokens
+    except:
+        gr.Error("Can't convert tokens.")
+        return tokens
+
+def coords_from_response(response):
+    # y1, x1, y2, x2
+    pattern = r"<box>(\d+),\s*(\d+),\s*(\d+),\s*(\d+)</box>"
+
+    match = re.search(pattern, response)
+    if match:
+        # Unpack and change order
+        y1, x1, y2, x2 = [int(coord) for coord in match.groups()]
+        return (x1, y1, x2, y2)
+    else:
+        gr.Error("The string is malformed or does not match the expected pattern.")
+        
+def localize(image, query):
+    prompt= f"When presented with a box, perform OCR to extract text contained within it. If provided with text, generate the corresponding bounding box.\n{query}"
+    model_inputs = processor(text=prompt, images=[image])
+    model_inputs = {k: v.to(dtype=dtype if torch.is_floating_point(v) else v.dtype, device=device) for k,v in model_inputs.items()}
+    
+    generation_output = model.generate(**model_inputs, max_new_tokens=40)
+    prompt_len = model_inputs["input_ids"].shape[-1]
+    tokens = generation_output[0][prompt_len:]
+    tokens = tokens_to_box(tokens, image.size)
+    decoded = tokenizer.decode(tokens, skip_special_tokens=True)
+    coords = coords_from_response(decoded)
+    return image, [(coords, f"Location of \"{query}\"")]
 
 
 css = """
@@ -88,21 +140,44 @@ with gr.Blocks(css=css) as demo:
         
     with gr.Tab("Image Captioning"):
         with gr.Row():
-            captioning_input = gr.Image(label="Upload your Image", type="pil")
+            with gr.Column():
+                captioning_input = gr.Image(label="Upload your Image", type="pil")
+                detailed_captioning_checkbox = gr.Checkbox(label="Enable detailed captioning")
             captioning_output = gr.Textbox(label="Output")
         captioning_btn = gr.Button("Generate Caption")
 
         gr.Examples(
-            [["assets/captioning_example_1.png"], ["assets/captioning_example_2.png"]],
-            inputs = [captioning_input],
+            [["assets/captioning_example_1.png", False], ["assets/captioning_example_2.png", True]],
+            inputs = [captioning_input, detailed_captioning_checkbox],
             outputs = [captioning_output],
             fn=caption,
             cache_examples=True,
             label='Click on any Examples below to get captioning results quickly 👇'
             )
         
-    captioning_btn.click(fn=caption, inputs=captioning_input, outputs=captioning_output)
+    captioning_btn.click(fn=caption, inputs=[captioning_input, detailed_captioning_checkbox], outputs=captioning_output)
     vqa_btn.click(fn=predict, inputs=[image_input, text_input], outputs=vqa_output)
 
+    with gr.Tab("Find Text in Screenshots"):
+        gr.Markdown("This demo is designed to locate text in desktop screenshots. Please, ensure to upload images of 1920x1080 for best results!")
+        with gr.Row():
+            with gr.Column():
+                localization_input = gr.Image(label="Upload your Image", type="pil")
+                query_input = gr.Textbox(label="Text to find")
+                localization_btn = gr.Button("Locate Text")
+            with gr.Column():
+                with gr.Row(height=800):
+                    localization_output = gr.AnnotatedImage(label="Text Position")
+
+        gr.Examples(
+            [["assets/localization_example_1.jpeg", "Share your repair"]],
+            inputs = [localization_input, query_input],
+            outputs = [localization_output],
+            fn=localize,
+            cache_examples=True,
+            label='Click on any Examples below to get localization results quickly 👇'
+            )
+    
+    localization_btn.click(fn=localize, inputs=[localization_input, query_input], outputs=localization_output)   
     
 demo.launch(server_name="0.0.0.0")