switch to click model

app.py

@@ -6,10 +6,9 @@ import torch
 import html
 from transformers import DonutProcessor, VisionEncoderDecoderModel
 
-pretrained_repo_name = 'ivelin/donut-refexp-combined-v1'
+pretrained_repo_name = 'ivelin/donut-refexp-click'
 pretrained_revision = 'main'
-# revision: '348ddad8e958d370b7e341acd6050330faa0500f' # Iou = 0.47
-# revision: '41210d7c42a22e77711711ec45508a6b63ec380f' # : IoU=0.42 
+# revision can be git commit hash, branch or tag
 # use 'main' for latest revision
 print(f"Loading model checkpoint: {pretrained_repo_name}")
 
@@ -31,7 +30,7 @@ def process_refexp(image: Image, prompt: str):
     pixel_values = processor(image, return_tensors="pt").pixel_values
 
     # prepare decoder inputs
-    task_prompt = "<s_refexp><s_prompt>{user_input}</s_prompt><s_target_bounding_box>"
+    task_prompt = "<s_refexp><s_prompt>{user_input}</s_prompt><s_target_center>"
     prompt = task_prompt.replace("{user_input}", prompt)
     decoder_input_ids = processor.tokenizer(
         prompt, add_special_tokens=False, return_tensors="pt").input_ids
@@ -61,37 +60,28 @@ def process_refexp(image: Image, prompt: str):
         fr"predicted decoder sequence before token2json: {html.escape(sequence)}")
     seqjson = processor.token2json(sequence)
 
-    # safeguard in case predicted sequence does not include a target_bounding_box token
-    bbox = seqjson.get('target_bounding_box')
-    if bbox is None:
+    # safeguard in case predicted sequence does not include a target_center token
+    center_point = seqjson.get('target_center')
+    if center_point is None:
         print(
-            f"token2bbox seq has no predicted target_bounding_box, seq:{seq}")
-        bbox = {"xmin": 0, "ymin": 0, "xmax": 0, "ymax": 0}
-        return bbox
+            f"predicted sequence has no target_center, seq:{sequence}")
+        center_point = {"x": 0, "y": 0}
+        return center_point
 
-    print(f"predicted bounding box with text coordinates: {bbox}")
-    # safeguard in case text prediction is missing some bounding box coordinates
+    print(f"predicted center_point with text coordinates: {center_point}")
+    # safeguard in case text prediction is missing some center point coordinates
     # or coordinates are not valid numeric values
     try:
-        xmin = float(bbox.get("xmin", 0))
+        x = float(center_point.get("x", 0))
     except ValueError:
-        xmin = 0
+        x = 0
     try:
-        ymin = float(bbox.get("ymin", 0))
+        y = float(center_point.get("y", 0))
     except ValueError:
-        ymin = 0
-    try:
-        xmax = float(bbox.get("xmax", 1))
-    except ValueError:
-        xmax = 1
-    try:
-        ymax = float(bbox.get("ymax", 1))
-    except ValueError:
-        ymax = 1
+        y = 0
     # replace str with float coords
-    bbox = {"xmin": xmin, "ymin": ymin, "xmax": xmax,
-            "ymax": ymax, "decoder output sequence": sequence}
-    print(f"predicted bounding box with float coordinates: {bbox}")
+    center_point = {"x": x, "y": y, "decoder output sequence": sequence}
+    print(f"predicted center_point with float coordinates: {center_point}")
 
     print(f"image object: {image}")
     print(f"image size: {image.size}")
@@ -99,26 +89,25 @@ def process_refexp(image: Image, prompt: str):
     print(f"image width, height: {width, height}")
     print(f"processed prompt: {prompt}")
 
-    # safeguard in case text prediction is missing some bounding box coordinates
-    xmin = math.floor(width*bbox["xmin"])
-    ymin = math.floor(height*bbox["ymin"])
-    xmax = math.floor(width*bbox["xmax"])
-    ymax = math.floor(height*bbox["ymax"])
+    # safeguard in case text prediction is missing some center point coordinates
+    x = math.floor(width*center_point["x"])
+    y = math.floor(height*center_point["y"])
 
     print(
-        f"to image pixel values: xmin, ymin, xmax, ymax: {xmin, ymin, xmax, ymax}")
-
-    shape = [(xmin, ymin), (xmax, ymax)]
+        f"to image pixel values: x, y: {x, y}")
 
-    # deaw bbox rectangle
+    # draw center point circle
     img1 = ImageDraw.Draw(image)
-    img1.rectangle(shape, outline="green", width=5)
-    img1.rectangle(shape, outline="white", width=2)
 
-    return image, bbox
+    r = 1
+    shape = [(x-r, y-r), (x+r, y+r)]
+    img1.ellipse(shape, outline="green", width=10)
+    img1.ellipse(shape, outline="white", width=5)
+
+    return image, center_point
 
 
-title = "Demo: Donut 🍩 for UI RefExp (by GuardianUI)"
+title = "Demo: Donut 🍩 for UI RefExp - Center Point (by GuardianUI)"
 description = "Gradio Demo for Donut RefExp task, an instance of `VisionEncoderDecoderModel` fine-tuned on [UIBert RefExp](https://huggingface.co/datasets/ivelin/ui_refexp_saved) Dataset (UI Referring Expression). To use it, simply upload your image and type a prompt and click 'submit', or click one of the examples to load them. See the model training <a href='https://colab.research.google.com/github/ivelin/donut_ui_refexp/blob/main/Fine_tune_Donut_on_UI_RefExp.ipynb' target='_parent'>Colab Notebook</a> for this space. Read more at the links below."
 article = "<p style='text-align: center'><a href='https://arxiv.org/abs/2111.15664' target='_blank'>Donut: OCR-free Document Understanding Transformer</a> | <a href='https://github.com/clovaai/donut' target='_blank'>Github Repo</a></p>"
 examples = [["example_1.jpg", "select the setting icon from top right corner"],