fix

app.py

@@ -1,4 +1,5 @@
 import gradio as gr
+import ast
 import numpy as np
 from PIL import Image, ImageDraw
 import torch
@@ -26,28 +27,29 @@ def get_image_data(image_path):
     image_input = image_to_tensor(image_path)
     return image_input
 
-def get_intervention_vector(selected_cells_bef, selected_cells_aft):
-    left = np.reshape(np.zeros((1, 14 * 14)), (14, 14))
-    right = np.reshape(np.zeros((1, 14 * 14)), (14, 14))
-
-    for (i, j) in selected_cells_bef:
-        left[i, j] = 1.
-    for (i, j) in selected_cells_aft:
-        right[i, j] = 1.
+def parse_bool_string(s):
+    try:
+        bool_list = ast.literal_eval(s)
+        if not isinstance(bool_list, list):
+            raise ValueError("The input string must represent a list.")
+        return bool_list
+    except (SyntaxError, ValueError) as e:
+        raise ValueError(f"Invalid input string: {e}")
 
+def get_intervention_vector(selected_cells_bef, selected_cells_aft):
 
     left_map = np.zeros((1, 14 * 14  + 1))
     right_map = np.zeros((1, 14 * 14  + 1))
 
-    left_map[0, 1:] = np.reshape(left, (1, 14 * 14))
-    right_map[0, 1:] = np.reshape(right, (1, 14 * 14))
+    left_map[0, 1:] = np.reshape(selected_cells_bef, (1, 14 * 14))
+    right_map[0, 1:] = np.reshape(selected_cells_aft, (1, 14 * 14))
 
 
-    if len(selected_cells_bef) == 0:
-        left_map[0, 0] = 0.0
+    if np.count_nonzero(selected_cells_bef) == 0:
+        left_map[0, 0] = 1.0
 
-    if len(selected_cells_aft) == 0:
-        right_map[0, 0] = 0.0
+    if np.count_nonzero(selected_cells_aft) == 0:
+        right_map[0, 0] = 1.0
 
 
     return left_map, right_map
@@ -85,7 +87,7 @@ def greedy_decode(model, tokenizer, video, video_mask, gt_left_map, gt_right_map
     return input_caption_ids[:, 1:].tolist(), left_map, right_map
 
 # Dummy prediction function
-def predict_image(image_bef, image_aft, selected_cells_bef, selected_cells_aft):
+def predict_image(image_bef, image_aft, json_data_bef, json_data_aft):
     if image_bef is None:
         return "No image provided", "", ""
     if image_aft is None:
@@ -98,6 +100,9 @@ def predict_image(image_bef, image_aft, selected_cells_bef, selected_cells_aft):
 
     tokenizer = ClipTokenizer()
 
+    selected_cells_bef = np.asarray(parse_bool_string(json_data_bef), np.int32)
+    selected_cells_aft = np.asarray(parse_bool_string(json_data_aft), np.int32)
+
     left_map, right_map = get_intervention_vector(selected_cells_bef, selected_cells_aft)
 
     left_map, right_map = torch.from_numpy(left_map).unsqueeze(0), torch.from_numpy(right_map).unsqueeze(0) 
@@ -228,12 +233,12 @@ with gr.Blocks() as demo:
     3. Click the 'Predict' button to get model predictions
     """)
 
-    selected_cells_bef = gr.State([])
-    selected_cells_aft = gr.State([])
-
     height = gr.State(value=320)
     width = gr.State(value=480)
 
+    sel_attn_bef = gr.Textbox("", visible=False)
+    sel_attn_aft = gr.Textbox("", visible=False)
+
     
     with gr.Row():
         with gr.Column(scale=1):
@@ -282,8 +287,9 @@ with gr.Blocks() as demo:
     # Connect the predict button to the prediction function
     predict_btn.click(
         fn=predict_image,
-        inputs=[image_bef, image_aft, selected_cells_bef, selected_cells_aft],
-        outputs=[prediction, selected_info_bef, selected_info_aft]
+        inputs=[image_bef, image_aft, sel_attn_bef, sel_attn_aft],
+        outputs=[prediction, selected_info_bef, selected_info_aft],
+        _js="(sel_attn_bef, sel_attn_aft) => { return [read_js_Data()]; }"
     )
 
     image_bef.change(

js/interactive_grid.js

@@ -304,7 +304,7 @@ function read_js_Data() {
     console.log("read_js_Data");
     console.log("read_js_Data");
     console.log("read_js_Data");
-    return grid;
+    return grid_bef, grid_aft;
 }