Update app.py

app.py

@@ -27,7 +27,7 @@ num_classes = bundle["num_classes"]
 class_names = bundle["class_names"]
 state_dict = bundle["state_dict"]
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
+device = "cpu"
 model = timm.create_model(architecture, pretrained=False, num_classes=num_classes)
 model.load_state_dict(state_dict)
 model.eval().to(device)
@@ -110,10 +110,10 @@ def perform_text_classification_and_format(text: str):
 
         predicted_text_label = str(loaded_predictor_from_hub.predict(text_df_processed).iloc[0])
         if predicted_text_label == "1":
-            has_high_concept = "是"
+            has_high_concept = "Yes"
             confidence = text_classification_probabilities["High Concept"]
         else:
-            has_high_concept = "否"
+            has_high_concept = "No"
             confidence = text_classification_probabilities["No High Concept"]
 
         text_classification_formatted = f"High Concept: {has_high_concept} (Confidence: {confidence:.2f})"
@@ -141,27 +141,28 @@ def generate_prompt_only(image_classification_results, text_classification_proba
     if image_classification_results and "error" not in image_classification_results:
         design_stage = max(image_classification_results, key=image_classification_results.get)
 
-    has_high_concept = "否"
+    has_high_concept = "No"
     confidence = text_classification_probabilities.get("No High Concept", 0.0)
     if predicted_text_label == "1":
-        has_high_concept = "是"
+        has_high_concept = "Yes"
         confidence = text_classification_probabilities.get("High Concept", 0.0)
 
     llm_attitude = llm_attitude_mapping.get(design_stage, llm_attitude_mapping["random"])
 
     prompt = f"""You are an abstract architecture critique interpreter.
 Your audience is a low-level architecture student.
-已知用户处于{design_stage}设计阶段，所以你的态度应该要{llm_attitude}。
-已知用户输入的结果（是/否）含有抽象建筑学概念：{has_high_concept}。
-牢记规则：
-- 撰写一段英文，严格控制在250-350字。
-- 文末必须以完整句子收尾。
-- 不得重复任何观点或句子。
-- 禁止使用警句、口号或平行句式。
-- 不得出现“最终输出”、‘输出结束’、“无后续文本”等元注释。
-- 禁止添加自我反思或系统性备注。
-- 段落末句结束后立即终止输出。
-以下是用户输入的文本内容：{text}你需要用儿童都懂的语言，举生活中的例子给用户解释抽象概念，并且给出可操作的建议。
+The user is currently in the {design_stage} design stage, so your attitude should be {llm_attitude}.
+The user's input contains abstract architectural concepts (Yes/No): {has_high_concept}.
+Rules:
+- Write a paragraph in English, strictly between 250-350 words.
+- End with a complete sentence.
+- Do not repeat any ideas or sentences.
+- Do not use slogans, mottos, or parallel structures.
+- Do not include phrases like 'final output', 'end of output', or meta-comments.
+- Do not add self-reflection or systematic remarks.
+- Stop immediately after the last sentence of the paragraph.
+Here is the user input text: {text}
+You must use simple language that a child could understand, provide everyday life examples to explain abstract concepts, and give actionable suggestions.
 """
     return prompt
 
@@ -191,16 +192,20 @@ with gr.Blocks() as demo_step_by_step:
     gr.Markdown("# Architecture Feedback Generator (Step-by-Step)")
 
     with gr.Row():
-        image_input = gr.Image(type="pil", label="Upload Architectural Image")
-        text_input = gr.Textbox(label="Enter Text Description or Question")
-
-    classify_button = gr.Button("Perform Classification & Generate Prompt")
-    image_output_label = gr.Label(num_top_classes=len(class_names), label="Image Classification Results")
-    text_output_textbox = gr.Textbox(label="Text Classification Results")
-    text_classification_probabilities_state = gr.State()
-    prompt_output_textbox = gr.Textbox(label="Generated Prompt for LLM", interactive=True)
-    generate_feedback_button = gr.Button("Generate Feedback from Prompt")
-    llm_output_text = gr.Textbox(label="Generated Feedback")
+        with gr.Column():
+            image_input = gr.Image(type="pil", label="Upload Architectural Image")
+            text_input = gr.Textbox(label="Enter Text Description or Question", lines=4)
+            classify_button = gr.Button("Perform Classification & Generate Prompt")
+
+        with gr.Column():
+            image_output_label = gr.Label(num_top_classes=len(class_names), label="Image Classification Results")
+            text_output_textbox = gr.Textbox(label="Text Classification Results", lines=2)
+            text_classification_probabilities_state = gr.State()
+            prompt_output_textbox = gr.Textbox(label="Generated Prompt for LLM", interactive=True, lines=8)
+            generate_feedback_button = gr.Button("Generate Feedback from Prompt")
+
+        with gr.Column():
+            llm_output_text = gr.Textbox(label="Generated Feedback", lines=12)
 
     def dynamic_generate_prompt(img_res, txt_prob, txt):
         predicted_label = "1" if txt_prob.get("High Concept",0) > txt_prob.get("No High Concept",0) else "0"