Update app.py

app.py

@@ -1,7 +1,7 @@
 import gradio as gr
 import spaces
 import torch
-from transformers import AutoTokenizer, AutoModelForSequenceClassification
+from transformers import AutoConfig, AutoTokenizer, AutoModelForSequenceClassification
 import torch.nn.functional as F
 import torch.nn as nn
 import re
@@ -9,75 +9,113 @@ import requests
 from urllib.parse import urlparse
 import xml.etree.ElementTree as ET
 
-# Model repository and device setup
-model_path = 'ssocean/NAIP'
+model_path = r'ssocean/NAIP'
 device = 'cuda' if torch.cuda.is_available() else 'cpu'
 
-# Globals
+global model, tokenizer
 model = None
 tokenizer = None
 
 def fetch_arxiv_paper(arxiv_input):
-    """Fetch title & abstract from arXiv URL or ID."""
+    """Fetch paper details from arXiv URL or ID using requests."""
     try:
+        # Extract arXiv ID from URL or use directly
         if 'arxiv.org' in arxiv_input:
             parsed = urlparse(arxiv_input)
-            arxiv_id = parsed.path.split('/')[-1].replace('.pdf', '')
+            path = parsed.path
+            arxiv_id = path.split('/')[-1].replace('.pdf', '')
         else:
             arxiv_id = arxiv_input.strip()
+
+        # Fetch metadata using arXiv API
         api_url = f'http://export.arxiv.org/api/query?id_list={arxiv_id}'
-        resp = requests.get(api_url)
-        if resp.status_code != 200:
-            return {"title":"", "abstract":"", "success":False, "message":"arXiv API error"}
-        root = ET.fromstring(resp.text)
-        ns = {'atom':'http://www.w3.org/2005/Atom'}
-        entry = root.find('.//atom:entry', ns)
+        response = requests.get(api_url)
+        
+        if response.status_code != 200:
+            return {
+                "title": "",
+                "abstract": "",
+                "success": False,
+                "message": "Error fetching paper from arXiv API"
+            }
+
+        # Parse the response XML
+        root = ET.fromstring(response.text)
+        
+        # ArXiv API uses Atom namespace
+        ns = {'arxiv': 'http://www.w3.org/2005/Atom'}
+        
+        # Extract title and abstract
+        entry = root.find('.//arxiv:entry', ns)
         if entry is None:
-            return {"title":"", "abstract":"", "success":False, "message":"Paper not found"}
-        title = entry.find('atom:title', ns).text.strip()
-        abstract = entry.find('atom:summary', ns).text.strip()
-        return {"title":title, "abstract":abstract, "success":True, "message":"Fetched successfully"}
+            return {
+                "title": "",
+                "abstract": "",
+                "success": False,
+                "message": "Paper not found"
+            }
+            
+        title = entry.find('arxiv:title', ns).text.strip()
+        abstract = entry.find('arxiv:summary', ns).text.strip()
+        
+        return {
+            "title": title,
+            "abstract": abstract,
+            "success": True,
+            "message": "Paper fetched successfully!"
+        }
     except Exception as e:
-        return {"title":"", "abstract":"", "success":False, "message":f"Error: {e}"}
+        return {
+            "title": "",
+            "abstract": "",
+            "success": False,
+            "message": f"Error fetching paper: {e}"
+        }
 
 @spaces.GPU(duration=60, enable_queue=True)
 def predict(title, abstract):
-    """Predict a normalized impact score (0–1) from title & abstract."""
+    """Predict a normalized academic impact score (0–1) from title & abstract."""
+    title = title.replace("\n", " ").strip().replace("''", "'")
+    abstract = abstract.replace("\n", " ").strip().replace("''", "'")
     global model, tokenizer
 
     if model is None:
-        # Try loading full-precision on GPU first
-        try:
-            model = AutoModelForSequenceClassification.from_pretrained(
-                model_path,
-                num_labels=1,
-                torch_dtype=torch.float32,
-                device_map="auto"
-            )
-        except RuntimeError:
-            # Fallback to CPU-only
-            model = AutoModelForSequenceClassification.from_pretrained(
-                model_path,
-                num_labels=1,
-                torch_dtype=torch.float32,
-                device_map="cpu"
-            )
+        # Load config and disable any quantization
+        config = AutoConfig.from_pretrained(model_path)
+        config.quantization_config = None
+
+        # Load model in full float32, then move to device
+        model = AutoModelForSequenceClassification.from_pretrained(
+            model_path,
+            config=config,
+            num_labels=1,
+            torch_dtype=torch.float32,
+            device_map=None,
+            low_cpu_mem_usage=False
+        )
+        model.to(device)
+
         tokenizer = AutoTokenizer.from_pretrained(model_path)
         model.eval()
-
-    prompt = (
+        
+    text = (
         f"Given a certain paper,\n"
-        f"Title: {title.strip()}\n"
-        f"Abstract: {abstract.strip()}\n"
+        f"Title: {title}\n"
+        f"Abstract: {abstract}\n"
         f"Predict its normalized academic impact (0~1):"
     )
-    inputs = tokenizer(prompt, return_tensors="pt", truncation=True, max_length=1024)
-    inputs = {k: v.to(device) for k, v in inputs.items()}
-    with torch.no_grad():
-        logits = model(**inputs).logits
-    prob = torch.sigmoid(logits).item()
-    score = min(1.0, prob + 0.05)  # +0.05 adjustment
-    return round(score, 4)
+
+    try:
+        inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=1024)
+        inputs = {k: v.to(device) for k, v in inputs.items()}
+        with torch.no_grad():
+            outputs = model(**inputs)
+        prob = torch.sigmoid(outputs.logits).item()
+        score = min(1.0, prob + 0.05)
+        return round(score, 4)
+    except Exception as e:
+        print(f"Prediction error: {e}")
+        return 0.0  # default on error
 
 def get_grade_and_emoji(score):
     if score >= 0.900: return "AAA 🌟"
@@ -90,32 +128,55 @@ def get_grade_and_emoji(score):
     if score >= 0.300: return "CC ✏️"
     return "C 📑"
 
+example_papers = [
+    {
+        "title": "Attention Is All You Need",
+        "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train.",
+        "score": 0.982,
+        "note": "💫 Revolutionary paper that introduced the Transformer architecture, fundamentally changing NLP and deep learning."
+    },
+    {
+        "title": "Language Models are Few-Shot Learners",
+        "abstract": "Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task. While typically task-agnostic in architecture, this method still requires task-specific fine-tuning datasets of thousands or tens of thousands of examples. By contrast, humans can generally perform a new language task from only a few examples or from simple instructions - something which current NLP systems still largely struggle to do. Here we show that scaling up language models greatly improves task-agnostic, few-shot performance, sometimes even reaching competitiveness with prior state-of-the-art fine-tuning approaches.",
+        "score": 0.956,
+        "note": "🚀 Groundbreaking GPT-3 paper that demonstrated the power of large language models."
+    },
+    {
+        "title": "An Empirical Study of Neural Network Training Protocols",
+        "abstract": "This paper presents a comparative analysis of different training protocols for neural networks across various architectures. We examine the effects of learning rate schedules, batch size selection, and optimization algorithms on model convergence and final performance. Our experiments span multiple datasets and model sizes, providing practical insights for deep learning practitioners.",
+        "score": 0.623,
+        "note": "📚 Solid research paper with useful findings but more limited scope and impact."
+    }
+]
+
 def validate_input(title, abstract):
-    """Ensure title ≥3 words, abstract ≥50 words, and ASCII-only."""
-    non_ascii = re.compile(r'[^\x00-\x7F]')
+    title = title.replace("\n", " ").strip().replace("''", "'")
+    abstract = abstract.replace("\n", " ").strip().replace("''", "'")
+    non_latin_pattern = re.compile(r'[^\u0000-\u007F]')
     if len(title.split()) < 3:
-        return False, "Title must be at least 3 words."
+        return False, "The title must be at least 3 words long."
     if len(abstract.split()) < 50:
-        return False, "Abstract must be at least 50 words."
-    if non_ascii.search(title):
-        return False, "Title contains non-ASCII characters."
-    if non_ascii.search(abstract):
-        return False, "Abstract contains non-ASCII characters."
-    return True, "Inputs look good."
+        return False, "The abstract must be at least 50 words long."
+    if non_latin_pattern.search(title):
+        return False, "The title contains invalid characters. Only English letters and symbols are allowed."
+    if non_latin_pattern.search(abstract):
+        return False, "The abstract contains invalid characters. Only English letters and symbols are allowed."
+    return True, "Inputs are valid!"
 
 def update_button_status(title, abstract):
-    valid, msg = validate_input(title, abstract)
+    valid, message = validate_input(title, abstract)
     if not valid:
-        return gr.update(value="Error: " + msg), gr.update(interactive=False)
-    return gr.update(value=msg), gr.update(interactive=True)
+        return gr.update(value="Error: " + message), gr.update(interactive=False)
+    return gr.update(value=message), gr.update(interactive=True)
 
 def process_arxiv_input(arxiv_input):
     if not arxiv_input.strip():
         return "", "", "Please enter an arXiv URL or ID"
-    res = fetch_arxiv_paper(arxiv_input)
-    if res["success"]:
-        return res["title"], res["abstract"], res["message"]
-    return "", "", res["message"]
+    result = fetch_arxiv_paper(arxiv_input)
+    if result["success"]:
+        return result["title"], result["abstract"], result["message"]
+    else:
+        return "", "", result["message"]
 
 css = """
 .gradio-container {
@@ -130,11 +191,17 @@ css = """
     -webkit-background-clip: text;
     -webkit-text-fill-color: transparent;
 }
+.sub-title {
+    text-align: center;
+    color: #4b5563;
+    font-size: 1.5rem !important;
+    margin-bottom: 2rem !important;
+}
 .input-section {
-    background: #ffffff;
+    background: white;
     padding: 2rem;
     border-radius: 1rem;
-    box-shadow: 0 4px 6px rgba(0,0,0,0.1);
+    box-shadow: 0 4px 6px -1px rgb(0 0 0 / 0.1);
 }
 .result-section {
     background: #f8fafc;
@@ -142,7 +209,13 @@ css = """
     border-radius: 1rem;
     margin-top: 2rem;
 }
-.methodology-section, .example-section {
+.methodology-section {
+    background: #ecfdf5;
+    padding: 2rem;
+    border-radius: 1rem;
+    margin-top: 2rem;
+}
+.example-section {
     background: #fff7ed;
     padding: 2rem;
     border-radius: 1rem;
@@ -154,134 +227,144 @@ css = """
     margin: 1rem 0;
 }
 .arxiv-input {
-    background: #f3f4f6;
+    margin-bottom: 1.5rem;
     padding: 1rem;
+    background: #f3f4f6;
     border-radius: 0.5rem;
-    margin-bottom: 1.5rem;
 }
 .arxiv-link {
     color: #2563eb;
     text-decoration: underline;
     font-size: 0.9em;
+    margin-top: 0.5em;
 }
 .arxiv-note {
-    color: #666666;
+    color: #666;
     font-size: 0.9em;
     margin-top: 0.5em;
     margin-bottom: 0.5em;
 }
 """
 
-example_papers = [
-    {
-        "title": "Attention Is All You Need",
-        "abstract": "The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train.",
-        "score": 0.982,
-        "note": "Revolutionary paper introducing the Transformer architecture."
-    },
-    {
-        "title": "Language Models are Few-Shot Learners",
-        "abstract": "Recent work has demonstrated substantial gains on many NLP tasks and benchmarks by pre-training on a large corpus of text followed by fine-tuning on a specific task. While typically task-agnostic in architecture, this method still requires task-specific fine-tuning datasets of thousands or tens of thousands of examples. By contrast, humans can generally perform a new language task from only a few examples or from simple instructions—something which current NLP systems still largely struggle to do. Here we show that scaling up language models greatly improves task-agnostic, few-shot performance, sometimes even reaching competitiveness with prior state-of-the-art fine-tuning approaches.",
-        "score": 0.956,
-        "note": "Groundbreaking GPT-3 paper on few-shot learning."
-    },
-    {
-        "title": "An Empirical Study of Neural Network Training Protocols",
-        "abstract": "This paper presents a comparative analysis of different training protocols for neural networks across various architectures. We examine the effects of learning rate schedules, batch size selection, and optimization algorithms on model convergence and final performance. Our experiments span multiple datasets and model sizes, providing practical insights for deep learning practitioners.",
-        "score": 0.623,
-        "note": "Solid empirical comparison of training protocols."
-    }
-]
-
 with gr.Blocks(theme=gr.themes.Default(), css=css) as iface:
-    gr.Markdown("<div class='main-title'>Papers Impact: AI-Powered Research Impact Predictor</div>")
-    gr.HTML("""
-<a href="https://visitorbadge.io/status?path=https%3A%2F%2FVIDraft-PaperImpact.hf.space">
-  <img src="https://api.visitorbadge.io/api/visitors?path=https%3A%2F%2FVIDraft-PaperImpact.hf.space&countColor=%23263759" />
-</a>
-""")
+    gr.Markdown(
+        """
+        # Papers Impact: AI-Powered Research Impact Predictor 
+        ## https://discord.gg/openfreeai
+        """
+    )
+    gr.HTML("""<a href="https://visitorbadge.io/status?path=https%3A%2F%2FVIDraft-PaperImpact.hf.space">
+<img src="https://api.visitorbadge.io/api/visitors?path=https%3A%2F%2FVIDraft-PaperImpact.hf.space&countColor=%23263759" />
+</a>""")
 
     with gr.Row():
         with gr.Column(elem_classes="input-section"):
             with gr.Group(elem_classes="arxiv-input"):
-                gr.Markdown("### Import from arXiv")
+                gr.Markdown("### 📑 Import from arXiv")
                 arxiv_input = gr.Textbox(
                     lines=1,
-                    placeholder="e.g. 2504.11651",
-                    label="arXiv URL or ID",
+                    placeholder="Enter arXiv URL or ID (e.g., 2504.11651)",
+                    label="arXiv Paper URL/ID",
                     value="2504.11651"
                 )
                 gr.Markdown("""
                 <p class="arxiv-note">
-                  Click to use the default example or visit <a href="https://arxiv.org" class="arxiv-link" target="_blank">arxiv.org</a>
+                Click input field to use example paper or browse papers at 
+                <a href="https://arxiv.org" target="_blank" class="arxiv-link">arxiv.org</a>
                 </p>
                 """)
                 fetch_button = gr.Button("🔍 Fetch Paper Details", variant="secondary")
-
-            gr.Markdown("### Or Enter Paper Details Manually")
+            
+            gr.Markdown("### 📝 Or Enter Paper Details Manually")
+            
             title_input = gr.Textbox(
                 lines=2,
-                placeholder="Enter paper title (minimum 3 words)...",
+                placeholder="Enter Paper Title (minimum 3 words)...",
                 label="Paper Title"
             )
             abstract_input = gr.Textbox(
                 lines=5,
-                placeholder="Enter paper abstract (minimum 50 words)...",
+                placeholder="Enter Paper Abstract (minimum 50 words)...",
                 label="Paper Abstract"
             )
-            validation_status = gr.Textbox(label="Validation Status", interactive=False)
+            validation_status = gr.Textbox(label="✔️ Validation Status", interactive=False)
             submit_button = gr.Button("🎯 Predict Impact", interactive=False, variant="primary")
-
+        
         with gr.Column(elem_classes="result-section"):
-            score_output = gr.Number(label="Impact Score")
-            grade_output = gr.Textbox(label="Grade", elem_classes="grade-display")
+            with gr.Group():
+                score_output = gr.Number(label="🎯 Impact Score")
+                grade_output = gr.Textbox(label="🏆 Grade", value="", elem_classes="grade-display")
 
     with gr.Row(elem_classes="methodology-section"):
-        gr.Markdown("""
-### Scientific Methodology
-- **Training Data**: Papers from CS.CV, CS.CL (NLP), and CS.AI fields  
-- **Optimization**: NDCG optimization with Sigmoid activation & MSE loss  
-- **Validation**: Cross-validated on historical citation data  
-- **Architecture**: Transformer-based text encoder  
-- **Metrics**: Citation-pattern analysis & research influence
-        """)
+        gr.Markdown(
+            """
+            ### 🔬 Scientific Methodology
+            - **Training Data**: Model trained on extensive dataset of published papers from CS.CV, CS.CL(NLP), and CS.AI fields
+            - **Optimization**: NDCG optimization with Sigmoid activation and MSE loss function
+            - **Validation**: Cross-validated against historical paper impact data
+            - **Architecture**: Advanced transformer-based deep textual analysis
+            - **Metrics**: Quantitative analysis of citation patterns and research influence
+            """
+        )
 
-    gr.Markdown("""
-### Rating Scale
-| Grade | Score Range | Description         | Emoji |
-|-------|-------------|---------------------|-------|
-| AAA   | 0.900–1.000 | Exceptional Impact  | 🌟    |
-| AA    | 0.800–0.899 | Very High Impact    | ⭐    |
-| A     | 0.650–0.799 | High Impact         | ✨    |
-| BBB   | 0.600–0.649 | Above Average       | 🔵    |
-| BB    | 0.550–0.599 | Moderate Impact     | 📘    |
-| B     | 0.500–0.549 | Average Impact      | 📖    |
-| CCC   | 0.400–0.499 | Below Average       | 📝    |
-| CC    | 0.300–0.399 | Low Impact          | ✏️    |
-| C     | <0.300      | Limited Impact      | 📑    |
-    """)
+    with gr.Row():
+        gr.Markdown(
+            """
+            ### 📊 Rating Scale
+            | Grade | Score Range | Description | Indicator |
+            |-------|-------------|-------------|-----------|
+            | AAA | 0.900-1.000 | Exceptional Impact | 🌟 |
+            | AA | 0.800-0.899 | Very High Impact | ⭐ |
+            | A | 0.650-0.799 | High Impact | ✨ |
+            | BBB | 0.600-0.649 | Above Average Impact | 🔵 |
+            | BB | 0.550-0.599 | Moderate Impact | 📘 |
+            | B | 0.500-0.549 | Average Impact | 📖 |
+            | CCC | 0.400-0.499 | Below Average Impact | 📝 |
+            | CC | 0.300-0.399 | Low Impact | ✏️ |
+            | C | < 0.299 | Limited Impact | 📑 |
+            """
+        )
 
     with gr.Row(elem_classes="example-section"):
-        gr.Markdown("### Example Papers")
+        gr.Markdown("### 📋 Example Papers")
         for paper in example_papers:
-            gr.Markdown(f"""
-#### {paper['title']}
-**Score**: {paper['score']} | **Grade**: {get_grade_and_emoji(paper['score'])}  
-{paper['abstract']}  
-*{paper['note']}*  
----
-            """)
+            gr.Markdown(
+                f"""
+                #### {paper['title']} 
+                **Score**: {paper.get('score', 'N/A')} | **Grade**: {get_grade_and_emoji(paper.get('score', 0))}
+                {paper['abstract']}
+                *{paper['note']}*
+                ---
+                """
+            )
 
-    # Event handlers
-    title_input.change(update_button_status, [title_input, abstract_input], [validation_status, submit_button])
-    abstract_input.change(update_button_status, [title_input, abstract_input], [validation_status, submit_button])
-    fetch_button.click(process_arxiv_input, [arxiv_input], [title_input, abstract_input, validation_status])
+    title_input.change(
+        update_button_status,
+        inputs=[title_input, abstract_input],
+        outputs=[validation_status, submit_button]
+    )
+    abstract_input.change(
+        update_button_status,
+        inputs=[title_input, abstract_input],
+        outputs=[validation_status, submit_button]
+    )
+    
+    fetch_button.click(
+        process_arxiv_input,
+        inputs=[arxiv_input],
+        outputs=[title_input, abstract_input, validation_status]
+    )
 
-    def run_prediction(t, a):
-        s = predict(t, a)
-        return s, get_grade_and_emoji(s)
+    def process_prediction(title, abstract):
+        score = predict(title, abstract)
+        grade = get_grade_and_emoji(score)
+        return score, grade
 
-    submit_button.click(run_prediction, [title_input, abstract_input], [score_output, grade_output])
+    submit_button.click(
+        process_prediction,
+        inputs=[title_input, abstract_input],
+        outputs=[score_output, grade_output]
+    )
 
 if __name__ == "__main__":
     iface.launch()