Create app.py

app.py

@@ -0,0 +1,152 @@
+import gradio as gr
+import pandas as pd
+import numpy as np
+from sklearn.model_selection import train_test_split
+from sklearn.linear_model import LogisticRegression
+from datasets import load_dataset
+import random
+
+print("📦 Phase 1: Loading Pre-calculated Assets...")
+df_cached = pd.read_csv("emodb_full_zeroshot_predictions.csv")
+X_embeddings = np.load("emodb_full_embeddings.npy")
+
+print("🧠 Phase 2: Dynamically Training Both Linear Classification Heads...")
+# --- Head A: The Global 80/20 Head ---
+labels = df_cached['True_Emotion'].values
+indices = np.arange(len(labels))
+X_train, X_test, y_train, y_test, idx_train, idx_test = train_test_split(
+    X_embeddings, labels, indices, test_size=0.20, random_state=42, stratify=labels
+)
+global_head = LogisticRegression(max_iter=1000, class_weight='balanced', random_state=42)
+global_head.fit(X_train, y_train)
+
+# --- Head B: The Cross-Speaker Head (Train on Speaker 31 & 34) ---
+train_speakers = ['Speaker_31.0', 'Speaker_34.0']
+cross_train_mask = df_cached['Speaker_Info'].isin(train_speakers)
+X_train_cross = X_embeddings[cross_train_mask]
+y_train_cross = df_cached[cross_train_mask]['True_Emotion'].values
+
+cross_head = LogisticRegression(max_iter=1000, class_weight='balanced', random_state=42)
+cross_head.fit(X_train_cross, y_train_cross)
+
+print("🌍 Phase 3: Attaching to EmoDB on Hugging Face Hub for Audio Streaming...")
+hf_dataset = load_dataset("harritaylor/er_emodb", split="train")
+
+# --- UI Functions ---
+def process_sample(index):
+    idx = int(index)
+    sample = hf_dataset[idx]
+    sr = sample['audio']['sampling_rate']
+    audio_array = sample['audio']['array']
+    row = df_cached.iloc[idx]
+    
+    # Run Inference on the vector using both heads
+    vector = X_embeddings[idx].reshape(1, -1)
+    global_pred = global_head.predict(vector)[0]
+    cross_pred = cross_head.predict(vector)[0]
+    
+    # Calculate Probabilities for visual feedback
+    probs = global_head.predict_proba(vector)[0]
+    prob_dict = {global_head.classes_[i]: float(probs[i]) for i in range(len(probs))}
+    
+    return (
+        (sr, audio_array),
+        row['Speaker_Info'],
+        row['True_Emotion'],
+        row['Model_Prediction'],
+        global_pred,
+        cross_pred,
+        prob_dict
+    )
+
+def pick_random_index():
+    return random.randint(0, 534)
+
+# --- GRADIO INTERFACE ---
+with gr.Blocks(theme=gr.themes.Default(primary_hue="orange", secondary_hue="neutral")) as demo:
+    gr.Markdown("# 🚀 Audio LLM Hidden Space Decoder")
+    gr.Markdown("### Evidence that an Audio LLM's internal mathematical representations vastly outshine its text outputs.")
+    
+    with gr.Tabs():
+        # TAB 1: INTERACTIVE EXPLORER
+        with gr.TabItem("📊 Dataset Explorer & Evaluation"):
+            gr.Markdown("Pick an index manually or hit 'Pick Random Sample' to stream audio directly from EmoDB and compare all three execution layers.")
+            
+            with gr.Row():
+                index_slider = gr.Slider(minimum=0, maximum=534, step=1, value=0, label="Select Audio Index")
+                random_btn = gr.Button("🎲 Pick Random Sample", variant="secondary")
+                analyze_btn = gr.Button("⚡ Analyze Sample", variant="primary")
+                
+            with gr.Row():
+                audio_player = gr.Audio(label="Audio Playback (Streamed from HF)")
+                speaker_out = gr.Textbox(label="Speaker ID")
+                true_out = gr.Textbox(label="Ground Truth (Human Label)")
+                
+            with gr.Row():
+                zs_out = gr.Textbox(label="1. Zero-Shot Text Generation Prediction")
+                global_out = gr.Textbox(label="2. Global 80/20 Embedding Head Prediction")
+                cross_out = gr.Textbox(label="3. Leave-Speakers-Out Head Prediction")
+                
+            confidence_chart = gr.Label(label="Global Embedding Classifier Class Probabilities")
+            
+            # Button mappings
+            analyze_btn.click(
+                process_sample, 
+                inputs=[index_slider], 
+                outputs=[audio_player, speaker_out, true_out, zs_out, global_out, cross_out, confidence_chart]
+            )
+            random_btn.click(pick_random_index, outputs=[index_slider]).then(
+                process_sample, 
+                inputs=[index_slider], 
+                outputs=[audio_player, speaker_out, true_out, zs_out, global_out, cross_out, confidence_chart]
+            )
+
+        # TAB 2: TECHNICAL REPORT & STATS
+        with gr.TabItem("📜 Methodological Report & Statistics"):
+            gr.Markdown("""
+            ## Decoding Audio LLM Hidden Spaces
+            **An Empirical Comparison of Information Extraction Methods on Qwen2-Audio-7B-Instruct**
+            
+            ### 📌 Executive Summary
+            When forcing a large multimodal model to output speech interpretations as text tokens, a massive **information bottleneck** occurs. This dashboard showcases that extracting the raw mathematical embeddings hidden behind the model's text decoder unlocks an entirely new layer of granular acoustic intelligence.
+            
+            ### 📊 Comparative Performance Summary
+            """)
+            
+            # Main comparison table
+            gr.Markdown("""
+            | Evaluation Architecture | Test Method | Dataset Coverage | Accuracy |
+            | :--- | :--- | :--- | :--- |
+            | **Zero-Shot Text Prompting** | Direct Generation | Full Dataset (535 files) | **67.3%** |
+            | **Linear Embedding Classifier** | Stratified 80/20 Split | Unseen 20% Subset | **97.2%** |
+            | **Linear Embedding Classifier** | Cross-Speaker Generalization | 6 Unseen Speakers (Blind) | **92.2%** |
+            """)
+            
+            gr.Markdown("""
+            ### 🌍 Cross-Speaker Generalization Breakdown
+            To determine if the internal representation generalizes across unique human vocal anatomies, accents, and pitches, we trained a linear classifier **strictly on 2 speakers** (Speaker 31 and 34) and evaluated blindly on the remaining **6 unseen speakers**. 
+            
+            The results confirm a highly robust, universal acoustic map:
+            """)
+            
+            # Speaker breakdown table
+            gr.Markdown("""
+            | Unseen Test Speaker ID | Extracted Audio Samples | Downstream Classification Accuracy |
+            | :--- | :--- | :--- |
+            | **Speaker_21.0** | 43 samples | **88.4%** |
+            | **Speaker_32.0** | 99 samples | **91.9%** |
+            | **Speaker_26.0** | 55 samples | **85.5%** |
+            | **Speaker_30.0** | 35 samples | **91.4%** |
+            | **Speaker_35.0** | 69 samples | **97.1%** |
+            | **Speaker_25.0** | 56 samples | **96.4%** |
+            | **COMBINED BLIND AVERAGE** | **357 samples** | **92.2%** |
+            """)
+            
+            gr.Markdown("""
+            ### 🔑 Primary Insights & Observations
+            1. **The Linear Advantage:** Complex non-linear architectures (XGBoost, Random Forests) easily fall prey to overfitting due to high dimensionality ($4096\\text{D}$) and low sample sizes. Simple `LogisticRegression` bounds generalize beautifully.
+            2. **Acoustic Edge Cases:** Misclassifications are bounded tightly by the physics of sound. The embedding head's rare failures occur strictly between acoustic "twins" like *Boredom/Neutral* (shared low-energy profiles) or *Anger/Fear* (shared high-energy profiles). 
+            3. **The Synergistic Save:** In rare instances where raw audio signals blur high-arousal acoustics, the textual deep reasoning layers of Qwen occasionally navigate structural nuances to succeed where raw vectors misalign.
+            """)
+
+demo.launch()