app.py

import gradio as gr
import random
import json
import re
import os
import numpy as np
from collections import Counter
from sklearn.feature_extraction.text import TfidfVectorizer
import functools
from concurrent.futures import ThreadPoolExecutor
import threading
import nltk
from nltk.corpus import wordnet
from nltk.stem import WordNetLemmatizer

# Add at the beginning of your script, after imports
import os
import nltk

# Get the current directory
current_dir = os.getcwd()
print(f"Current directory: {current_dir}")

# Point NLTK to the data directories in your current directory
nltk_data_path = os.path.join(current_dir, "nltk_data")
print(f"Setting NLTK data path to: {nltk_data_path}")

# Add the path to NLTK's search paths
nltk.data.path.insert(0, nltk_data_path)  # Insert at position 0 to search here first

# Print all paths for debugging
print(f"NLTK will search in: {nltk.data.path}")

# Try to load the taggers from your local directory
try:
    # Try to directly load the tagger model
    from nltk.tag.perceptron import PerceptronTagger
    tagger = PerceptronTagger()
    print("Successfully loaded PerceptronTagger")
except Exception as e:
    print(f"Error loading tagger: {e}")

# nltk.download('averaged_perceptron_tagger_eng')

# Add the header constant at the top of your file
WEBSITE = ("""<div class="embed_hidden" style="text-align: center;">
    <h1>SINC: Spatial Composition of 3D Human Motions for Simultaneous Action Generation</h1>
    <h2 style="margin: 1em 0; font-size: 2em;">
        <span style="font-weight: normal; font-style: italic;">ICCV 2023</span>
    </h2>
    <h3>
        <a href="https://atnikos.github.io/" target="_blank" rel="noopener noreferrer">Nikos Athanasiou</a><sup>*1</sup>,
        <a href="https://mathis.petrovich.fr/" target="_blank" rel="noopener noreferrer">Mathis Petrovich</a><sup>*1,2</sup>,
        <br>
        <a href="https://ps.is.mpg.de/person/black" target="_blank" rel="noopener noreferrer">Michael J. Black</a><sup>1</sup>,
        <a href="https://gulvarol.github.io/" target="_blank" rel="noopener noreferrer">Gül Varol</a><sup>2</sup>
    </h3>
    <h3>
        <sup>1</sup>MPI for Intelligent Systems, Tübingen, Germany<br>
        <sup>2</sup>LIGM, École des Ponts, Univ Gustave Eiffel, CNRS, France
    </h3>
</div>
<div style="display:flex; gap: 0.3rem; justify-content: center; align-items: center;" align="center">
<a href='https://arxiv.org/abs/2304.10417'><img src='https://img.shields.io/badge/Arxiv-2304.10417-A42C25?style=flat&logo=arXiv&logoColor=A42C25'></a>
<a href='https://sinc.is.tue.mpg.de'><img src='https://img.shields.io/badge/Project-Page-%23df5b46?style=flat&logo=Google%20chrome&logoColor=%23df5b46'></a>
</div>
    <h2 align="center">
        Download 
        <a href="https://drive.google.com/drive/folders/1ks9wvNN_arrgBcd0GxN5nRLf5ASPkUgc?usp=sharing" target="_blank" rel="noopener noreferrer"> SINC synthetic data</a>,
        if you want to train your models with spatial composition from AMASS. 
        <br>
        The data you are exploring in this demo are 
        the data created using the 
        code <a href='https://github.com/atnikos/sinc/blob/main/create_synthetic_babel.py' target="_blank" rel="noopener noreferrer">to compose motions from AMASS in our repo.</a><sup>**</sup>
    </h2>
""")

# Action examples
ACTION_EXAMPLES = [
    "walk forward on balance beam", "walk counterclockwise", "sit on chair", "kick a ball", "jump up",
    "hold on to rail with right hand", "pick up an object", 'wave with the right hand', 'throw a ball', 'bow'
]
ACTION_EXAMPLES_SIMULTANEOUS = [
   "walk forward on balance beam while holding rail with right hand",
   "walk counterclockwise while waving with left hand", 
   "sit on chair and wave with left hand",
   "pick up an object while bowing",
   "walk forward on balance beam while waving left hand"
]

# Global cache for expensive operations
SIMILARITY_CACHE = {}
SEARCH_RESULTS_CACHE = {}
GPT_SIMILARITY_CACHE = {}
GPT_SEARCH_RESULTS_CACHE = {}
SYNONYM_CACHE = {}
MAX_WORKERS = 4  # For ThreadPoolExecutor

# Cache for TF-IDF
TFIDF_VECTORIZER = None
TFIDF_MATRIX = None
MOTION_TEXTS = []
MOTION_KEYS = []
GPT_TEXTS = []
GPT_KEYS = []

# Initialize lemmatizer
lemmatizer = WordNetLemmatizer()

# Movement action word mappings - manually defined synonyms for common motion words
ACTION_SYNONYMS = {
    'walk': ['move', 'stroll', 'pace', 'stride', 'wander', 'stalk', 'amble', 'saunter', 'tread', 'step'],
    'run': ['sprint', 'jog', 'dash', 'race', 'bolt', 'scamper', 'rush', 'hurry'],
    'jump': ['leap', 'hop', 'spring', 'bounce', 'vault', 'bound', 'skip'],
    'turn': ['rotate', 'spin', 'twist', 'revolve', 'pivot', 'swivel', 'whirl'],
    'wave': ['signal', 'gesture', 'flap', 'flutter', 'waggle', 'shake', 'brandish'],
    'sit': ['perch', 'recline', 'rest', 'squat'],
    'stand': ['rise', 'upright', 'erect', 'vertical'],
    'throw': ['toss', 'hurl', 'fling', 'chuck', 'lob', 'pitch', 'cast'],
    'grab': ['grasp', 'clutch', 'seize', 'grip', 'hold', 'take', 'catch'],
    'pick': ['lift', 'raise', 'hoist', 'elevate'],
    'kick': ['boot', 'punt', 'strike'],
    'bow': ['bend', 'stoop', 'incline', 'nod'],
    'dance': ['twirl', 'sway', 'shimmy', 'boogie', 'groove', 'swing'],
    'balance': ['steady', 'stabilize', 'poise', 'equilibrium'],
    'forward': ['ahead', 'onward', 'frontward', 'forth'],
    'backward': ['back', 'rearward', 'reverse', 'retreat'],
    'clockwise': ['right', 'rightward', 'rightways'],
    'counterclockwise': ['left', 'leftward', 'leftways', 'anticlockwise'],
    'hold': ['grip', 'grasp', 'clutch', 'clasp', 'clench', 'possess']
}

# Build reverse mapping for faster lookups
REVERSE_SYNONYMS = {}
for word, synonyms in ACTION_SYNONYMS.items():
    REVERSE_SYNONYMS[word] = word  # A word is its own synonym
    for synonym in synonyms:
        REVERSE_SYNONYMS[synonym] = word

def get_wordnet_pos(word):
    """Map POS tag to first character used by WordNet lemmatizer
       with fallback for errors"""
    try:
        tag = nltk.tag.pos_tag([word])[0][1][0].upper()
        tag_dict = {"J": wordnet.ADJ,
                    "N": wordnet.NOUN,
                    "V": wordnet.VERB,
                    "R": wordnet.ADV}
        return tag_dict.get(tag, wordnet.NOUN)
    except Exception as e:
        print(f"POS tagging error for word '{word}': {e}")
        # Default to NOUN as fallback
        return wordnet.NOUN

def get_synonyms(word):
    """Get all synonyms for a word using WordNet and our custom action mappings"""
    if word in SYNONYM_CACHE:
        return SYNONYM_CACHE[word]

    synonyms = set()
    
    # Add the word itself
    synonyms.add(word)
    
    # Check our custom action mappings first (faster and more domain-specific)
    if word in REVERSE_SYNONYMS:
        canonical_word = REVERSE_SYNONYMS[word]
        synonyms.add(canonical_word)
        synonyms.update(ACTION_SYNONYMS.get(canonical_word, []))
    
    # Then check WordNet (more general but can be noisy)
    try:
        word_lemma = lemmatizer.lemmatize(word, get_wordnet_pos(word))
        for syn in wordnet.synsets(word_lemma):
            for lemma in syn.lemmas():
                synonyms.add(lemma.name().lower().replace('_', ' '))
    except Exception as e:
        print(f"Error getting WordNet synonyms for '{word}': {e}")
    
    SYNONYM_CACHE[word] = synonyms
    return synonyms

def expand_query_with_synonyms(query):
    """Expand a query with synonyms for each term"""
    try:
        words = nltk.word_tokenize(query.lower())
    except Exception as e:
        print(f"Tokenization error: {e}")
        # Fallback to simple split if tokenization fails
        words = query.lower().split()
        
    expanded_terms = []
    
    for word in words:
        if len(word) > 2:  # Only expand words with length > 2 to avoid stop words
            synonyms = get_synonyms(word)
            expanded_terms.extend(synonyms)
        else:
            expanded_terms.append(word)
    
    # Join back into a space-separated string
    return ' '.join(expanded_terms)

def create_example_buttons(textbox, loftexts):
    """Creates clickable buttons for example actions"""
    return gr.Examples(
        examples=loftexts,
        inputs=textbox,
        label="Example Actions"
    )

# Load motion data
def load_json_dict(file_path):
    with open(file_path, "r") as f:
        return json.load(f)

# Load data at startup
print("Loading motion data...")
motion_dict = load_json_dict("for_website_v4.json")
motion_dict = {
    key: value for key, value in motion_dict.items()
    if "guide forward walk" not in value['source_annot'].lower()
    and "guide forward walk" not in value['target_annot'].lower()
}

print("Loading GPT labels...")
GPT_LABELS_LIST = load_json_dict('gpt3-labels-list.json')
GPT_LABELS_LIST = {k: v[2] for k, v in GPT_LABELS_LIST.items()}

# TF-IDF based similarity implementation with synonym expansion
def initialize_tfidf():
    """Initialize TF-IDF vectorizer and precompute matrices"""
    global TFIDF_VECTORIZER, TFIDF_MATRIX, MOTION_TEXTS, MOTION_KEYS
    
    print("Initializing TF-IDF vectorizer...")
    
    # Extract text descriptions from the motion dictionary for TF-IDF
    MOTION_TEXTS = []
    MOTION_KEYS = []
    
    for key, motion in motion_dict.items():
        # Combine source and target annotations
        text = f"{motion['source_annot']} {motion['target_annot']}".lower()
        MOTION_TEXTS.append(text)
        MOTION_KEYS.append(key)
    
    # Initialize the TF-IDF vectorizer
    TFIDF_VECTORIZER = TfidfVectorizer(
        lowercase=True, 
        stop_words='english',
        ngram_range=(1, 2),  # Include bigrams for better matching
        max_features=20000,  # Increased to accommodate synonym expansions
        min_df=1             # Lower threshold to catch less frequent terms
    )
    
    # Fit and transform to get TF-IDF vectors
    TFIDF_MATRIX = TFIDF_VECTORIZER.fit_transform(MOTION_TEXTS)
    
    print(f"TF-IDF matrix created with shape {TFIDF_MATRIX.shape}")
    
    # Also create GPT labels matrix
    initialize_gpt_tfidf()

def initialize_gpt_tfidf():
    """Initialize TF-IDF for GPT labels"""
    global GPT_TEXTS, GPT_KEYS
    
    print("Initializing TF-IDF for GPT labels...")
    
    GPT_TEXTS = []
    GPT_KEYS = []
    
    for key, text in GPT_LABELS_LIST.items():
        GPT_TEXTS.append(text.lower())
        GPT_KEYS.append(key)

def compute_tfidf_similarity(query, top_k=10):
    """Compute similarity using TF-IDF vectors with synonym expansion"""
    global TFIDF_VECTORIZER, TFIDF_MATRIX, MOTION_TEXTS, MOTION_KEYS
    
    # Original query for cache key
    original_query = query.lower().strip()
    
    # Check cache first
    cache_key = f"tfidf_{original_query}_{top_k}"
    if cache_key in SIMILARITY_CACHE:
        return SIMILARITY_CACHE[cache_key]
    
    try:
        # Expand query with synonyms
        expanded_query = expand_query_with_synonyms(original_query)
        
        # Transform query to TF-IDF space
        query_vector = TFIDF_VECTORIZER.transform([expanded_query])
        
        # Compute cosine similarity between query and all texts
        # Using matrix multiplication for sparse matrices
        similarities = (query_vector @ TFIDF_MATRIX.T).toarray().flatten()
        
        # Get indices of top_k highest similarity scores
        top_indices = np.argsort(similarities)[-top_k:][::-1]
        
        # Get the corresponding entries and scores
        top_entries = [motion_dict[MOTION_KEYS[idx]] for idx in top_indices]
        top_scores = [similarities[idx] for idx in top_indices]
        
        result = (top_entries, top_scores)
    except Exception as e:
        print(f"Error in TF-IDF similarity computation: {e}")
        # Fallback to random motions if TF-IDF fails
        result = (get_random_motions(top_k), ['NA']*top_k)
    
    SIMILARITY_CACHE[cache_key] = result
    return result

def compute_gpt_tfidf_similarity(query):
    """Compute similarity between query and GPT labels using TF-IDF with synonym expansion"""
    global TFIDF_VECTORIZER, GPT_TEXTS, GPT_KEYS
    
    # Original query for cache key
    original_query = query.lower().strip()
    
    # Check cache first
    cache_key = f"gpt_tfidf_{original_query}"
    if cache_key in GPT_SIMILARITY_CACHE:
        return GPT_SIMILARITY_CACHE[cache_key]
    
    try:
        # Expand query with synonyms
        expanded_query = expand_query_with_synonyms(original_query)
        
        # Transform query and all GPT texts to TF-IDF space
        query_vector = TFIDF_VECTORIZER.transform([expanded_query])
        gpt_vectors = TFIDF_VECTORIZER.transform(GPT_TEXTS)
        
        # Compute cosine similarity between query and all GPT texts
        similarities = (query_vector @ gpt_vectors.T).toarray().flatten()
        
        # Get the index of highest similarity score
        best_idx = np.argmax(similarities)
        best_key = GPT_KEYS[best_idx]
        best_text = GPT_LABELS_LIST[best_key]
        best_sim = similarities[best_idx]
        
        result = (best_key, best_text, best_sim)
    except Exception as e:
        print(f"Error in GPT TF-IDF similarity computation: {e}")
        # Fallback to first GPT label if computation fails
        if GPT_KEYS:
            result = (GPT_KEYS[0], GPT_LABELS_LIST[GPT_KEYS[0]], 0.5)
        else:
            result = (None, None, 0)
    
    GPT_SIMILARITY_CACHE[cache_key] = result
    return result

# Precompile regex pattern
WORD_PATTERN = re.compile(r'\b\w+\b')

# Cache the word lists to avoid repeated tokenization
SOURCE_WORDS_CACHE = {}
TARGET_WORDS_CACHE = {}

def get_words(text):
    """Tokenize text and cache the results"""
    if text in SOURCE_WORDS_CACHE:
        return SOURCE_WORDS_CACHE[text]
    words = set(WORD_PATTERN.findall(text.lower()))
    SOURCE_WORDS_CACHE[text] = words
    return words

def exact_string_search(action1, action2):
    """Search for exact string matches first"""
    exact_results = []
    
    action1_lower = action1.lower().strip()
    action2_lower = action2.lower().strip()
    
    for k, v in motion_dict.items():
        source_lower = v["source_annot"].lower()
        target_lower = v["target_annot"].lower()
        
        # Check for exact matches in either annotation
        cond1 = action1_lower in source_lower or action1_lower in target_lower
        cond2 = action2_lower in source_lower or action2_lower in target_lower
        
        if cond1 and cond2:
            exact_results.append(v)
    
    return exact_results

def search_motions_two_actions(action1, action2):
    """Enhanced substring search with synonym expansion"""
    # Create a cache key for this query
    cache_key = f"{action1.lower().strip()}_{action2.lower().strip()}"
    
    # Check if we already have results for this query
    if cache_key in SEARCH_RESULTS_CACHE:
        return SEARCH_RESULTS_CACHE[cache_key]
    
    try:
        # Convert actions into lists of words
        action1_words = set(action1.lower().strip().split())
        action2_words = set(action2.lower().strip().split())
        
        # Expand with synonyms
        expanded_action1_words = set()
        for word in action1_words:
            if len(word) > 2:  # Only consider words longer than 2 chars
                expanded_action1_words.update(get_synonyms(word))
            else:
                expanded_action1_words.add(word)
        
        expanded_action2_words = set()
        for word in action2_words:
            if len(word) > 2:  # Only consider words longer than 2 chars
                expanded_action2_words.update(get_synonyms(word))
            else:
                expanded_action2_words.add(word)
        
        results = []
        for k, v in motion_dict.items():
            # Get or compute tokenized words from cache
            if v["source_annot"] not in SOURCE_WORDS_CACHE:
                SOURCE_WORDS_CACHE[v["source_annot"]] = set(WORD_PATTERN.findall(v["source_annot"].lower()))
            
            if v["target_annot"] not in TARGET_WORDS_CACHE:
                TARGET_WORDS_CACHE[v["target_annot"]] = set(WORD_PATTERN.findall(v["target_annot"].lower()))
            
            source_words = SOURCE_WORDS_CACHE[v["source_annot"]]
            target_words = TARGET_WORDS_CACHE[v["target_annot"]]
            
            # For each word in action1, check if any of its synonyms match
            cond1 = False
            if action1_words:  # Only check if action1 has words
                matches = 0
                for word in action1_words:
                    word_matches = False
                    if len(word) <= 2:  # For short words, just check exact match
                        if word in source_words or word in target_words:
                            word_matches = True
                    else:  # For longer words, check all synonyms
                        for syn in get_synonyms(word):
                            if syn in source_words or syn in target_words:
                                word_matches = True
                                break
                    if word_matches:
                        matches += 1
                # Consider a match if at least 70% of words (or their synonyms) are found
                cond1 = (matches / len(action1_words)) >= 0.7 if action1_words else True
            else:
                cond1 = True
                
            # For each word in action2, check if any of its synonyms match
            cond2 = False
            if action2_words:  # Only check if action2 has words
                matches = 0
                for word in action2_words:
                    word_matches = False
                    if len(word) <= 2:  # For short words, just check exact match
                        if word in source_words or word in target_words:
                            word_matches = True
                    else:  # For longer words, check all synonyms
                        for syn in get_synonyms(word):
                            if syn in source_words or syn in target_words:
                                word_matches = True
                                break
                    if word_matches:
                        matches += 1
                # Consider a match if at least 70% of words (or their synonyms) are found
                cond2 = (matches / len(action2_words)) >= 0.7 if action2_words else True
            else:
                cond2 = True

            if cond1 and cond2:
                results.append(v)
    except Exception as e:
        print(f"Error in substring search: {e}")
        results = []
    
    # Cache the results
    SEARCH_RESULTS_CACHE[cache_key] = results
    
    return results

def search_motions_semantic(action1, action2, top_k=10):
    """Semantic search using TF-IDF similarity with synonym expansion"""
    query_text = (action1.strip() + " " + action2.strip()).strip().lower()
    if not query_text:
        return [], []
    
    # Check cache first
    cache_key = f"{query_text}_{top_k}"
    if cache_key in SEARCH_RESULTS_CACHE:
        return SEARCH_RESULTS_CACHE[cache_key]
    
    # Use TF-IDF similarity
    return compute_tfidf_similarity(query_text, top_k)

def get_random_motions(n_motions):
    all_vals = list(motion_dict.values())
    return random.sample(all_vals, min(n_motions, len(all_vals)))

def search_gpt_semantic(action, top_k=1):
    """Search GPT labels using TF-IDF similarity with synonym expansion"""
    query_text = action.strip().lower()
    if not query_text:
        return None, None, None
    
    # Check cache first
    if query_text in GPT_SEARCH_RESULTS_CACHE:
        return GPT_SEARCH_RESULTS_CACHE[query_text]
    
    # Use TF-IDF similarity for GPT labels
    result = compute_gpt_tfidf_similarity(query_text)
    GPT_SEARCH_RESULTS_CACHE[query_text] = result
    
    return result

def search_motions_combined(action1, action2, n_motions):
    """Improved combined search approach that prioritizes exact matches"""
    # Create a cache key for this query
    cache_key = f"{action1.lower().strip()}_{action2.lower().strip()}_{n_motions}"
    
    # Check if we already have results for this query
    if cache_key in SEARCH_RESULTS_CACHE:
        return SEARCH_RESULTS_CACHE[cache_key]
    
    # 1. First try exact string matches
    exact_results = exact_string_search(action1, action2)
    
    if len(exact_results) >= n_motions:
        # If we have enough exact matches, return them
        result = (random.sample(exact_results, n_motions), ['EXACT']*n_motions)
        SEARCH_RESULTS_CACHE[cache_key] = result
        return result
    
    # 2. If not enough exact matches, try the enhanced substring search with synonyms
    string_results = search_motions_two_actions(action1, action2)
    
    # Filter out any results that are already in exact_results
    string_results = [r for r in string_results if r not in exact_results]
    
    # Combine exact_results with string_results
    combined_results = list(exact_results)
    combined_scores = ['EXACT'] * len(exact_results)
    
    if len(combined_results) + len(string_results) >= n_motions:
        # If we have enough combined results, use them
        needed = n_motions - len(combined_results)
        if needed > 0:
            combined_results.extend(random.sample(string_results, needed))
            combined_scores.extend(['SUBSTR'] * needed)
        
        result = (combined_results[:n_motions], combined_scores[:n_motions])
    else:
        # 3. If still not enough, add all substring matches and then use semantic search
        combined_results.extend(string_results)
        combined_scores.extend(['SUBSTR'] * len(string_results))
        
        # Use semantic search for the remaining needed motions
        needed = n_motions - len(combined_results)
        if needed > 0:
            sem_list, sem_score_list = search_motions_semantic(action1, action2, top_k=2*needed)
            
            # Filter out duplicates
            used_combo = {m["motion_combo"] for m in combined_results}
            
            for item, score in zip(sem_list, sem_score_list):
                if item["motion_combo"] not in used_combo:
                    combined_results.append(item)
                    combined_scores.append(score)
                    used_combo.add(item["motion_combo"])
                if len(combined_results) == n_motions:
                    break
            
            # Still short? Fill with random
            if len(combined_results) < n_motions:
                needed2 = n_motions - len(combined_results)
                rnd = get_random_motions(needed2)
                for r in rnd:
                    if r["motion_combo"] not in used_combo:
                        combined_results.append(r)
                        combined_scores.append('RANDOM')
                        used_combo.add(r["motion_combo"])
                    if len(combined_results) == n_motions:
                        break
        
        result = (combined_results[:n_motions], combined_scores[:n_motions])
    
    # Cache the results
    SEARCH_RESULTS_CACHE[cache_key] = result
    
    return result

def safe_video_update(motion_data, semantic_score, visible=True):
    """Optimized video update with match type display"""
    
    # Prepare the annotation text based on the match type
    if semantic_score == 'EXACT':
        match_info = "Exact Match"
    elif semantic_score == 'SUBSTR':
        match_info = "Substring Match"
    elif semantic_score == 'RANDOM':
        match_info = "Random Result"
    else:
        # For semantic matches, round to 2 decimal places
        ssim = str(round(semantic_score, 2)) if semantic_score != 'NA' else ''
        match_info = f"Semantic Match (sim: {ssim})"
    
    actual_annot = f"{motion_data['annotation']} | {match_info}"
    
    return [
        gr.update(value=url, visible=visible)
        for url in (motion_data["motion_combo"], 
                    motion_data["motion_a"], 
                    motion_data["motion_b"])
    ] + [gr.update(value=actual_annot, visible=visible)]

def update_videos(motions, n_visible, semantic_scores):
    """Update video components with motion data, with parallel video processing"""
    updates = []
    if not motions:
        updates.append(gr.update(value='incompatible combination', visible=True))
        remaining = 7
        for _ in range(remaining):
            updates.extend([
                gr.update(value=None, visible=False),
                gr.update(value=None, visible=False),
                gr.update(value=None, visible=False),
                gr.update(value=None, visible=False)
            ])
    else:
        try:
            # Prepare all updates in parallel using ThreadPoolExecutor
            with ThreadPoolExecutor(max_workers=min(8, n_visible)) as executor:
                # Submit all video update tasks
                future_updates = [
                    executor.submit(safe_video_update, motion, semantic_scores[jj], True)
                    for jj, motion in enumerate(motions[:n_visible])
                ]
                
                # Collect all updates as they complete
                for future in future_updates:
                    updates.extend(future.result())
            
            remaining = 8 - len(motions[:n_visible])
            for _ in range(remaining):
                updates.extend([
                    gr.update(value=None, visible=False),
                    gr.update(value=None, visible=False),
                    gr.update(value=None, visible=False),
                    gr.update(value=None, visible=False)
                ])
        except Exception as e:
            print(f"Error updating videos: {e}")
            # Fallback if parallel processing fails
            updates = []
            for i in range(8):
                if i < len(motions[:n_visible]):
                    motion = motions[i]
                    score = semantic_scores[i]
                    
                    # Handle different score types
                    if score == 'EXACT':
                        match_info = "Exact Match"
                    elif score == 'SUBSTR':
                        match_info = "Substring Match"
                    elif score == 'RANDOM':
                        match_info = "Random Result"
                    else:
                        # For semantic matches, round to 2 decimal places
                        ssim = str(round(score, 2)) if score != 'NA' else ''
                        match_info = f"Semantic Match (sim: {ssim})"
                    
                    actual_annot = f"{motion['annotation']} | {match_info}"
                    updates.extend([
                        gr.update(value=motion["motion_combo"], visible=True),
                        gr.update(value=motion["motion_a"], visible=True),
                        gr.update(value=motion["motion_b"], visible=True),
                        gr.update(value=actual_annot, visible=True)
                    ])
                else:
                    updates.extend([
                        gr.update(value=None, visible=False),
                        gr.update(value=None, visible=False),
                        gr.update(value=None, visible=False),
                        gr.update(value=None, visible=False)
                    ])
    
    return updates

def parse_gpt_labels(text):
    """Parse GPT labels from text"""
    if text.startswith("Answer: "):
        text = text[len("Answer: "):]  # Remove the "Answer: " prefix
    return text.split("\n")  # Split by newline

def failure_update(message, n_motions=None):
    """Create UI updates for failure cases"""
    updates = []
    # For the first motion: hide videos and display the message in the text box
    updates.append(gr.update(value=None, visible=False))  # video_combo for motion 1
    updates.append(gr.update(value=None, visible=False))  # video_a for motion 1
    updates.append(gr.update(value=None, visible=False))  # video_b for motion 1
    updates.append(gr.update(value=message, visible=True))  # annotation text for motion 1
    
    # For the remaining 7 motions, hide all components
    for _ in range(7):
        updates.extend([
            gr.update(value=None, visible=False),
            gr.update(value=None, visible=False),
            gr.update(value=None, visible=False),
            gr.update(value=None, visible=False)
        ])
    
    return updates

def handle_interaction(action1, action2, n_motions):
    """Handle user interaction with caching for faster responses"""
    # Create a cache key for the entire interaction
    cache_key = f"interaction_{action1.strip().lower()}_{action2.strip().lower()}_{n_motions}"
    
    # Check if we have cached results for this interaction
    if cache_key in SEARCH_RESULTS_CACHE:
        return SEARCH_RESULTS_CACHE[cache_key]
    
    try:
        if not action1.strip() and not action2.strip():
            # Both empty => random
            motions = get_random_motions(n_motions)
            result = update_videos(motions, n_motions, ['NA'] * len(motions))
        else:
            # Process GPT labels in parallel
            with ThreadPoolExecutor(max_workers=2) as executor:
                # Submit tasks for processing both actions in parallel
                if action1 in GPT_LABELS_LIST:
                    future_act1 = executor.submit(lambda: parse_gpt_labels(GPT_LABELS_LIST[action1]))
                else:
                    future_act1 = executor.submit(search_gpt_semantic, action1, 1)
                
                if action2 in GPT_LABELS_LIST:
                    future_act2 = executor.submit(lambda: parse_gpt_labels(GPT_LABELS_LIST[action2]))
                else:
                    future_act2 = executor.submit(search_gpt_semantic, action2, 1)
                
                # Get results
                try:
                    if action1 in GPT_LABELS_LIST:
                        gpt_act1 = future_act1.result()
                    else:
                        best_key, best_text, best_sim = future_act1.result()
                        if not best_text:
                            result = failure_update("Action 1 not recognized.")
                            SEARCH_RESULTS_CACHE[cache_key] = result
                            return result
                        gpt_act1 = parse_gpt_labels(best_text)
                    
                    if action2 in GPT_LABELS_LIST:
                        gpt_act2 = future_act2.result()
                    else:
                        best_key, best_text, best_sim = future_act2.result()
                        if not best_text:
                            result = failure_update("Action 2 not recognized.")
                            SEARCH_RESULTS_CACHE[cache_key] = result
                            return result
                        gpt_act2 = parse_gpt_labels(best_text)
                except Exception as e:
                    print(f"Error processing GPT labels: {e}")
                    result = failure_update("Error processing actions. Please try again.")
                    SEARCH_RESULTS_CACHE[cache_key] = result
                    return result
            
            # Check for conflicts
            if bool(set(gpt_act1) & set(gpt_act2)):
                failure_message = "Incompatible action pair. Please select actions that are not conflicting."
                result = failure_update(failure_message)
            else:
                motions, sem_mot_scores = search_motions_combined(action1, action2, n_motions)
                result = update_videos(motions, n_motions, sem_mot_scores)
    
    except Exception as e:
        print(f"Error in handle_interaction: {e}")
        result = failure_update("An error occurred. Please try again.")
    
    # Cache the result
    SEARCH_RESULTS_CACHE[cache_key] = result
    
    return result

# Custom CSS
CUSTOM_CSS = """
button.compact-button {
    width: auto !important;       /* Let the button shrink to fit text */
    min-width: unset !important;  /* Remove any forced min-width */
    padding: 4px 8px !important;
    font-size: 20px !important;
    line-height: 1 !important;
}
"""

# Build the Gradio UI
with gr.Blocks(css=CUSTOM_CSS) as demo:
    gr.HTML(WEBSITE)
    with gr.Tabs():
        with gr.Tab("SINC-Synth exploration"):
            with gr.Row():
                with gr.Column():
                    with gr.Row():
                        with gr.Column():
                            action1_textbox = gr.Textbox(
                                label="Action 1",
                                placeholder="Select an action or type the first action, e.g. 'walk'",
                            )
                            create_example_buttons(action1_textbox, ACTION_EXAMPLES[:5])
                        with gr.Column():
                            action2_textbox = gr.Textbox(
                                label="Action 2",
                                placeholder="Select an action or type the second action, e.g. 'wave'"
                            )
                            create_example_buttons(action2_textbox,  ACTION_EXAMPLES[5:])
                        with gr.Column():
                            n_motions_radio = gr.Radio(
                                    choices=[2, 4, 6, 8],
                                    label="Number of motions to be shown from the SINC-Synthetic data",
                                    value=2,
                                    show_label=True,
                                    container=True,
                                )
                            with gr.Row():
                                search_button = gr.Button("Search", 
                                                            elem_classes=["compact-button"])
                                random_button = gr.Button("Random", 
                                                            elem_classes=["compact-button"])

            # up to 8 motions
            motion_components = []
            videos_per_row = 2
            max_motions = 8
            num_rows = (max_motions + videos_per_row - 1) // videos_per_row  # Ceiling division

            for i in range(num_rows):
                with gr.Row():
                    for j in range(videos_per_row):
                        motion_index = i * videos_per_row + j
                        if motion_index >= max_motions:
                            break
                            
                        with gr.Column():
                            video_combo = gr.Video(
                                label=f"Motion {motion_index + 1}",
                                visible=False, 
                                width=480,
                                height=384,
                                loop=True
                            )
                            with gr.Row():
                                video_a = gr.Video(
                                    label="Motion A",
                                    visible=False, 
                                    width=320,
                                    height=256,
                                    loop=True
                                )
                                video_b = gr.Video(
                                    label="Motion B",
                                    visible=False, 
                                    width=320,
                                    height=256,
                                    loop=True
                                )
                            text = gr.Textbox(
                                visible=False,
                                interactive=False
                            )
                            motion_components.extend([video_combo, video_a, video_b, text])

            search_button.click(
                fn=handle_interaction,
                inputs=[action1_textbox, action2_textbox, n_motions_radio],
                outputs=motion_components
            )
            random_button.click(
                fn=lambda n: handle_interaction("", "", n),
                inputs=[n_motions_radio],
                outputs=motion_components
            )

            gr.HTML(("""
                    <div style='text-align: center; margin-top: 20px; font-size: 16px;'>
                    <p><sup>**</sup>Our data in the official paper are using on the fly compositions,
                    which means than are not computed and filtered offline. This is a minimally 
                    processed version of ~124k motions ranging between 3-7 seconds.</p>
                    <p>Made with ❤️ by Nikos Athanasiou</p>
                    </div>
                    """)
                    )
        with gr.Tab("Simultaneous Motion Generation with SINC model"):
            gr.HTML("<h2>Motion Generation from Text [TBD. Currenly under construction.]</h2>")
            with gr.Row():
                text_input_gen = gr.Textbox(
                    label="Motion Description",
                    placeholder="Describe the motion, e.g. 'A person walking forward while waving'"
                )
                create_example_buttons(text_input_gen, ACTION_EXAMPLES_SIMULTANEOUS)

                generate_button = gr.Button("Generate Motion",
                                            elem_classes=["compact-button"])
            
            with gr.Row():
                output_video = gr.Video(
                    label="Generated Motion",
                    visible=True,
                    width=320,
                    height=180
                )
            
            def generate_motion(text):
                # Placeholder function - replace with actual model inference
                # Return None instead of a string path to avoid schema conversion issues
                return None
                
            generate_button.click(
                fn=generate_motion,
                inputs=[text_input_gen],
                outputs=[output_video]
            )

# Initialize TF-IDF at startup
initialize_tfidf()

# Precompute synonyms for common action words
print("Precomputing synonyms for common action words...")
for action in ACTION_SYNONYMS:
    get_synonyms(action)

# Video prefetching
def prefetch_videos():
    """Prefetch some common videos to warm up the cache"""
    print("Prefetching common videos...")
    try:
        # Get a small set of common videos to prefetch
        random_motions = get_random_motions(4)
        common_actions = [("walk", "wave"), ("sit", "bow"), ("jump", "throw")]
        
        with ThreadPoolExecutor(max_workers=8) as executor:
            futures = []
            # Add random motions to prefetch list
            for motion in random_motions:
                futures.append(executor.submit(
                    lambda m: (m["motion_combo"], m["motion_a"], m["motion_b"]), 
                    motion
                ))
            
            # Add common action combinations
            for act1, act2 in common_actions:
                motions, _ = search_motions_combined(act1, act2, 2)
                if motions:
                    for motion in motions:
                        futures.append(executor.submit(
                            lambda m: (m["motion_combo"], m["motion_a"], m["motion_b"]),
                            motion
                        ))
            
            # Wait for all prefetch operations to complete
            for future in futures:
                future.result()
        
        print("Video prefetching complete")
    except Exception as e:
        print(f"Error in video prefetching: {e}")

# Start prefetching in a separate thread to not block startup
threading.Thread(target=prefetch_videos).start()

# Print ready message
print("Demo ready! Optimized code running with exact matching prioritized over synonym-enhanced TF-IDF similarity.")

# Launch the demo
demo.launch(server_name="0.0.0.0", server_port=7860, share=False)