app.py

import gradio as gr
import cv2
import numpy as np
import pandas as pd
import random
import os
import shutil
from pathlib import Path
from datasets import load_dataset
from huggingface_hub import hf_hub_download
from tqdm import tqdm

HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HUGGING_FACE_HUB_TOKEN")

# Step 6: Add success/failure filtering

def sample_trajectories(dataset_repo, config_name, is_robot, quality_filter, num_samples, max_to_check=10000):
    """Sample random trajectories from HuggingFace dataset with quality filter."""
    try:
        if config_name:
            dataset = load_dataset(dataset_repo, config_name, split="train", streaming=True, token=HF_TOKEN)
        else:
            dataset = load_dataset(dataset_repo, split="train", streaming=True, token=HF_TOKEN)
        
        matching = []
        for i, sample in enumerate(dataset):
            if i >= max_to_check:
                break
            
            # Check robot/human
            if sample.get("is_robot", False) != is_robot:
                continue
            
            # Check quality/success if filter is applied
            if quality_filter != "All":
                quality_label = sample.get("quality_label", "")
                partial_success = sample.get("partial_success", None)
                
                if quality_filter == "Success":
                    # Check for success indicators
                    if quality_label and "success" not in quality_label.lower():
                        if partial_success is None or partial_success < 1:
                            continue
                elif quality_filter == "Failure":
                    # Check for failure indicators
                    if quality_label and "success" in quality_label.lower():
                        continue
                    if partial_success is not None and partial_success >= 1:
                        continue
            
            matching.append(sample)
        
        if len(matching) == 0:
            return []
        if len(matching) <= num_samples:
            random.shuffle(matching)
            return matching
        return random.sample(matching, num_samples)
    except Exception as e:
        print(f"Error sampling: {e}")
        return []

def download_video(trajectory, dataset_repo, config_name=None):
    """Download video for a trajectory."""
    video_path = trajectory.get("frames")
    if not video_path:
        return None
    
    cache_dir = Path("video_cache")
    repo_key = f"{dataset_repo}_{config_name}" if config_name else dataset_repo
    repo_key = repo_key.replace("/", "_").replace("\\", "_")
    dataset_cache_dir = cache_dir / repo_key
    dataset_cache_dir.mkdir(parents=True, exist_ok=True)
    
    local_video_path = dataset_cache_dir / Path(video_path).name
    if local_video_path.exists():
        return str(local_video_path)
    
    try:
        downloaded_path = hf_hub_download(
            repo_id=dataset_repo,
            repo_type="dataset",
            filename=video_path,
            local_dir=str(dataset_cache_dir),
            local_dir_use_symlinks=False,
            token=HF_TOKEN,
        )
        if Path(downloaded_path).exists():
            return downloaded_path
        return None
    except Exception as e:
        print(f"Error downloading: {e}")
        return None

def extract_frame(video_path, frame_num):
    """Extract a specific frame from video."""
    if not video_path or not os.path.exists(video_path):
        return None, "No video loaded", "0.0%"
    
    cap = cv2.VideoCapture(video_path)
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    
    if frame_num >= total_frames:
        frame_num = total_frames - 1
    
    cap.set(cv2.CAP_PROP_POS_FRAMES, frame_num)
    ret, frame = cap.read()
    cap.release()
    
    if ret:
        frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        percent = (frame_num / total_frames * 100) if total_frames > 0 else 0
        return frame_rgb, f"Frame {frame_num}/{total_frames-1}", f"{percent:.1f}%"
    return None, "Error reading frame", "0.0%"

# Global state
current_trajectories = []
current_idx = 0
labels_df = pd.DataFrame(columns=[
    "dataset_repo", "config_name", "trajectory_id", "is_robot", "quality_label",
    "task", "manual_end_frame", "manual_end_percent", "notes"
])

def load_labels():
    """Load existing labels from CSV."""
    global labels_df
    if Path("labels.csv").exists():
        labels_df = pd.read_csv("labels.csv")
        # Add quality_label column if missing
        if 'quality_label' not in labels_df.columns:
            labels_df['quality_label'] = ''

def save_labels():
    """Save labels to CSV."""
    global labels_df
    labels_df.to_csv("labels.csv", index=False)

def load_dataset_trajectories(dataset_repo, config_name, quality_filter, num_human, num_robot):
    """Load and download trajectories from dataset."""
    global current_trajectories, current_idx
    
    config = config_name.strip() if config_name else None
    
    try:
        human_trajs = sample_trajectories(dataset_repo, config, is_robot=False, quality_filter=quality_filter, num_samples=int(num_human))
        robot_trajs = sample_trajectories(dataset_repo, config, is_robot=True, quality_filter=quality_filter, num_samples=int(num_robot))
        all_trajs = human_trajs + robot_trajs
        
        if not all_trajs:
            return f"No {quality_filter.lower()} trajectories found", None, "No video", "", "0.0%", None, ""
        
        current_trajectories = []
        for traj in all_trajs:
            local_path = download_video(traj, dataset_repo, config)
            if local_path:
                traj["local_video_path"] = local_path
                traj["dataset_repo"] = dataset_repo
                traj["config_name"] = config
                current_trajectories.append(traj)
        
        current_idx = 0
        
        if current_trajectories:
            first_traj = current_trajectories[0]
            video_path = first_traj.get("local_video_path")
            task = first_traj.get("task", "No task description")
            is_robot_str = "Robot" if first_traj.get("is_robot") else "Human"
            quality = first_traj.get("quality_label", "Unknown")
            
            cap = cv2.VideoCapture(video_path)
            max_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) - 1
            cap.release()
            
            traj_info = f"Trajectory 1/{len(current_trajectories)} | Type: {is_robot_str} | Quality: {quality}"
            frame, frame_text, percent = extract_frame(video_path, 0)
            
            return (
                f"✅ Loaded {len(current_trajectories)} {quality_filter.lower()} trajectories ({len(human_trajs)} human, {len(robot_trajs)} robot)",
                gr.update(maximum=max_frames, value=0),
                video_path,
                task,
                percent,
                frame,
                traj_info
            )
        
        return "No videos downloaded", None, None, "", "0.0%", None, ""
        
    except Exception as e:
        return f"❌ Error: {str(e)}", None, None, "", "0.0%", None, ""

def save_label(dataset_repo, config_name, end_frame, notes):
    """Save label for current trajectory."""
    global current_trajectories, current_idx, labels_df
    
    if not current_trajectories or current_idx >= len(current_trajectories):
        return "No trajectory loaded"
    
    traj = current_trajectories[current_idx]
    video_path = traj.get("local_video_path")
    
    if not video_path:
        return "No video path"
    
    cap = cv2.VideoCapture(video_path)
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    cap.release()
    
    end_percent = (int(end_frame) / total_frames * 100) if total_frames > 0 else 0
    
    # Check if label exists
    mask = (
        (labels_df['dataset_repo'] == dataset_repo) &
        (labels_df['config_name'] == (config_name or "")) &
        (labels_df['trajectory_id'] == traj.get('id'))
    )
    
    if mask.any():
        idx = labels_df[mask].index[0]
        labels_df.at[idx, 'manual_end_frame'] = int(end_frame)
        labels_df.at[idx, 'manual_end_percent'] = end_percent
        labels_df.at[idx, 'notes'] = notes
        save_labels()
        return f"✅ Updated: Frame {int(end_frame)} ({end_percent:.1f}%)"
    
    new_row = pd.DataFrame([{
        "dataset_repo": dataset_repo,
        "config_name": config_name or "",
        "trajectory_id": traj.get('id'),
        "is_robot": traj.get('is_robot', False),
        "quality_label": traj.get('quality_label', ''),
        "task": traj.get('task', ''),
        "manual_end_frame": int(end_frame),
        "manual_end_percent": end_percent,
        "notes": notes
    }])
    
    labels_df = pd.concat([labels_df, new_row], ignore_index=True)
    save_labels()
    return f"✅ Saved: Frame {int(end_frame)} ({end_percent:.1f}%)"

def navigate_next():
    """Go to next trajectory."""
    global current_idx
    
    if not current_trajectories or current_idx >= len(current_trajectories) - 1:
        return "No more trajectories", None, "", "0.0%", None, ""
    
    current_idx += 1
    traj = current_trajectories[current_idx]
    video_path = traj.get("local_video_path")
    task = traj.get("task", "No task description")
    is_robot_str = "Robot" if traj.get("is_robot") else "Human"
    quality = traj.get("quality_label", "Unknown")
    
    cap = cv2.VideoCapture(video_path)
    max_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) - 1
    cap.release()
    
    traj_info = f"Trajectory {current_idx+1}/{len(current_trajectories)} | Type: {is_robot_str} | Quality: {quality}"
    frame, frame_text, percent = extract_frame(video_path, 0)
    
    return gr.update(maximum=max_frames, value=0), video_path, task, percent, frame, traj_info

def navigate_prev():
    """Go to previous trajectory."""
    global current_idx
    
    if not current_trajectories or current_idx <= 0:
        return "No previous trajectories", None, "", "0.0%", None, ""
    
    current_idx -= 1
    traj = current_trajectories[current_idx]
    video_path = traj.get("local_video_path")
    task = traj.get("task", "No task description")
    is_robot_str = "Robot" if traj.get("is_robot") else "Human"
    quality = traj.get("quality_label", "Unknown")
    
    cap = cv2.VideoCapture(video_path)
    max_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT)) - 1
    cap.release()
    
    traj_info = f"Trajectory {current_idx+1}/{len(current_trajectories)} | Type: {is_robot_str} | Quality: {quality}"
    frame, frame_text, percent = extract_frame(video_path, 0)
    
    return gr.update(maximum=max_frames, value=0), video_path, task, percent, frame, traj_info

def analyze_patterns(dataset_repo, config_name):
    """Analyze patterns - requires all trajectories to be labeled."""
    global current_trajectories, labels_df
    
    if not current_trajectories:
        return {"error": "No trajectories loaded"}
    
    traj = current_trajectories[current_idx]
    is_robot = traj.get('is_robot', False)
    
    expected_count = len([t for t in current_trajectories if t.get('is_robot', False) == is_robot])
    
    filtered = labels_df[
        (labels_df['dataset_repo'] == dataset_repo) &
        (labels_df['config_name'] == (config_name or "")) &
        (labels_df['is_robot'] == is_robot)
    ]
    
    labeled_count = len(filtered)
    
    if labeled_count < expected_count:
        return {
            "error": True,
            "message": f"Only {labeled_count}/{expected_count} {'robot' if is_robot else 'human'} trajectories labeled. Label all before analyzing.",
            "labeled_count": labeled_count,
            "expected_count": expected_count
        }
    
    if labeled_count < 3:
        return {
            "error": True,
            "message": "Need at least 3 labels",
            "labeled_count": labeled_count
        }
    
    percents = filtered['manual_end_percent'].values
    
    result = {
        "pattern_found": np.std(percents) < 10,
        "mean_percent": round(float(np.mean(percents)), 2),
        "median_percent": round(float(np.median(percents)), 2),
        "std_percent": round(float(np.std(percents)), 2),
        "min_percent": round(float(np.min(percents)), 2),
        "max_percent": round(float(np.max(percents)), 2),
        "quantile_90": round(float(np.percentile(percents, 90)), 2),
        "count": labeled_count,
        "suggested_label": round(float(np.mean(percents)))
    }
    
    if result["pattern_found"]:
        result["message"] = f"✅ Pattern detected! Low variance ({result['std_percent']}%)"
    else:
        result["message"] = f"⚠️ High variance ({result['std_percent']}%) - no clear pattern"
    
    return result

# Load existing labels on startup
load_labels()

with gr.Blocks(title="Trajectory End Point Labeler") as demo:
    gr.Markdown("# Trajectory End Point Labeler")
    gr.Markdown("Label trajectory end points and analyze patterns")
    
    with gr.Row():
        with gr.Column(scale=1):
            dataset_input = gr.Textbox(
                label="Dataset Repository",
                value="jesbu1/epic_rfm",
                placeholder="jesbu1/epic_rfm"
            )
            config_input = gr.Textbox(
                label="Config Name (optional)",
                placeholder="Leave empty if no config"
            )
            quality_filter = gr.Radio(
                choices=["All", "Success", "Failure"],
                value="All",
                label="Trajectory Quality Filter"
            )
            num_human = gr.Number(label="Human Samples", value=10, precision=0)
            num_robot = gr.Number(label="Robot Samples", value=10, precision=0)
            load_btn = gr.Button("Load Dataset", variant="primary")
            status = gr.Textbox(label="Status", interactive=False)
        
        with gr.Column(scale=2):
            traj_info = gr.Textbox(label="Current Trajectory", interactive=False)
            task_display = gr.Textbox(label="Task Description", interactive=False)
            
            with gr.Row():
                prev_btn = gr.Button("← Previous")
                next_btn = gr.Button("Next →")
            
            video_player = gr.Video(label="Trajectory Video")
            frame_slider = gr.Slider(minimum=0, maximum=63, step=1, value=0, label="Frame Number")
            frame_display = gr.Image(label="Current Frame")
            frame_info = gr.Textbox(label="Frame Info", interactive=False)
            
            with gr.Row():
                end_frame_input = gr.Number(label="End Frame", value=0, precision=0)
                end_percent = gr.Textbox(label="End Percent", interactive=False)
            
            notes_input = gr.Textbox(label="Notes (optional)", placeholder="Add notes...")
            save_btn = gr.Button("Save Label", variant="primary")
            save_status = gr.Textbox(label="Save Status", interactive=False)
    
    # Pattern analysis section
    with gr.Row():
        with gr.Column():
            gr.Markdown("### Pattern Analysis")
            gr.Markdown("Requires all trajectories of same type to be labeled")
            analyze_btn = gr.Button("Analyze Pattern", variant="secondary")
            pattern_output = gr.JSON(label="Pattern Metrics")
    
    # Load dataset
    load_btn.click(
        load_dataset_trajectories,
        inputs=[dataset_input, config_input, quality_filter, num_human, num_robot],
        outputs=[status, frame_slider, video_player, task_display, end_percent, frame_display, traj_info]
    )
    
    # Navigate
    next_btn.click(
        navigate_next,
        outputs=[frame_slider, video_player, task_display, end_percent, frame_display, traj_info]
    )
    
    prev_btn.click(
        navigate_prev,
        outputs=[frame_slider, video_player, task_display, end_percent, frame_display, traj_info]
    )
    
    # Frame navigation
    frame_slider.change(
        extract_frame,
        inputs=[video_player, frame_slider],
        outputs=[frame_display, frame_info, end_percent]
    )
    
    video_player.change(
        lambda v: extract_frame(v, 0) if v else (None, "No video", "0.0%"),
        inputs=[video_player],
        outputs=[frame_display, frame_info, end_percent]
    )
    
    # Update percent when end frame changes
    end_frame_input.change(
        lambda v, f: (None, "No video", "0.0%")[2] if not v else f"{(int(f) / int(cv2.VideoCapture(v).get(cv2.CAP_PROP_FRAME_COUNT)) * 100):.1f}%" if os.path.exists(v) and int(cv2.VideoCapture(v).get(cv2.CAP_PROP_FRAME_COUNT)) > 0 else "0.0%",
        inputs=[video_player, end_frame_input],
        outputs=[end_percent]
    )
    
    # Save label
    save_btn.click(
        save_label,
        inputs=[dataset_input, config_input, end_frame_input, notes_input],
        outputs=[save_status]
    )
    
    # Pattern analysis
    analyze_btn.click(
        analyze_patterns,
        inputs=[dataset_input, config_input],
        outputs=[pattern_output]
    )

if __name__ == "__main__":
    demo.launch()