wm/dataset/visualize_dataset.py

import os
import torch
import gradio as gr
import random
import pandas as pd
import numpy as np
from wm.dataset.data_config import DATASET_REGISTRY

def load_dataset_samples(dataset_name):
    if dataset_name not in DATASET_REGISTRY:
        return [None] * 10 + [None] * 10 + [f"Error: Dataset {dataset_name} not found."]
    
    config = DATASET_REGISTRY[dataset_name]
    root_dir = config["root_dir"]
    metadata_path = os.path.join(root_dir, "metadata.pt")
    
    if not os.path.exists(metadata_path):
        return [None] * 10 + [None] * 10 + [f"Error: Metadata not found at {metadata_path}. Wait for conversion to finish if this is a new dataset."]
    
    try:
        metadata = torch.load(metadata_path, weights_only=False)
    except Exception as e:
        return [None] * 10 + [None] * 10 + [f"Error loading metadata: {e}"]
    
    num_samples = min(10, len(metadata))
    indices = random.sample(range(len(metadata)), num_samples)
    
    videos = []
    actions_text = []
    
    for idx in indices:
        entry = metadata[idx]
        video_rel_path = entry["video_path"]
        video_full_path = os.path.join(root_dir, video_rel_path)
        
        if not os.path.exists(video_full_path):
            videos.append(None)
            actions_text.append(f"Video not found: {video_full_path}")
            continue
            
        videos.append(video_full_path)
        
        # Format actions
        if "actions" in entry:
            act = entry["actions"]
        elif "commands" in entry:
            # Handle RECON commands
            lin = entry["commands"]["linear_velocity"]
            ang = entry["commands"]["angular_velocity"]
            act = torch.stack([lin, ang], dim=-1)
        else:
            act = torch.zeros((1, 0))
            
        if isinstance(act, torch.Tensor):
            act_np = act.numpy()
        else:
            act_np = np.array(act)
            
        # Create a string representation for the textbox
        actions_str = f"Shape: {act_np.shape}\n"
        actions_str += str(act_np)
        actions_text.append(actions_str)
        
    # Pad to 10 if necessary
    while len(videos) < 10:
        videos.append(None)
        actions_text.append("")
        
    return videos + actions_text + [f"Loaded {num_samples} samples from {dataset_name}"]

def create_visualizer():
    with gr.Blocks(title="Dataset Visualizer") as demo:
        gr.Markdown("# 📊 World Model Dataset Visualizer")
        
        with gr.Row():
            dataset_dropdown = gr.Dropdown(
                choices=list(DATASET_REGISTRY.keys()), 
                label="Select Dataset",
                value="language_table"
            )
            refresh_btn = gr.Button("🔄 Random Sample 10", variant="primary")
            
        status_text = gr.Markdown("Select a dataset and click refresh.")
        
        video_components = []
        action_components = []
        
        for i in range(5):
            with gr.Row():
                with gr.Column():
                    v = gr.Video(label=f"Sample {i*2}", interactive=False)
                    a = gr.Textbox(label=f"Actions {i*2}", lines=10)
                    video_components.append(v)
                    action_components.append(a)
                with gr.Column():
                    v = gr.Video(label=f"Sample {i*2+1}", interactive=False)
                    a = gr.Textbox(label=f"Actions {i*2+1}", lines=10)
                    video_components.append(v)
                    action_components.append(a)
                    
        def on_refresh(name):
            results = load_dataset_samples(name)
            # results is [v0, v1, ..., v9, a0, a1, ..., a9, status]
            return results
            
        refresh_btn.click(
            fn=on_refresh,
            inputs=[dataset_dropdown],
            outputs=video_components + action_components + [status_text],
            api_name=False
        )
        
    return demo

if __name__ == "__main__":
    demo = create_visualizer()
    demo.launch(server_name="0.0.0.0", server_port=7861, share=True)