app.py

import os
import sys
import torch
import numpy as np
import gradio as gr
import matplotlib.pyplot as plt
from matplotlib.patches import Rectangle, FancyBboxPatch
import h5py
import requests
from tqdm import tqdm
import json
from pathlib import Path

# Add spaces GPU decorator if available
try:
    import spaces
    GPU_AVAILABLE = True
except ImportError:
    GPU_AVAILABLE = False
    # Create dummy decorator if spaces not available
    class spaces:
        @staticmethod
        def GPU(func):
            return func

# Add current directory to path for imports
sys.path.insert(0, os.path.dirname(__file__))

# Import DeepCAD modules (assuming they're in the repository)
try:
    from config import ConfigAE
    from trainer import TrainerAE
    from dataset import CADDataset
except ImportError:
    print("Warning: Could not import DeepCAD modules. Creating minimal config...")
    # Create minimal config class if import fails
    class ConfigAE:
        def __init__(self):
            self.exp_name = "pretrained"
            self.device = "cuda" if torch.cuda.is_available() else "cpu"
            self.n_commands = 60
            self.n_args = 256
            self.dim = 256
            self.n_layers = 4

# Constants
MODEL_URL = "http://www.cs.columbia.edu/cg/deepcad/pretrained.tar"
CHECKPOINT_DIR = "pretrained_model"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

def download_pretrained_model():
    """Download the pretrained DeepCAD model if not already present."""
    checkpoint_path = os.path.join(CHECKPOINT_DIR, "ckpt_1000.pt")
    
    if os.path.exists(checkpoint_path):
        print(f"✓ Pretrained model found at {checkpoint_path}")
        return checkpoint_path
    
    print("Downloading pretrained model...")
    os.makedirs(CHECKPOINT_DIR, exist_ok=True)
    
    try:
        response = requests.get(MODEL_URL, stream=True)
        response.raise_for_status()
        
        tar_path = os.path.join(CHECKPOINT_DIR, "pretrained.tar")
        total_size = int(response.headers.get('content-length', 0))
        
        with open(tar_path, 'wb') as f, tqdm(
            total=total_size,
            unit='B',
            unit_scale=True,
            desc="Downloading"
        ) as pbar:
            for chunk in response.iter_content(chunk_size=8192):
                if chunk:
                    f.write(chunk)
                    pbar.update(len(chunk))
        
        # Extract tar file
        import tarfile
        with tarfile.open(tar_path, 'r') as tar:
            tar.extractall(CHECKPOINT_DIR)
        
        os.remove(tar_path)
        print("✓ Model downloaded and extracted successfully!")
        return checkpoint_path
        
    except Exception as e:
        print(f"Error downloading model: {e}")
        return None

class SimpleCADDecoder(torch.nn.Module):
    """Simplified CAD decoder for inference."""
    
    def __init__(self, dim=256, n_commands=60, n_args=256):
        super().__init__()
        self.dim = dim
        self.n_commands = n_commands
        self.n_args = n_args
        
        # Simple decoder architecture
        self.fc_layers = torch.nn.Sequential(
            torch.nn.Linear(dim, dim * 2),
            torch.nn.ReLU(),
            torch.nn.Linear(dim * 2, dim * 2),
            torch.nn.ReLU(),
            torch.nn.Linear(dim * 2, n_commands * n_args)
        )
    
    def forward(self, z):
        """Decode latent vector to CAD sequence."""
        batch_size = z.size(0)
        out = self.fc_layers(z)
        out = out.view(batch_size, self.n_commands, self.n_args)
        return out

class DeepCADModel:
    """Wrapper for DeepCAD model with inference capabilities."""
    
    def __init__(self, checkpoint_path=None):
        self.device = DEVICE
        self.dim = 256
        self.n_commands = 60
        self.n_args = 256
        
        # Initialize model
        self.model = SimpleCADDecoder(
            dim=self.dim,
            n_commands=self.n_commands,
            n_args=self.n_args
        ).to(self.device)
        
        # Load checkpoint if provided
        if checkpoint_path and os.path.exists(checkpoint_path):
            try:
                checkpoint = torch.load(checkpoint_path, map_location=self.device)
                # Try to load decoder weights
                if 'decoder' in checkpoint:
                    self.model.load_state_dict(checkpoint['decoder'])
                elif 'model' in checkpoint:
                    self.model.load_state_dict(checkpoint['model'])
                print(f"✓ Loaded checkpoint from {checkpoint_path}")
            except Exception as e:
                print(f"Warning: Could not load checkpoint: {e}")
                print("Using randomly initialized model...")
        
        self.model.eval()
    
    def generate_from_latent(self, z):
        """Generate CAD sequence from latent vector."""
        with torch.no_grad():
            if isinstance(z, np.ndarray):
                z = torch.from_numpy(z).float()
            z = z.to(self.device)
            if len(z.shape) == 1:
                z = z.unsqueeze(0)
            
            output = self.model(z)
            return output.cpu().numpy()
    
    def random_generation(self, seed=None):
        """Generate a random CAD sequence."""
        if seed is not None:
            np.random.seed(seed)
            torch.manual_seed(seed)
        
        # Sample random latent vector from normal distribution
        z = torch.randn(1, self.dim).to(self.device)
        return self.generate_from_latent(z)

def visualize_cad_sequence(cad_sequence, title="Generated CAD Model"):
    """
    Visualize CAD sequence as a 2D projection.
    Since we can't use pythonocc-core, we'll create a simplified visualization.
    """
    fig = plt.figure(figsize=(12, 8))
    
    # Main plot: 2D projection of CAD operations
    ax1 = plt.subplot(2, 2, (1, 3))
    ax1.set_title(title, fontsize=14, fontweight='bold')
    ax1.set_xlim(-5, 5)
    ax1.set_ylim(-5, 5)
    ax1.set_aspect('equal')
    ax1.grid(True, alpha=0.3)
    ax1.set_xlabel('X')
    ax1.set_ylabel('Y')
    
    # Parse and visualize the sequence
    sequence = cad_sequence[0] if len(cad_sequence.shape) == 3 else cad_sequence
    
    # Extract meaningful features from the sequence
    # Each command has multiple arguments representing geometric operations
    colors = plt.cm.viridis(np.linspace(0, 1, len(sequence)))
    
    for i, command in enumerate(sequence):
        # Interpret command parameters as geometric primitives
        # This is a simplified interpretation
        if np.abs(command).max() > 0.01:  # Skip near-zero commands
            # Extract position and size parameters
            x = command[0] * 4  # Scale to viewport
            y = command[1] * 4
            width = np.abs(command[2]) * 2 + 0.3
            height = np.abs(command[3]) * 2 + 0.3
            
            # Draw a rectangle representing this operation
            rect = FancyBboxPatch(
                (x - width/2, y - height/2),
                width, height,
                boxstyle="round,pad=0.05",
                edgecolor=colors[i],
                facecolor=colors[i],
                alpha=0.3,
                linewidth=2
            )
            ax1.add_patch(rect)
            
            # Add operation number
            if i % 5 == 0:  # Label every 5th operation
                ax1.text(x, y, str(i), ha='center', va='center',
                        fontsize=8, color='black', fontweight='bold')
    
    # Command histogram
    ax2 = plt.subplot(2, 2, 2)
    ax2.set_title('Command Distribution', fontsize=12)
    command_magnitudes = np.linalg.norm(sequence, axis=1)
    ax2.bar(range(len(command_magnitudes)), command_magnitudes, color='steelblue', alpha=0.7)
    ax2.set_xlabel('Command Index')
    ax2.set_ylabel('Magnitude')
    ax2.grid(True, alpha=0.3)
    
    # Parameter statistics
    ax3 = plt.subplot(2, 2, 4)
    ax3.set_title('Parameter Statistics', fontsize=12)
    param_stats = {
        'Mean': np.mean(np.abs(sequence)),
        'Std': np.std(sequence),
        'Max': np.max(np.abs(sequence)),
        'Non-zero': np.sum(np.abs(sequence) > 0.01) / sequence.size
    }
    bars = ax3.bar(param_stats.keys(), param_stats.values(), color='coral', alpha=0.7)
    ax3.set_ylabel('Value')
    ax3.grid(True, alpha=0.3)
    
    # Add value labels on bars
    for bar in bars:
        height = bar.get_height()
        ax3.text(bar.get_x() + bar.get_width()/2., height,
                f'{height:.3f}',
                ha='center', va='bottom', fontsize=9)
    
    plt.tight_layout()
    return fig

def save_cad_sequence(cad_sequence, filename="generated_cad.h5"):
    """Save CAD sequence to H5 file."""
    with h5py.File(filename, 'w') as f:
        f.create_dataset('cad_sequence', data=cad_sequence)
        f.attrs['format'] = 'DeepCAD vectorized representation'
    return filename

# Initialize model globally
print("Initializing DeepCAD model...")
checkpoint_path = download_pretrained_model()
model = DeepCADModel(checkpoint_path)
print(f"✓ Model initialized on {DEVICE}")

# Add GPU decorator to the generation function
@spaces.GPU(duration=60)  # Reserve GPU for 60 seconds
def generate_cad(seed, temperature):
    """Generate CAD model from seed and temperature."""
    try:
        # Set random seed for reproducibility
        if seed >= 0:
            np.random.seed(seed)
            torch.manual_seed(seed)
        
        # Generate random latent vector with temperature scaling
        z = torch.randn(1, model.dim) * temperature
        
        # Generate CAD sequence
        cad_sequence = model.generate_from_latent(z)
        
        # Create visualization
        fig = visualize_cad_sequence(
            cad_sequence,
            title=f"Generated CAD Model (seed={seed}, temp={temperature:.2f})"
        )
        
        # Save to H5 file
        h5_filename = f"generated_cad_seed{seed}.h5"
        save_cad_sequence(cad_sequence, h5_filename)
        
        # Create info text
        info_text = f"""
        **Generation Info:**
        - Seed: {seed}
        - Temperature: {temperature:.2f}
        - Device: {DEVICE}
        - Sequence shape: {cad_sequence.shape}
        - Non-zero commands: {np.sum(np.abs(cad_sequence) > 0.01)}
        
        **Note:** The visualization shows a 2D projection of the CAD operations.
        Download the H5 file to use with full DeepCAD evaluation tools.
        """
        
        return fig, h5_filename, info_text
        
    except Exception as e:
        import traceback
        error_msg = f"Error during generation:\n{str(e)}\n\n{traceback.format_exc()}"
        print(error_msg)
        # Return empty plot and error message
        fig = plt.figure(figsize=(8, 6))
        plt.text(0.5, 0.5, "Generation Failed\nSee console for details",
                ha='center', va='center', fontsize=14, color='red')
        plt.axis('off')
        return fig, None, error_msg

# Create Gradio interface
with gr.Blocks(title="DeepCAD: CAD Model Generation", theme=gr.themes.Soft()) as demo:
    gr.Markdown("""
    # 🔧 DeepCAD: Deep Generative Network for CAD Models
    
    Generate Computer-Aided Design (CAD) models using deep learning! This demo uses the DeepCAD model 
    from the ICCV 2021 paper by Wu, Xiao, and Zheng.
    
    **How it works:**
    1. Adjust the seed for different random generations
    2. Control temperature to adjust variation (higher = more creative, lower = more conservative)
    3. Click Generate to create a new CAD model
    4. Download the H5 file for use with full DeepCAD tools
    
    **Note:** Visualization is simplified (2D projection). For full 3D STEP export, use the downloaded 
    H5 file with the original DeepCAD repository tools.
    """)
    
    with gr.Row():
        with gr.Column(scale=1):
            gr.Markdown("### 🎛️ Generation Parameters")
            
            seed_input = gr.Slider(
                minimum=0,
                maximum=10000,
                value=42,
                step=1,
                label="Random Seed",
                info="Set seed for reproducible generation"
            )
            
            temperature_input = gr.Slider(
                minimum=0.1,
                maximum=2.0,
                value=1.0,
                step=0.1,
                label="Temperature",
                info="Controls generation diversity"
            )
            
            generate_btn = gr.Button("🚀 Generate CAD Model", variant="primary", size="lg")
            
            gr.Markdown("### 📊 Quick Stats")
            info_output = gr.Markdown()
            
            gr.Markdown("### 💾 Download")
            h5_output = gr.File(label="Download H5 File")
        
        with gr.Column(scale=2):
            gr.Markdown("### 🎨 Visualization")
            plot_output = gr.Plot(label="CAD Model Visualization")
    
    gr.Markdown("""
    ---
    ### 📚 References
    - **Paper:** [DeepCAD: A Deep Generative Network for Computer-Aided Design Models](https://arxiv.org/abs/2105.09492)
    - **Authors:** Rundi Wu, Chang Xiao, Changxi Zheng (Columbia University)
    - **Conference:** ICCV 2021
    - **Code:** [GitHub Repository](https://github.com/ChrisWu1997/DeepCAD)
    
    ### ℹ️ About
    This is a simplified deployment for demonstration. For full functionality including:
    - 3D STEP file export
    - Complete evaluation metrics
    - Training your own models
    
    Please visit the [official GitHub repository](https://github.com/ChrisWu1997/DeepCAD).
    """)
    
    # Connect the generate button
    generate_btn.click(
        fn=generate_cad,
        inputs=[seed_input, temperature_input],
        outputs=[plot_output, h5_output, info_output]
    )
    
    # Add examples
    gr.Examples(
        examples=[
            [42, 1.0],
            [123, 0.8],
            [999, 1.2],
            [2024, 1.5],
        ],
        inputs=[seed_input, temperature_input],
        outputs=[plot_output, h5_output, info_output],
        fn=generate_cad,
        cache_examples=False,
    )

# Launch the app
if __name__ == "__main__":
    print("\n" + "="*50)
    print("🚀 Starting DeepCAD Gradio Interface")
    print(f"📍 Device: {DEVICE}")
    print("="*50 + "\n")
    
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=False
    )