polymesh-ai 0.3.3

Transformer Library for 3D Mesh Processing

polymesh-ai Library Usage Guide

Complete guide for using polymesh-ai - the transformer library for 3D mesh processing.

📦 Installation

Install from PyPI (Recommended)

# Basic installation
pip install polymesh-ai

# Full installation with all features
pip install polymesh-ai[full]

# Development installation
pip install polymesh-ai[dev]

Install from Source

git clone https://github.com/MatN23/polymesh-ai.git
cd polymesh-ai
pip install -e .

Installation Options

• Basic: pip install polymesh-ai - Core functionality only
• Full: pip install polymesh-ai[full] - Includes wandb, matplotlib, scipy, scikit-learn
• Dev: pip install polymesh-ai[dev] - Development tools (pytest, black, flake8, jupyter)
• Docs: pip install polymesh-ai[docs] - Documentation tools (sphinx, themes)

🚀 Quick Start

1. Basic Mesh Classification

import torch
import polymesh-ai
from polymesh-ai import VertexTokenizer, MeshTransformer
import numpy as np

# Create sample mesh data (you'll need actual mesh data)
# This is just for demonstration - replace with your mesh loading code
class SimpleMesh:
    def __init__(self):
        # Create a simple triangle mesh
        self.vertices = [
            SimpleVertex([0.0, 0.0, 0.0], [0, 0, 1]),
            SimpleVertex([1.0, 0.0, 0.0], [0, 0, 1]), 
            SimpleVertex([0.5, 1.0, 0.0], [0, 0, 1])
        ]

class SimpleVertex:
    def __init__(self, position, normal=None):
        self.position = np.array(position, dtype=np.float32)
        self.normal = np.array(normal, dtype=np.float32) if normal else None

# Load or create your mesh
mesh = SimpleMesh()

# Initialize tokenizer
tokenizer = VertexTokenizer(
    include_normals=True,
    include_colors=False
)

# Create model
model = MeshTransformer(
    feature_dim=6,      # 3 position + 3 normal
    d_model=256,
    nhead=8,
    num_layers=6
)

# Tokenize and process
tokens = tokenizer.tokenize(mesh)
print(f"Generated {len(tokens)} tokens")

# Run classification
with torch.no_grad():
    output = model(tokens, task='classification')
    print(f"Classification output shape: {output.shape}")

2. Advanced Model with Adaptive Attention

from polymesh-ai import AdaptiveMeshTransformer

# Create adaptive model that switches attention mechanisms
model = AdaptiveMeshTransformer(
    d_model=512,
    nhead=8,
    num_layers=6,
    dim_feedforward=2048,
    dropout=0.1
)

# Prepare tensor inputs for adaptive model
batch_size, seq_len = 1, len(tokens)
features = torch.tensor([token.features for token in tokens]).unsqueeze(0)
positions = torch.tensor([token.position for token in tokens]).unsqueeze(0)

# Forward pass
output = model(features, positions)
print(f"Adaptive model output shape: {output.shape}")

🎯 Core Components

Tokenization Strategies

Vertex Tokenizer

from polymesh-ai import VertexTokenizer

# Basic vertex tokenization
tokenizer = VertexTokenizer(
    include_normals=True,
    include_colors=True,
    quantize_positions=False
)

tokens = tokenizer.tokenize(mesh)

Face Tokenizer

from polymesh-ai import FaceTokenizer

# Face-based tokenization
face_tokenizer = FaceTokenizer(
    max_face_vertices=4,
    include_face_normal=True
)

face_tokens = face_tokenizer.tokenize(mesh)

Patch Tokenizer

from polymesh-ai import PatchTokenizer

# Patch-based hierarchical tokenization
patch_tokenizer = PatchTokenizer(
    patch_size=16,
    overlap=4,
    feature_aggregation='mean'
)

patch_tokens = patch_tokenizer.tokenize(mesh)

Attention Mechanisms

Geometric Attention

from polymesh-ai import GeometricAttention

# Distance-aware attention
geo_attention = GeometricAttention(
    d_model=256,
    nhead=8,
    max_distance=5.0,
    distance_bins=32
)

Graph Attention

from polymesh-ai import GraphAttention

# Graph-based attention for mesh connectivity
graph_attention = GraphAttention(
    d_model=256,
    nhead=8,
    edge_dim=16
)

Multi-Scale Attention

from polymesh-ai import MultiScaleAttention

# Hierarchical multi-scale processing
multiscale_attention = MultiScaleAttention(
    d_model=256,
    scales=[1, 2, 4, 8],
    nhead=8
)

Sparse Attention

from polymesh-ai import SparseAttention

# Efficient attention for large meshes
sparse_attention = SparseAttention(
    d_model=256,
    nhead=8,
    neighborhood_size=16,
    sparse_pattern='spatial'
)

🏋️ Training Pipeline

Complete Training Setup

from polymesh-ai import (
    MeshTransformerTrainingPipeline,
    MeshTransformerDataset,
    MeshAugmentation
)

# Training configuration
config = {
    'model_type': 'adaptive',
    'tokenizer_type': 'vertex',
    'd_model': 256,
    'nhead': 8,
    'num_layers': 6,
    'learning_rate': 1e-4,
    'batch_size': 16,
    'max_epochs': 100,
    'early_stopping_patience': 15,
    'device': 'cuda' if torch.cuda.is_available() else 'cpu',
    'use_wandb': False,  # Set True for experiment tracking
    'task_type': 'classification'
}

# Initialize training pipeline
pipeline = MeshTransformerTrainingPipeline(config)

# Prepare your dataset
train_data = [
    {'mesh': mesh1, 'label': 0, 'id': 'mesh_001'},
    {'mesh': mesh2, 'label': 1, 'id': 'mesh_002'},
    # ... more training data
]

val_data = [
    {'mesh': val_mesh1, 'label': 0, 'id': 'val_001'},
    # ... validation data
]

# Create datasets with augmentation
train_dataset = MeshTransformerDataset(
    train_data,
    pipeline.tokenizer,
    task_type='classification',
    augmentation_fn=MeshAugmentation.random_augment,
    max_seq_len=512
)

val_dataset = MeshTransformerDataset(
    val_data,
    pipeline.tokenizer,
    task_type='classification',
    max_seq_len=512
)

# Train the model
pipeline.train(train_dataset, val_dataset)

Data Augmentation

from polymesh-ai import MeshAugmentation

# Individual augmentations
rotated_mesh = MeshAugmentation.random_rotation(mesh, angle_range=30.0)
scaled_mesh = MeshAugmentation.random_scale(mesh, scale_range=(0.8, 1.2))
translated_mesh = MeshAugmentation.random_translation(mesh, translation_range=0.1)
noisy_mesh = MeshAugmentation.add_noise(mesh, noise_std=0.01)

# Combined random augmentation
augmented_mesh = MeshAugmentation.random_augment(mesh)

🎨 Use Cases & Applications

3D Shape Classification

# Configure for classification
config = {
    'model_type': 'standard',
    'task_type': 'classification',
    'feature_dim': 6,  # Position + Normal
    'd_model': 512,
    'num_layers': 8
}

pipeline = MeshTransformerTrainingPipeline(config)

# Your mesh dataset with labels
mesh_classes = ['chair', 'table', 'lamp', 'sofa']
# Train classifier...

Mesh Reconstruction/Autoencoder

# Configure for reconstruction
config = {
    'model_type': 'standard',
    'task_type': 'reconstruction',
    'feature_dim': 6,
    'd_model': 512,
    'learning_rate': 1e-5,
    'max_epochs': 200
}

pipeline = MeshTransformerTrainingPipeline(config)

# Train autoencoder for mesh compression/denoising
model = pipeline.model
with torch.no_grad():
    reconstructed = model(tokens, task='reconstruction')

Mesh Generation

# Configure for generation
config = {
    'model_type': 'adaptive',
    'task_type': 'generation',
    'd_model': 768,
    'num_layers': 12
}

# Pre-training tasks
from polymesh-ai import MeshTransformerPreTrainer

trainer = MeshTransformerPreTrainer(model, tokenizer)

# Masked mesh modeling
masked_input, positions, targets = trainer.masked_mesh_modeling(tokens, mask_ratio=0.15)

# Mesh completion
partial_input, complete_target = trainer.mesh_completion_task(partial_tokens, complete_tokens)

🔧 Advanced Features

Custom Attention Mechanisms

from polymesh-ai import MeshTransformerLayer

# Create custom transformer layers
custom_layer = MeshTransformerLayer(
    d_model=256,
    nhead=8,
    attention_type='geometric'  # or 'graph', 'multiscale', 'sparse'
)

3D Positional Encoding

from polymesh-ai import MeshPositionalEncoding

# Custom 3D positional encoding
pos_encoding = MeshPositionalEncoding(
    d_model=256,
    max_freq=10.0,
    num_freq_bands=10
)

# Apply to 3D positions
positions = torch.randn(batch_size, seq_len, 3)
pos_embeddings = pos_encoding(positions)

Model Checkpointing

# Save model checkpoint
pipeline.save_checkpoint('my_model_epoch_50.pth')

# Load checkpoint
pipeline.load_checkpoint('my_model_epoch_50.pth')

# Save just the model state
torch.save(pipeline.model.state_dict(), 'model_weights.pth')

📊 Monitoring & Logging

WandB Integration

# Enable WandB logging
config = {
    'use_wandb': True,
    'wandb_project': 'mesh-transformer-experiments',
    # ... other config
}

# Your training will automatically log to WandB
pipeline = MeshTransformerTrainingPipeline(config)

Custom Metrics

# Track custom metrics during training
def compute_custom_metrics(predictions, targets):
    # Your custom metric computation
    return {'custom_accuracy': accuracy, 'custom_f1': f1_score}

# Extend the training pipeline with custom metrics

🐛 Troubleshooting

Common Issues

1. Memory Issues with Large Meshes

# Use sparse attention for large meshes
model = MeshTransformer(
    feature_dim=6,
    d_model=256,
    # ... other params
)

# Or chunk your processing
def process_large_mesh(mesh, chunk_size=1000):
    tokens = tokenizer.tokenize(mesh)
    results = []
    for i in range(0, len(tokens), chunk_size):
        chunk = tokens[i:i+chunk_size]
        result = model(chunk)
        results.append(result)
    return torch.cat(results, dim=0)

2. GPU Memory Management

# Clear cache between batches
torch.cuda.empty_cache()

# Use mixed precision training
config['use_amp'] = True  # If implemented

3. Custom Mesh Loading

# Implement your own mesh class compatible with tokenizers
class CustomMesh:
    def __init__(self, vertices, faces=None):
        self.vertices = [CustomVertex(v) for v in vertices]
        self.faces = faces
        
class CustomVertex:
    def __init__(self, position, normal=None, color=None):
        self.position = np.array(position, dtype=np.float32)
        self.normal = np.array(normal, dtype=np.float32) if normal else None
        self.color = np.array(color, dtype=np.float32) if color else None

📚 Examples & Tutorials

Check the library's examples/ directory for:

• Basic classification tutorial
• Mesh autoencoder training
• Custom tokenizer implementation
• Advanced attention mechanism usage
• Large-scale training pipelines

🤝 Contributing

The library is actively developed! Contribute by:

1. Reporting issues on GitHub
2. Submitting pull requests
3. Adding new attention mechanisms
4. Improving documentation
5. Creating examples and tutorials

---

Happy mesh processing with polymesh-ai! 🎉