🧬 PokeVAE — Variational Autoencoder for Pokémon Stats, Types & Abilities

Model Summary

PokeVAE is a Variational Autoencoder (VAE) trained on Pokémon base data to learn a continuous latent representation of Pokémon base stats, types, and abilities.

The model enables:

• Random Pokémon-like generation

This project is designed for creative exploration, procedural generation, and educational experimentation with VAEs.

---

Model Details

• Model type: Variational Autoencoder (VAE)
• Framework: PyTorch
• Latent dimension: 16
• Input modalities:
  • Base stats (standardized)
  • Pokémon types (multi-label)
  • Pokémon abilities (single-label)
• Outputs:
  • Base stats (continuous)
  • Types (multi-label logits)
  • Ability (categorical logits)

---

Architecture

Encoder

• Fully connected MLP
• Input dimension:
  • 6 (stats) + 18 (types) + 286 (abilities) = 310
  • Hidden dimension: 128
• Outputs:
  • Mean vector μ ∈ ℝ^16
  • Log-variance vector log σ² ∈ ℝ^16

Decoder

• MLP with dropout
• Hidden layers: [128, 64]
• Output heads:
  • Stats head: linear (regression)
  • Type head: linear (binary logits)
  • Talent head: linear (categorical logits)

---

Training

Dataset

• Pokémon dataset loaded from pokemons.json
• Features:
• Base stats: HP, Attack, Defense, Sp. Attack, Sp. Defense, Speed
• Types: multi-hot encoding
• Abilities: one-hot encoding (first listed ability only)

Preprocessing

• Base stats are standardized using sklearn.StandardScaler
• Types are treated as a multi-label classification problem
• Abilities are treated as categorical classification

---

Loss Function

The total loss is a weighted sum of reconstruction losses and KL divergence:

L = MSE(stats)

BCE(types)

0.50 × CE(ability)

β × KL

Where:

• MSE: Mean Squared Error for base stats
• BCE: Binary Cross-Entropy with logits for types
• CE: Cross-Entropy for abilities
• KL: Kullback–Leibler divergence
• β: linearly annealed during the first 100 epochs
• β_max: 0.012

KL divergence is clamped to prevent posterior collapse.

---

Optimization

• Optimizer: Adam
• Learning rate: 3e-4
• Batch size: 64
• Training epochs: 600

---

Inference & Generation

Random Sampling

Samples from the latent prior and decodes into Pokémon-like entries:

• Ensures at least one type is assigned
• Ability selected via argmax
• Stats are inverse-transformed for readability

Implemented in inference.py.

Intended Uses

Example of uses

• Procedural Pokémon-like content generation
• Latent space exploration
• Educational demonstrations of VAEs
• Creative tooling (fusions, variants)

Out-of-Scope Use

• Competitive Pokémon balancing
• Canonical or official Pokémon creation
• Real-world decision making

---

Evaluation

This model is not quantitatively benchmarked.

Evaluation is qualitative and exploratory:

• Plausibility of generated stats
• Smoothness of latent interpolations
• Diversity of generated forms

---

Limitations

• Only the first listed ability is modeled
• No hard constraints on stat realism beyond BST normalization
• Type correlations are learned implicitly
• Dataset biases directly affect generations

---

Ethical Considerations

This model generates fictional content inspired by Pokémon data.

It is intended strictly for educational and creative use.
All Pokémon-related concepts, names, and data are the property of their respective rights holders.

---

Reproducibility

Training

You can change training hyperparameters in config.json.

Start training:

python train.py

Saved artifacts:

Model weights
StandardScaler statistics
Type and ability vocabularies
Training configuration

Requirements

See requirements.txt:

PyTorch ≥ 2.1
scikit-learn
NumPy

Model Card Authors

Author: tiboitel Model name: poke-vae