cynstl/atelierai-models

AtelierAI Trained Model Checkpoints

This repository hosts the trained PyTorch model checkpoints and evaluation metrics for AtelierAI: Text-to-Conditional-Fashion Synthesis Studio.

Models Included

1. Conditional GAN Generator (cgan_generator.pt):

   • Architecture: Deep Convolutional GAN with Projection Discriminator (Miyato & Sugiyama, 2018)
   • Conditioning: 10 Fashion-MNIST apparel classes from latent noise $z \sim \mathcal{N}(0, I)$ (100-dim)
   • Resolution: $1 imes 28 imes 28$ grayscale
   • Empirical Fidelity: 97.20% Oracle Match Rate (Tested on 1,000 samples)
   • Training: 25 epochs on NVIDIA GeForce RTX 3060 CUDA

2. Conditional VAE (vae.pt):

   • Architecture: Convolutional Conditional Variational Autoencoder
   • Latent Space: 32-dimensional continuous Gaussian prior
   • Empirical Fidelity: 74.90% Oracle Match Rate
   • Training: 25 epochs with BCE reconstruction and KL-divergence loss

3. Auxiliary CNN Classifier (fashion_classifier.pt):

   • Architecture: 4-layer Convolutional Neural Network with BatchNorm & Dropout
   • Test Accuracy: 90.76% on real Fashion-MNIST test dataset

4. Fine-Tuned DistilBERT (distilbert_fashion/):

   • Base Model: distilbert-base-uncased
   • Fine-tuned on natural language apparel prompt intent classification

Licensing & Attribution

• Fashion-MNIST: Zalando SE (MIT License)
• AtelierAI Checkpoints: MIT License