README.md

metadata

datasets:
  - marcelhuber/KermanyV3_flattened_postprocessed
language:
  - en
tags:
  - medical

KermanyV3 GAN Model

Description

The KermanyV3 GAN is a Generative Adversarial Network trained on the KermanyV3_flattened_posptrocessed dataset to generate synthetic medical images.

Files Included

• fakes000840.png: Sample of synthetic images generated by the GAN at snapshot 000840.
• log.txt: Log file containing training progress and details.
• metric-fid50k_full.jsonl: JSONL file with FID (Fréchet Inception Distance) metric scores for evaluating image quality.
• metric-kid50k_full.jsonl: JSONL file with KID (Kernel Inception Distance) metric scores for evaluating image quality.
• network-snapshot-000840.pkl: Snapshot of the GAN's trained parameters at iteration 000840.
• stats.jsonl: JSONL file containing statistics and metrics related to the training process.
• training_options.json: JSON file detailing the training hyperparameters and options.

Dataset Used

KermanyV3_flattened_posptrocessed

Model Performance

• FID Score: 4.954847615273339
• KID Score: 0.003050379662162

Training Details

• Training Iterations: 840
• Training Time: 8h 40m
• Architecture: StyleGAN2-ADA-PyTorch
• Batch Size: 16

Usage Instructions

1. Load the pre-trained GAN parameters from network-snapshot-000840.pkl.
2. Use the generator to generate synthetic images.
3. To evaluate image quality, refer to metric-fid50k_full.jsonl for FID scores and metric-kid50k_full.jsonl for KID scores.
4. For additional training information, refer to log.txt, stats.jsonl, and training_options.json.