Boosting Generative Networks for Incremental Learning

Training a ResNet classifier using pre-trained BigGAN models

[Report]

Using the base implementation of BigGAN provided by Andrew Brock. Achieves ~88.5% accuracy on CIFAR-10 (when using data transformations and classifier filtering), versus ~94.3 when trained using real data for the same number of optimization steps. With five GANs, reaches ~91% accuracy.

Summary

• Report
• Usage
• Reproducing results
• Scripts
• References

Report

Click HERE to download the final research report.

Usage

Training a classifier

Training a classifier requires:

• pre-trained classifier weights in: ./classifier/weights/model_name.pth
• pre-trained BigGAN weights in: ./weights/weights_folder_name/

To run the script:

python3 train_classifier.py [options]

Parameters are as follows:

Input/Output

• model: Weights file to use for the GAN (of the form: ./weights/model_name/G_ema.pth if single GAN, ./weights/model_name/gan_multi_n/G_ema.pth if n GANs are used)
• classifier_model: Weights file to use for the filtering classifier (of the form: ./classifiers/weights/class_model_name.pth)
• ofile: Output file name (default: trained_net)

Training

• batch_size: Size of each batch (same for generation/filtering/training, default: 64)
• num_batches: Number of batches per class to train the classifier with (default: 1)
• epochs: Number of epochs to train the classifier for (default: 10)

Classifier filtering

• filter_samples: Enable classifier-filtering of generated images (default: False)
• threshold: Threshold probability for classifier filtering (default: 0.9)

Multi-GANs

• multi_gans: Sample using multiple GANs (default: None, integer value)
• gan_weights: If using multi-GANs, specify weights for each GAN (default: sample from each GAN with equiprobability)

Other

• truncate: Sample latent z from a truncated normal (default: no truncation, float format).
• fixed_dset: Use a fixed generated dataset for training (of size: batch_size*num_batches*num_classes, default: False)
• transform: Apply image transformations to generated images (default: False)

Sampling from GAN weights

To sample from GAN weights:

python3 sample.py [options]

Parameters are as follows:

Input/Output

• model: Same as above
• ofile: Output file name (default: trained_net)
• torch_format: Save NPZ images as float tensors instead of uint8 (default: False)

Generation

• num_samples: Number of samples to generate (default: 10)
• class: Class to sample from (in [[0,K-1]] for K classes, default: sample sequentially num_samples/k for all classes.)
• random_k: Sample classes randomly (default: False)
• multi_gans: Generate samples using multiple GANs (default: None, integer value)

Other

• transform, truncate: Same as above

Scripts

Some bash scripts are already in the folder ./scripts/, to run classifier training sessions with various parameters.

References

1. [Brock et al., 2018] Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale GAN training for high fidelity natural image synthesis, 2018.

2. [Pham et al., 2019] Thanh Dat Pham, Anuvabh Dutt, Denis Pellerin, and Georges Quénot. Classifier Training from a Generative Model. In CBMI 2019 - 17th International Conference on Content-Based Multimedia Indexing, Dublin, Ireland, September 2019.