Adversarial Feature Generation for Zero-Shot Learning

We try to generate synthetic data for unseen classes conditioned on class attributes. Training is done using wasserstein generative adversarial network.

Architecture Diagram

Loss Function

Train

python train.py --logdir run_dir --train-dir /data/CUBNew

Parameters:

• iterations
• batch_size
• dropout
• train_dir
• logdir
• z_dim
• g_steps
• d_steps
• lr
• wgan
• log_interval

Results

Data

Link - (https://drive.google.com/file/d/1XIFik0Cv1MTWtQEQQUZygOwL-X-ZajKY/view?usp=sharing)