Using DCGAN to generate flowers that has never been sean before in data.
This project leverages the 102 Flowers Dataset. The dataset was prepared by manually selecting images that contain only one flower facing upwards. Images were cropped to a square shape, centered around the flower so that the center of the flower aligns with the center of the image, and resized to a resolution of 128x128 pixels. This approach ensured better results as the model could focus on the flower's shape and characteristics rather than background distractions or multiple flowers in a frame. Finding the right images did limit the final dataset size. To counteract this, some data augmentation techniques were applied.
The model used is DCGAN, with generator and discriminator architectures depicted below.
The criterion for training is BCELoss because it is effective for binary classification tasks inherent to the discriminator in GANs. The model was trained for 1000 epochs with a batch size of 128. Both the generator and the discriminator had a learning rate set to 0.0005. The latent space has a size of 300. To optimize performance, a learning rate scheduler was employed, reducing the learning rate by 0.9 every 50 epochs.
