lab5

Let's Play GANs

In this lab, we implement conditional GAN to generate synthetic images in multi-label conditions. Given a latent vector $z$, which is randomly generated, and a dependent vector, we trained a generator to generate objects with specific colors and shapes; conditional GAN should do this.

Data

We are using CLEVR as our dataset. For dealing with the labels in the JSON file, we use get_iCLEVR_data() function in dataset.py. For images, we apply transformations that resize the image to 64x64, transform the value into a tensor and normalize the pixel values.

Training

Install dependencies

conda env create -f environment.yml
conda activate gan

Sample running code

python train.py \
    --batch-size 128 \
    --im-size 64 \
    --z-size 128 \
    --g-conv-dim 300 \
    --d-conv-dim 100 \
    --device "cuda:0" \
    --g-lr 0.0001 \
    --d-lr 0.0004 \
    --beta1 0.0 \
    --beta2 0.9 \
    --epochs 300 \
    --eval-iter 50 \
    --num-cond 24 \
    --adv-loss "wgan-gp" \
    --c-size 100 \
    --lambda-gp 10 \
    --net "sagan"

Results

• Model: Self-attention GAN
• Testing accuracy: 77.22%