Training Variational Autoencoder using celebA and MNIST dataset.
To train celebA dataset, download imgimg_align_celeba.zip and list_eval_partition.txt, unzip, place it under data/celebA/.
For training, you can modify hyperparameters e.g. learning rate, latent-dim, batch size, lr decay schedule...
(Take a look at parser_utils.py).
Learning curve will be saved at "{model}{dataset}.csv" by default (in this case e.g. cnn_celebA.csv).
While training, reconstruction results for the test dataset are saved in results/{dataset}/{model} direcotry, (e.g. results/celebA/cnn).
Pretrained model is provided in "{model}{dataset}.pt" format (e.g. cnn_celebA.pt).
python train.py --model cnn --dataset celebA --num-epochs 300 --latent-dim 128
Original image on the top, reconstruction on the bottom.
Generation (decode) from random Gaussian distribution.
python generate_samples.py --model cnn --dataset celebA --saved-path cnn_celebA.pt --latent-dim 128
python plot_learning_curve.py --loss Loss --learning-curve-csv cnn_celebA.csv
python plot_learning_curve.py --loss Reconstruction_Loss --learning-curve-csv cnn_celebA.csv
python plot_learning_curve.py --loss KL_Loss --learning-curve-csv cnn_celebA.csv
| Loss (Reconstruction + KL) | Reconstruction Loss | KL Loss |
|---|---|---|
 |
 |
 |
(Both training loss and Test loss decrease, I guess you can train more epochs if you want)
python train.py --model fc --dataset celebA --num-epochs 300 --latent-dim 128
Original image on the top, reconstruction on the bottom.
Generation (decode) from random Gaussian distribution.
python generate_samples.py --model fc --dataset celebA --saved-path fc_celebA.pt --latent-dim 128
python plot_learning_curve.py --loss Loss --learning-curve-csv fc_celebA.csv
python plot_learning_curve.py --loss Reconstruction_Loss --learning-curve-csv fc_celebA.csv
python plot_learning_curve.py --loss KL_Loss --learning-curve-csv fc_celebA.csv
| Loss (Reconstruction + KL) | Reconstruction Loss | KL Loss |
|---|---|---|
 |
 |
 |
python train.py --model cnn --dataset mnist --num-epochs 1000 --latent-dim 16
Original image on the top, reconstruction on the bottom.
Generation (decode) from random Gaussian distribution.
python generate_samples.py --model cnn --dataset mnist --saved-path cnn_mnist.pt --latent-dim 16
python plot_learning_curve.py --loss Loss --learning-curve-csv cnn_mnist.csv
python plot_learning_curve.py --loss Reconstruction_Loss --learning-curve-csv cnn_mnist.csv
python plot_learning_curve.py --loss KL_Loss --learning-curve-csv cnn_mnist.csv
| Loss (Reconstruction + KL) | Reconstruction Loss | KL Loss |
|---|---|---|
 |
 |
 |
Using t-SNE, maps 16 dimensional latent space to 2D.
python plot_latent_space.py --saved-path cnn_mnist.pt --latent-dim 16 --dataset mnist --model cnn
