Latent Interpolation on MNIST

I found doing linear interpolation in the latent space quite gratifying to watch. So I have created this respository which allows you to create a countdown or countup GIF with MNIST images using smooth transitions that comes from the latent space.

Sample Output

You can choose any range of numbers from 0-9 to create the GIF.

Countdown 9-0	Count Up 0-9	Countdown 5-0	Count Up 0-5	Countdown 9-5	Count Up 5-9

Quick Usage

You need to first download this pre-trained model to begin generating your own GIFs.

Then simply run the following:

python main.py --model-path <path-to-saved-model>

You can find below a detailed list of the available command line arguments that you can use to customize your GIF.

CLI Arguments

Training Arguments

Long	Short	Description	Default
--train	-t	flag to train model from scratch	False
--epochs	-e	number of epochs used for training	250
--batch-size	-b	batch size used in training	400
--save-interval	-s	save model every interval	50

Generator Arguments

Long	Short	Description	Default
--generate	-g	flag to generate gif from an existing model	True
--imsize	-i	image size of generated gif	128
--start-num	-a	count down/up starts with this number	9
--end-num	-z	count down/up ends with this number	0
--n-steps	-n	number of steps used in latent interpolation	30
--out-path	-o	output folder path to save generated gif	./output
--out-name	-l	name of generated gif	countdown.gif
--model-path	-m	path to pre-trained Keras model	models/model-250.h5