Art_GAN

Inspiration

I've started to take a lot of data science machine learning type courses and wanted to start to dig deeper into machine learning and challenge myself to scrap together my own data to build my own neural network architecture.

What it does

Simple web front end where you can generate a new art image from the art generator.

How I built it

I found about 100,000 artwork images online which was stored in about 5.5GB. I first compressed this data down to 100x100 pixel RGB images which compressed the data down to about 350MB and reduced it to a dimensionality space where we had enough data to learn a model for the feature space. I then put a more simple neural network architecture in google colab to help me test my javascript front-end I would use to load in my tensorflow NN model and generate an input space of a gaussian distribution to predict the output for the model. Once I was satisfied with my javascript front-end I made the GAN with the knowledge I gained from my classes and reading over research papers. I edited the number of layers/type of layers for the network until I reached a point where the discriminator was not learning too fast and the generator had a complex enough model to learn something from trying to trick the discriminator.

Challenges I ran into

Producing the input space array in a javascript array. Loading my python Tensorflow model into javascript with Tensorflow js. Figuring out what sized of data points I should reduce my feature space too. Finding the right learning rates and model complexity to get somewhat decent output from the model. I think there is still a lot of fine-tuning that could be done to make the model better and it could benefit from using AWS or Google Cloud so I am not limited to the amount of RAM that I have in Google CoLab.

What I learned

How to build a neural network architecture with your own data.

Built With

• colab
• javascript
• jquery
• python
• tensorflow