search

afislonge / deepfake-detection-project

Our project aims to develop a machine learning model to compare and contrast images sourced from Flickr, a platform with user-uploaded photographs, against those generated by ThisPersonDoesNotExist, which uses generative adversarial networks (GANs) to create lifelike human images. We'll evaluate the nuances in authenticity and quality
3 stars 1 forks source link

VGG19 model design #40

Open JuanS286 opened 6 months ago

JuanS286 commented 6 months ago

Definition of a model leveraging transfer learning, in this case using the VGG19 model.

JuanS286 commented 6 months ago

First model design was made today. Also I defined the sample data to train and test the models, the data is in the "Sample" folder inside the Drive folder of the project. I arrange the data in train, validation and test folders and inside each one of them there are two folders, 'Fake' and 'Real'. I set 600 images for testing, 200 for validation and 200 for testing with equally balanced classes.

The model is currently overfitting but there is a foundation and some stuff can be done to improve it, also having in mid that it is currently working with a sample of the data.

I am going to commit it in its current version to the repository, but you can also find the notebook inside the Colab notebooks folder inside Drive.

JuanS286 commented 6 months ago

Today I trained the second model, using a sample of 2000 image by class. The model overfits less but the accuracy is not great.

Code is committed as a jupyter notebook and html for submission purposes.