Transferring Discriminator weights to CNN for image classficiation (Microsoft BIG 2015 challenge)

Question

We are currenyly working with BIG 2015 dataset, where we have created black and white images using the OP code files given. Then we have used a generator to create similar looking images to the actual dataset images. We have used discriminator in tandem to improved the generated images. Now we have to create a CNN to which is trained on original + generated images, then classify unseen data into 10 classes of malwares.

The CNN should have pre-trained weights from discriminator and I am looking for a way on how to do that?

Our images are 10873, 128x128 images which are B/W.

We have trained our discriminator and basically added the layers into CNN (not weights). And we are getting accuracy of 30%. We need help to reduce the overfitting as the train-test split is giving 96% accuracy.