I am making a four-class classifier and I have a dataset of four classes containing 10000 images in the training set and 8 in the validation set. So, I took 1000 images from the training set and 192 images from the training set, which are not used in training and add it to the valid set to make a proper ratio. Now when I had only 8 images in the validation set, it is giving me a good accuracy and valid loss was good too, but after adding 192 extra images, the loss decreasing has almost stopped and accuracy is about 55% max. I have tried with many learning rates, but nothing changing. Can you guys help me out to increasing the accuracy and decreasing the test loss?