When fine-tuning a Resnet50 model should I do the standard train validation split and train like any normal CNN or should I just do the training and not the validation if I am going to use this model as a base model in another more advanced model? I think I will need to do the train/val split but wanted to get a good answer for this from other people as I have seen on Github people just training with no validation. Thanks for all the input.
While using an already trained Resnet50 model or in any other Transfer learning cases, you have to train the model on your dataset but not from scratch, just remove few last layers of the pre-trained model and froze the remaining layers. Then you should train the model as you normally would with cross validation.