DenseNet-121 with finetuning achieves an overall classification accuracy of 94.17% while the model built from scratch is only 82.4% accurate.
Keras provides a relatively simple and fast implementation of neural network procedures while PyTorch requires low-level data and model manipulation.