Fine tuning a model in transfer learning

Question

I am working on a deep learning model for "Age Invariant Face detection". I started with preprocessing the data which included face cropping via MTCNN, image contrast correction and sharpening. I then moved to build the model architecture, I designed a custom sequential CNN model. The training and validation losses were close to 1.9 but the model was taking approx 8 hrs for one epoch. I needed fast results so I moved to try on transfer learning. I read a paper which said Resnet-18 performed well on the task so I went with Resnet-18 first. Tried on tuning it by varying learning rates, but for no good. Accuracy wasn't increasing beyond 40%. I then went with Resnet-50, currently, I am working on it. I tried fine-tuning by changing the learning rate manually, tried with different optimizers Adam, SGD. Used learning rate decay. My model started to overfit so then I added dropout, image augmentation to training data and again added weight_decay as it was still overfitting. Right now there is no overfit but the losses aren't decreasing. I am trying on doing my experiments over it, reading many articles per day but feels like I'm stuck.

Answer 1

Apparently, this is one of the very good cases where feature engineering comes to the light. Try as much as you can to make your images use less computational power by apply some classic feature engineering methods. Using MTCNN is a really good approach and I would deffo look into Feature Engineering, also here is a link that could provide you some AHAA moments performace wise. All the best