Need tips for training

I was training a Resnet-101 network for a binary image classification task. However, the training seems to be stuck as the training/val accuracy stays at 50%. Previously, I used a pre-trained Resnet-101 on imagenet and with that my network can raise the accuracy up to 80% in 20 epochs using sgd optimizer with a learning rate of 0.01 and momentum of 0.9. (strangely, adam optimizer doesn’t work for this task)

Now, I want to train a network without the help of pre-trained weights of imagenet. However, training becomes much harder then and no matter how I tune the hyperparameter, the accuracy just stays at 50% - stuck at a local maxima I guess.

What is the best practice to get the network out of the local maxima? Any ideas for optimizer/hyperparameter tuning methods I should be using?

As one of the broad and widely used branches of artificial intelligence, machine learning explores the methods and algorithms by which computers and systems can learn and learn.
You probably will use car learning several times a day, even without knowing it. Every time you do an internet search on Google or Bing, machine learning is done because their machine learning software understands how to rank web pages. When Facebook or the Apple Photo app knows your friends and pictures, this is also a machine learning. Each time you check your email and get rid of spam spam from having thousands of spam updates again, it’s why your computer has learned to detect spam from non-spam email. This is the same learning machine. This is a science that makes computers learn from a specific topic without having to go through an explicit program.
A statistical survey of machine learning and specifically machine learning with monitoring.
Trevor Hastie, Robert Tibshirani, and Jerome Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Prediction, Springer, 2003 (ISBN 0-387-95284-5)
Machine Learning Thomas G. Diet Erich - Department of Computer Science - Oregon State University
A simple and fluid reference for learning enhancement; Suitable for learning the basics:
Richard S. Sutton and Andrew G. Barto, Reinforcement Learning: An Introduction, MIT Press, 1998 (online version).

This compact volume is one of the most important classic references in machine learning:
Tom M. Mitchell, Machine Learning, McGraw-Hill Companies, Inc. , 1997. ISBN 0-07-042807-7
With these resources, I collected a series of video tutorials for practicing these topics to view these videos . visit my blog :
پرستاری بدون کنکور مشاوره خانواده