I've uploaded the partial gnt dataset, you can transfer it to any form as you need. Here, you can use resnet_tensorflow1.12/tfrecord_op/gnt2tfrecord.py to convert gnt to tfrecord within one minutes. After activating tensorflow-gpu, you can print python gnt2tfrecord.py -h to get the helpful information about parameters.
When you want to generate the training set, it's matter that to set the shuffle=1, or the network will be hard to converge.
If you are in Windows, Ubuntu or other systems which have the GUI, you can verify tfrecords by printing the images and theirs label. The python document involved is resnet_tensorflow1.12/tfrecord_op/test_tfrecord.py.
All of the three sections were wrote in ./main.py, and you can print python main.py -h to get the helpful information.
python main.py --mode train
python main.py --mode validation
Here, it needs you have the GUI in order that you can install pillow.
python main.py --mode inference --png_dir xxx
tensorboard –logdir=./log3755/train
Firstly, you can change the learning_rate, decay_step, optimizer in ./main.py, I suggested that you can set them to 0.0001, 0.9, adam. Secondly, weight_decay, batch_norm_decay, batch_norm_epsilon in ./model/resnet_utils.py also can influence the accuracy.