README

gpu-b01:/u02/wang/tf-dhl2
1. create_TFRecords.sh 
creates train.tfrecords and validation.tfrecords from ./images folder. 

2. train2.py
This is the latest and most advanced program. Do validation every 1000 step while trianing

3. check the scalers and images summary
tensorboard --logdir=./log/resized_train2/alexnet_v2_2 to check the training and validation summary of an Alexnet.
So far, Alexnet performs the best both in speed(cost 2000 seconds to train 15000 steps, around 4 times faster than vgg_16) and accuracy(93% on validation dataset)

4. actually, tf.resize_images can work well with neural networks. See the problem of display in the 'explore image_resizing' section in write_read_TFRecords.ipynb

Other basic program:
5. train.py 
start a training process with vgg_16
6. test.py 
checkout an checkpoint created during train.py and do the test.