A sample training procedure

1. use pretrained ResNet-50 model for fine tuning, retrain last 20 or so layers starting from 5a

   A. TODO: 
         a1. train with subset1, validate with subset2, test with test set; 
         a2. train with subset2, validate with subset3, test with test set; 
         a3. train with subset3, validate with subset4, test with test set; 
         a4. train with subset4, validate with subset1, test with test set; 
         a5. train with subset1&3, validate with subset2&4, test with test set; 
         a6. train with subset1&2, validate with subset3&4, test with test set; 
         record loss for each experiment, this is to verify that we have randomly divided the dataset and our network is robust
         
   B. TODO: 
         b1. train with full data set, validate with test data 
         b2. train with subset1&2&3, validate with test data 
         b3. train with subset1&2&4, validate with subset3
         record loss for each experiment, a) loss b) top5/3/1 accuracy
   C. TODO : experiment with different learning rate, step size, and SGD configurations and selectively repeat A.& B.
   D. TODO : experiment with different weight initialization, and selectively repeat A.& B.
   
2. train from scratch ResNet-50 model, repeat A. B. C. & D.(or use the best config from 1.) 
   a1-a6
   b1-b3
   c
   d

3. Add bounding box to model

4. stretch: play with Inception v4 net

NOTE: each task should be reproduced by 2 people on the team. Please add your name to the tasks.