My excise with Stanford Dogs Dataset

Did any one try with the Stanford Dogs Dataset, at http://vision.stanford.edu/aditya86/ImageNetDogs/? I tried this with resnet50 with 8 epoch of training. The final error rate I got is 0.104. Yet I am not sure how this compares to the result at the link above. They used something called mean accuracy. Can somebody explain the relation between mean accuracy and error rate?

|epoch|train_loss|valid_loss|error_rate|time|
|0|0.993842|0.516923|0.151883|05:10|
|1|0.789177|0.456484|0.145079|05:07|
|2|0.685540|0.430629|0.133414|05:07|
|3|0.613375|0.385798|0.121021|05:07|
|4|0.486800|0.376321|0.116646|05:07|
|5|0.407328|0.362916|0.109842|05:07|
|6|0.327958|0.340387|0.106197|05:07|
|7|0.302153|0.339336|0.104010|05:07|

Some other questions during my learning progress:

  1. After above mentioned 8 epoch of training, I used lr_find to find the learning rate, and plot that. Then if I do learn.save and learn.load, Did I get everything restored? I did an lr_find and plot again after learn.load, I can see the graph plotted is totally different with the first one.
  2. After I run lr_find, I run learn.unfreeze(), and learn.fit_one_cycle(4, max_lr=slice(1e-6,1e-3)). Yet as listed below, the result is not improving at all. This is even true for the dog and cat dataset in lesson 1. Can anyone explain that?

|epoch|train_loss|valid_loss|error_rate|time|
|0|0.421895|0.439848|0.131956|06:48|
|1|0.390584|0.419171|0.126853|06:47|
|2|0.279158|0.387257|0.114702|06:47|
|3|0.207735|0.369129|0.107655|06:47|

Also what should I do, if I want to further improve the result?

Thanks in advance for your time/support.

regards, Weizhong