Fine-tuning GoogLeNet

[Steven Clark]

Hi-

 

When fine-tuning the googlenet models, how do you handle the first 2 auxiliary loss classifiers? Do you reduce their learning rates, or delete them entirely?

 

I was fine-tuning for a 3-class classifier, and did the following:

 

in quick_solver.prototxt:

  -  decrease base_lr by a factor of 100

  -  decrease max_iter to 10,000

 

in train_val.prototxt:

  -  delete all loss1/* layers

  -  delete all loss2/* layers

  -  rename loss3/classifer layer to loss3/classifier_mod

  -  increase loss3/classifier_mod's learning rates by a factor of 10 (from 1, 2  to 10, 20)

 

Loss decreases and test top-1 accuracy hits ~97%, but I'm wondering if my approach is optimal/sensible.

 

cheers,

-Steven

[Ashutosh Singla]

Hi, 

I am using GooLeNet model for image classification and now I want to use pre-trained model. Do you know how can I use pretrained GoogLeNet model for fine tunning? What are the steps?

[beau norgeot]

Hi Steven,

I haven't tried fine-tuning googlenet through caffe. I think your approach is generally solid, however I'd be be concerned about time to convergence. You reduced the base_lr by a factor of 100 which will help to 'freeze' the weights that you want to keep. However, by increasing the lr for the loss3/classifier layer to 10/20 I think you're going to be learning parameters for your new data very slowly. If the original base_lr was 0.01, you've decreased that to .0001, for loss3 you've increased that back to .001 -- you're learning 10x slower than the base. If your learning is peaking you might consider bumping that rate up by a couple orders of magnitudes.

Have you tried comparing your caffe results with retraining using tensorFlow? (As a sanity/check baseline comparison)

-Beau