NEED a help on Fine Tuning Caffe Training - I'm stuck on "over-fitting"

[hamlet11ito]

I'm training the Caffe for new 20 classes of my own as the FineTuning training,  the same  scenario as the "Filckr_style" (i.e. using the already-trained parameter with ilsvrc2012 as the initial sets of model/weight parameters).  

                                        Seems like it being stuck as "over-training/over fitting":

(1)The first 1000 iteration is great: accuracy is 0.59 (and  loss went to down to 0.17 level from the initial value of 3.7). Then after, the accuracy reduces to 0.15 level and stay around the similar value  even at 20000 iteration.

Prediction for the new classes is not too bad for some classes, if I use the snapshot at 1000 iteration for ".caffemodel" file

If I use the snapshot at the 20000 iteration, the prediction for every class is approximately equal. i.e.garbage! This makes me to conclude as  "overfitting".

                                        Any one has similar experience?  Please let me know how did you resolve this.    

(2) I would like to try training from scratch but as  "finetune model".  What kind of weight file I should use instead of "bvlc_reference_caffenet.caffemodel" below?

 ./build/tools/caffe train -solver models/finetune_flickr_style/solver.prototxt  -weights models/bvlc_reference_caffenet/bvlc_referenece_caffenet.caffemodel - gpu 0

Thanks in advance for your help!