How should I do to train a compressed or little model on my own data?

[beihangzxm123]

I wanna train a compressed or little model on my own data whose annotation format is the same as voc dataset but has different classes with VOC,i just use the command:"./tools/train_net.py --gpu 0 --solver models/pvanet/example_finetune/solver.prototxt --weights models/pvanet/comp/test.model –iters 100000 --cfg models/pvanet/cfgs/train.yml –imdb voc_2007_trainval", is it a correct way? I found out that the performance of the detection with this model is so bad.The finetuned model using "models/pvanet/full/test.model" has almost the same bad performance that has so many false detections .

How should I do to train a compressed or little model on my own data?

Please give me some detailed advice,thanks so much!Looking forward to your reply!

[sanghoon]

Hi @beihangzxm123 , I'm not sure how big your dataset is, but if it's not that big I recommend you do ...

• start from the up-to-date PVANet detection model (VOC2007, VOC2012 or compressed) (there used to be a bug in the example prototxts)
• train the model for longer iterations (we've trained the models for more than 1M iterations)

[runningJ]

hi @beihangzxm123
Are you resolve the problem?

[MyVanitar]

@sanghoon @runningJ

I have almost the similar questions here #63 Please have a look

[pasxalinamed]

I have trained a model on my own data. How can I compress the output caffemodel? Is compression something that should have taken place during training?