search

fmassa / object-detection.torch

BSD 2-Clause "Simplified" License
116 stars 40 forks source link

Fine Tuning #21

Open mzahran001 opened 8 years ago

mzahran001 commented 8 years ago

How to fine tune your model? I don't have sufficient data to retrain your model from scratch.I want to fine tune your model on my data which has only two classes ?

fmassa commented 8 years ago

Hi,

If you don't have enough data, I'd advise you to train an SVM using the last fc7 layer from the network. There is a SVM trainer in here.

Using it will require you to convert your dataset to have labels in a format similar to Pascal VOC 2012 (or to write a dataset that has the same interface as DatasetPascal, which also includes computing bounding boxes.

brisker commented 7 years ago

@fmassa Hi, after I trained a model from scratch, I found that the epoch seems not enough, so I reload the model and try to finetune it for a bit more epoches. But why the training loss seems like training from scratch?? I just use this code: model=torch.load(...) local parameters, gradParameters = model:getParameters() ... The other codes are just the same as the code for training from scratch. Why this happens? I am really confused about this.