search

google / automl

Google Brain AutoML
Apache License 2.0
6.24k stars 1.45k forks source link

VERY VERY LOW AP VALUE AND OTHER EVALUATION INDEXES ON CUSTOM DATASETS #312

Closed qtw1998 closed 3 years ago

qtw1998 commented 4 years ago

Low params size caused THE Hyperparameter sensitive, so the very little fluctuation in params settings would bring a low ap result.... Powerful computing sources are needed to find the proper settings.🤯

mingxingtan commented 4 years ago

Hmm, I am not sure about the details, but because the #params is much smaller, you generally need to train longer.

qtw1998 commented 4 years ago

Hmm, I am not sure about the details, but because the #params is much smaller, you generally need to train longer.

image I've trained it for quite a long time 3days plus.... and the loss never get down below 0.4 image

qtw1998 commented 4 years ago

Hmm, I am not sure about the details, but because the #params is much smaller, you generally need to train longer.

could you please share some tricks to attach better training courses like hyperparams settings like anchor size changed after measuring the datasetting by kmeans.

denisb411 commented 4 years ago

Hmm, I am not sure about the details, but because the #params is much smaller, you generally need to train longer.

could you please share some tricks to attach better training courses like hyperparams settings like anchor size changed after measuring the datasetting by kmeans.

Did you got better results by changing some hparam? Could you please share with us? I'm having the same 'issue'.

zishanahmed08 commented 4 years ago

@qtw1998 Any suggestion on hparam changes?

qtw1998 commented 4 years ago

@qtw1998 Any suggestion on hparam changes?

I've changed to pytorch version..