Skip to first unread message
Hamid Bazargani
Feb 10, 2015, 10:15:56 AM2/10/15
to caffe...@googlegroups.com
Hello everyone,
I have a question about training process of convolutional neural network. I am trying to train a network's weights along with all feature maps for face detection purpose. My original training set includes 13k positive samples and 18k negatives. My question arises from bootstrap procedure for learning the network to reduce number of false positives. So every 1000 epochs, I reduce output threshold and apply the network on a set of background images. Samples with confidence greater than the threshold are fed to the training set to boost the performance. I reduced threshold from 1 , 0.9, 0.8, ...
and after threshold=0.3, I noticed that the trained network with added negative samples (hard negatives) even generates more false negatives compared with the previous state. I have this feeling that I might not add enough samples or might add too many . Now my question is, "Should I continue the training process for thrsh=0.2, with updated weights, or I should revert back to the previous state?"
"Is there any analytical way to determine the number of samples to be added?".
I am confused that why I don't get performance improvement after adding more samples. And what would be the possible solution in this case.
So I'd appreciate if anyone can share his/her feedback with me.
I have a question about training process of convolutional neural network. I am trying to train a network's weights along with all feature maps for face detection purpose. My original training set includes 13k positive samples and 18k negatives. My question arises from bootstrap procedure for learning the network to reduce number of false positives. So every 1000 epochs, I reduce output threshold and apply the network on a set of background images. Samples with confidence greater than the threshold are fed to the training set to boost the performance. I reduced threshold from 1 , 0.9, 0.8, ...
and after threshold=0.3, I noticed that the trained network with added negative samples (hard negatives) even generates more false negatives compared with the previous state. I have this feeling that I might not add enough samples or might add too many . Now my question is, "Should I continue the training process for thrsh=0.2, with updated weights, or I should revert back to the previous state?"
"Is there any analytical way to determine the number of samples to be added?".
I am confused that why I don't get performance improvement after adding more samples. And what would be the possible solution in this case.
So I'd appreciate if anyone can share his/her feedback with me.
Reply all
Reply to author
Forward
0 new messages