How to pass y_labels or scores to train model
deshwalmahesh opened this issue · comments
I want to mimic your model but using TID2013 dataset. I have built the model but do not know how to train the model. I use ImageDataGenertor of Keras
If I train the model using original scores with 1 output neuron, it'll be a regression problem and it'll be bad.
If I train using round(score) with sparse input and 10 output neurons, it gives me inf loss
If I train using categorical or raw, same thing happens. How should I pass the training y_scores to the model?
Hi, what we did for TID2013 is to infer a probability distribution from the reported mean score. Then we passed the inferred distribution as labels throughout training with an EMD loss function. I also experimented with a regression approach using just the mean score as you mentioned which worked quite well.
@clennan If I go with the regression approach and reduce the number of output neurons to 1, should I go with the linear activation function? Which function did you use in your approach?