Hello again.
I have started to play with old movies (Bw, Noirs...Sci-Fi/Horror B movies) and I have to say that collecting "good samples" is kind of difficult. (The faces and the expressions of the actors in the old times were kind of cliche: Try to find Bogey smiling or laughing...)
So... I was wandering if the graph of the raw losses can indicate -in some way- how good the training set is:
For examples... If I have a lot of spikes or -to say it in another way- a big oscillation around the smoothed loss... is this a sign that the trainer has issues learning and need more quality data?
I am still playing around with the software as a newby... and my only experience with neural network was at the university....30 years ago, so... try to be patient.
So... I was wandering if the graph of the raw losses can indicate -in some way- how good the training set is:
For examples... If I have a lot of spikes or -to say it in another way- a big oscillation around the smoothed loss... is this a sign that the trainer has issues learning and need more quality data?
Not really. The loss is just a measure of how the NN thinks it is recreating data. "Quality" is a bit of a subjective term, so hard to use in a real world context.
You could write something which parses an alignments file and categorizes pose and expression and looks for gaps in data. This won't tell you if you are using high quality images, but should tell you if there are gaps in your data
