Hi guys, after reading chapter 2 I’m really excited to start a similar mini project and start blogging about it. Can anyone tell me what how many images I need to successfully train my model. I’m thinking about making a fun dog or cat emotion classification recognition model. However, I don’t think there’s an image dataset for this so I would be locating and labeling the images myself.
Hi AyersAuthentic hope your having a wonderful day.
I like a quote from Jeremy which is “whenever your not sure about something, try it, experiment”
There are many factors involved, I have seen a sports racket image classifier with as few as 35 images per class with accuracy’s of over 90% this sports bat or racket classifier, could tell the difference between things such as tennis rackets, cricket bats, table tennis bats and other sporting rackets or bats.
These items differed quite significantly in appearance. The emotions of cats and dogs looks like quite a tricky one. I have a dog and I think it may be quite difficult to classify their expressions.
For example my research into dogs, has shown dogs with very dark and long hair get attacked more by other dogs, because other dogs find it difficult to tell what expressions they are showing, as the long dark hair obscures their expressions.
Start with say 200 images per class and increase the number of images and increase the size of the transfer model used e.g. start with resnet34 and try resnet50 and see if your results improve.
Can’t wait to see your blog!
Cheers mrfabulous1 
3 Likes
Thank! @mrfabulous1! All this is great info. I really appreciate the help. I’ll let you know when the blog gets posted.