Image normalization values about squeezenet HOT 1 CLOSED

Good question. It's been a while since we did this work, but here is what I remember of how we prepared the training data:

• Warp each image to 256x256
• During training, crop a random 227x227 patch in each 256x256 image (different patch in each epoch), and feed the crop into the network. I believe we also did data augmentation by flipping the image in the horizontal dimension with a probability of 0.5.
• During inference, crop the center 224x224 pixels.
• Note that SqueezeNet has average pooling at the end, so you could actually feed in almost any image size if you want.
• As for normalization, I believe we calculated the average R, G, and B values, and we subtracted that value from the input data. Some people do a normalization grid (so R,0,0 would have a different value than R,0,1 in the normalization grid), but we just computed 3 scalar normalization values.
• I don't think we did any other data augmentation or normalization techniques. The field has come a long way in the last 4 years, and I think modern techniques for almost every aspect of training and inference would be superior to the approach we took in SqueezeNet.

Good luck! I am going to close this issue, but you can still reply here if you have more questions.

from squeezenet.