Hi, every time I use PyTorch torchvison.module, I need to download the pre-trained model. However, I feel it does not work very well for me, and I am thinking to train my own pre-trained. But I guess the image does not need to be very high. If It does not hit the improvement, I can make my dataset like 5 or 3 times smaller. I use my dataset not only for classification.
Most of the classification models were trained with a spatial size of 224x224, with the exception of Inception, which was trained with 299x299.
What kind of dataset are you using which yields a bad performance?
Thank You for telling me that. It is a Japanese anime dataset. But the problem is when it is being used for coloring, the image-net won’t work for coloring. I have seem this on my work, and I can color the object with my pre-trained model only but not the image-net.