Pre-trained models

  1. while using cnn pre-trained models we are resizing the image, does this leads to some data lost
  2. can we not edit the models to work with small size of images
  1. We expect that there would be no loss while applying RandomResizedCrop.
  2. Yes we can do so.

can you please explain 2nd point

You can edit the models to work with different size of images but then you will have to change all the configuration of the model , like its padding , strides so that no loss of data happens.(output pixel = ((Input pixel + -Filter size + 2P)/S +1 ) And then you also will have to change the whole FNN/classification layer because the input number of the first linear layer will change because of different pixels , so instead we just prefer to change the pixels of the image rather than changing the whole model.

1 Like