Is the attached workflow correct in case of training and testing with different data sets? Is the attached workflow considered "transfer learning" since my images are not related to the images that the network was pre-trained?
1 Answer
Your workflow is correct and you are doing a type of transfer learning where only the last layer is fine-tuned. Image embedding widget calculates features that are the activations of a penultimate layer of Inception-v3, VGG or Painters network. All of those networks are trained on the different set of images. VGG and Inception-v3 are trained on ImageNet dataset and Painters network is trained on the set of artworks. In the transfer learning you are implementing with the workflow you fine-tune all of the connected classifiers for your images.