Aug 31, 2018 · 1 min read
I have one question, do you have done, or have suggestions for deployment when the training was performed using transfer learning? without diving into details and specific tools, my question is related to the conceptual pipeline, for example, would it be:
- Creating a new computation graph being a concatenation of the pre-trained graph(in this case inception) and your specific trained graph.
- Run them separately : forward-propagate the image through the pre-trained model(inception) and capture the resulting vector in your “driver” program/script , then sending this vector as input to your specific trained graph