In the ResNeSt paper they say on page 4:

"despite their great success in image classification, the meta network structures are distinct from each other, which makes it hard for downstream models to build upon." What are downstream models in this context?

Paper can be found at https://arxiv.org/abs/2004.08955


Downstream models are simply models that come after the model in question, in this case ResNet variants. Models for various topics within the computer vision domain often use a backbone to extract features from images, after which a downstream model is used to help to fit the model better to the task at hand. Tables 5, 6, and 7 in the linked paper give a good overview of the different ways backbones are often combined for topics such as object detection and segmentation.

  • $\begingroup$ Thanks! I have two questions: 1. When we use for example ResNet as the base model and use a downstream model on top of it, are ResNets parameters tuned during training as well? 2. What happens to the last fully connected layers in ResNet, are they cut? $\endgroup$ Aug 2 '20 at 17:32
  • 1
    $\begingroup$ There is no real answer to the first question as each authors can make those choice freely, however to my understanding it is common to first freeze the backbone and train the downstream model for a number of epochs, after which the model is trained end to end. Regarding the second question, the last fully connected layer is almost always cut but more layers can cut off depending on the model choice by the author. $\endgroup$
    – Oxbowerce
    Aug 2 '20 at 17:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.