I want to Reduce communication volume by sending only top k% of model parameters . Is there a way in which i can send those specific model parameters ?
1 Like
Hi @wehtam , thanks for your question. Yes you could select partial parameters of the model for communication. For example, in the quickstart-pytorch example, you could modify the get_weights() function to select specific parameters.
Yes I understand but if i need to work with weights at server side for aggregation etc , how can i reconstruct the full parameters ( by replacing the weights that were not sent by zeroes for example) ?