Hello Chen,
Thank you for your sharing of the code. I'm watching your code and paper. From your code, the probe_num_query data points from the target label are sampled according to their similarities with pre-selected data points of the target label (in the support set), while the is_sort_query is true. From the _sort_proto_similarity of the current version, you use one pre-selected data point to calculate the similarities and comment on one sentence using all the pre-selected data points (first two lines of the function). As similarity calculation is designed for all the pre-selected data points, I'm wondering which is the case you run in your experiments. In other words, how many pre-selected data points of the target label are used for finding the probe_num_query data points of the target label?
Hello Chen,
Thank you for your sharing of the code. I'm watching your code and paper. From your code, the probe_num_query data points from the target label are sampled according to their similarities with pre-selected data points of the target label (in the support set), while the is_sort_query is true. From the _sort_proto_similarity of the current version, you use one pre-selected data point to calculate the similarities and comment on one sentence using all the pre-selected data points (first two lines of the function). As similarity calculation is designed for all the pre-selected data points, I'm wondering which is the case you run in your experiments. In other words, how many pre-selected data points of the target label are used for finding the probe_num_query data points of the target label?