Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/59322 Implement sparse feature dropout (with replacement) that can drop out individual items in each sparse feature. For example, the existing sparse feature dropout with replacement drops out whole feature (e.g., a list of page ids) when the feature is selected for drop out. This itemwise dropout assigns probability and drops out to individual items in sparse features. Test Plan: ``` buck test mode/dev caffe2/torch/fb/sparsenn:test ``` https://www.internalfb.com/intern/testinfra/testrun/281475166777899/ ``` buck test mode/dev //dper3/dper3/modules/tests:sparse_itemwise_dropout_with_replacement_test ``` https://www.internalfb.com/intern/testinfra/testrun/6473924504443423 ``` buck test mode/opt caffe2/caffe2/python:layers_test ``` https://www.internalfb.com/intern/testinfra/testrun/2533274848456607 ``` buck test mode/opt caffe2/caffe2/python/operator_test:sparse_itemwise_dropout_with_replacement_op_test ``` https://www.internalfb.com/intern/testinfra/testrun/8725724318782701 Reviewed By: Wakeupbuddy Differential Revision: D27867213 fbshipit-source-id: 8e173c7b3294abbc8bf8a3b04f723cb170446b96
