Howard

Add compatibility of distributedwrapper for two stream input, so wrapped miner and loss can also use forward as `self.miner(query_embed, labels, doc_embed, labels.clone())` and `self.loss(query_embed, labels, doc_embed, labels.clone())`

Does DistributedWrapper support two stream input?

3

Hi, how does the DistributedWrapper support two stream input? I see the original implementation DistributedWrapper class only has three input with `def forward(self, embeddings, labels)` , but I want something...

Hi, I don't get why bottleneck is tiled with dimension of batch size here. Wouldn't there be cases that batch size can not be perfectly split, so the last dimension...