I would like to perform both the “mean” and “sum” method on one BoW feature. Docs show that `EmbeddingBag`

could be initialized with the same weights using the `from_pretrained`

method. But it seems the underlying `THPVariable_make_subclass`

function would initialize different weight tensors into the computational graph.

The most straightforward way to share weights might be to use the functional interface for the “second invocation”, where you replace the `self`

in the call to `F.embedding_bag`

taken from the forward:

```
def forward(self, input, offsets=None, per_sample_weights=None):
# type: (Tensor, Optional[Tensor], Optional[Tensor]) -> Tensor
return F.embedding_bag(input, self.weight, offsets,
self.max_norm, self.norm_type,
self.scale_grad_by_freq, self.mode, self.sparse,
per_sample_weights, self.include_last_offset)
```

That said, I would imagine that for the same *features*, it’s more efficient to use Embedding + sum/mean manually (or even just use EmbeddingBag with sum and scale yourself to go from sum to mean?)

Best regards

Thomas