https://pytorch.org/tutorials/intermediate/seq2seq_translation_tutorial.html

I don’t think we should change the shape of input, just keep it (seq,batch)

The input to `nn.GRU`

is defined as `[seq_len, batch, input_size]`

, which is why the `view`

is performed.

But the shape is 1 1 -1 which means 1 1 seq*batch*hidden.this makes no sense

2019年3月9日星期六，ptrblck via PyTorch Forums noreply@discuss.pytorch.org 写道：

The view operation reshapes the tensor to `[1, 1, "remaining size"]`

.

In case your tensor has 10 elements, `embedded`

will be reshaped to `[1, 1, 10]`

.

So the size is `1*1*seq_len*batch_size*hidden_size`

then the tensor will be fed into `GRU`

. But `GRU`

only accept tensor with size `(seq_len, batch, input_size)`

Basically yes, but since in the tutorial each word is fed ony by one, `seq_len`

and `batch_size`

will both be 1.

1 Like