This is a really simple patch that allows more flexibility with the size of the input tensor.
In the code example, the task used for the finetuning of the model takes a pair of sentences as input so the tensor is of shape (batch, sentence index, sequence length, data or position index). Now let's say I want to use the model to do language modelling, I would like my input shape to simply be (batch, sequence_length, data or position index) which is not possible with the current version of the code.
My commit fixes this. By using negative indices, the model can now take both tensor shapes as input.
thomwolf
commented
6 years ago
Hi,
This is a really simple patch that allows more flexibility with the size of the input tensor.
In the code example, the task used for the finetuning of the model takes a pair of sentences as input so the tensor is of shape
(batch, sentence index, sequence length, data or position index). Now let's say I want to use the model to do language modelling, I would like my input shape to simply be(batch, sequence_length, data or position index)which is not possible with the current version of the code.My commit fixes this. By using negative indices, the model can now take both tensor shapes as input.