Why won't PyTorch RNN accept unbatched input?

Question

I'm trying to train a PyTorch RNN to predict the next value in a 1D sequence. According to the PyTorch documentation page, I think I should be able to feed unbatched input to the RNN with shape [L,H_in] where L is the length of the sequence and H_in is the input length. That is, a 2D vector.

https://pytorch.org/docs/stable/generated/torch.nn.RNN.html

import torch

x1 = torch.tensor([[1.0], [2.0], [3.0], [4.0], [5.0], [6.0], [7.0], [8.0]])
x1_input = x1[0:-1, :]
x1_target = x1[1:, :]

rnn = torch.nn.RNN(1, 1, 1)

optimizer_prediction = torch.optim.Adam(rnn.parameters())
prediction_loss = torch.nn.L1Loss()
rnn.train()

epochs = 100
for i in range(0, epochs):
    output_x1 = rnn(x1_input)
    print(output_x1)
    error_x1 = prediction_loss(output_x1, x1_target)
    error_x1.backward()
    optimizer_prediction.step()
    optimizer_prediction.zero_grad()

However, PyTorch is complaining that it expects a 3D input vector (i.e. including a dimension for the batch):

RuntimeError: input must have 3 dimensions, got 2

What is the correct method for feeding unbatched input to an RNN?

Answer 1

I would recommend turning your input into a 3d array by adding a batch size of one with:

torch.unsqueeze(x1_input, dim=0).