PyTorch RNN, many to many learning, one to many test

vision
1

Hello! I need to create recurrent NN, during the process of training I have all the x_i and y_i

IMG_4517
IMG_45172238×1021 195 KB

but for test I have only x_1 and want to use y_1 as x_2, y_2 as x_3 and etc. How to do it via torch.nn.RNN?

2

Assuming your hidden dimension is the same size as your input dimension, you can do it like so:

rnn = torch.nn.RNN(16, 16)
x1 = torch.randn((1, 1, 16))

y1,h1 = rnn(x1)
y2,h2 = rnn(y1, h1)
...

The output of the RNN is a sequence of outputs produced and the last hidden state. The RNN layer can take in a sequence of inputs and a hidden state (default to all zeros for the first forward).

3

Can you explain, how should I train and use the net you suggested?

4

It’s hard to say without knowing the objective, but I would maybe do something like this:

rnn = torch.nn.RNN(16, 16)
x1 = torch.randn((1, 1, 16))    # Example input for one-to-many
seq_output_len = 5              # I want a sequence of five produced

yn = x1
hn = None
outputs = []
for _ in range(seq_output_len):
    yn, hn = rnn(yn, hn)        # Use previous output and hidden state
    outputs.append(yn)

output = torch.cat(outputs)     # 5 x 1 x 16
loss = do_something(output)
loss.backward()

This is a simplistic example which will vary on your use case

5

No, you did not understand, for training I already have all y_i and want to use them for training

6

For training, feed your entire sequence into your RNN and you will get a similar size sequence out. Compute a loss between that output and your target Y.

For testing, refer to the code I provided. output will be what you compare against the ground truth.