What is the difference between src and tgt in nn.Transformer for PyTorch?

Question

From the docs it says to create a transformer model like this:

transformer_model = nn.Transformer(nhead=16, num_encoder_layers=12)
src = torch.rand((10, 32, 512))
tgt = torch.rand((20, 32, 512)) # What is tgt??
out = transformer_model(src, tgt) 

What is tgt mean't to be? Should tgt be the same as the src?

Answer 1

The transformer structure is of two components, the encoder and the decoder. The src is the input to encoder and the tgt is the input to decoder.

For example doing a machine translation task that translates English sentence to French, the src is english sequence ids and tgt is french sequence ids.