I’m trying to develop an inference script for a Non autoregressive Transformer, I’ve used nn.transformer() for my model, I’ve got alright results during training, but I had no luck developing an inference script
Since the Decoder layer requires source and target, for my target I’ve used a tensor of size [batch, target_len] filled with start token/start token+pad token but the accuracy is close to 0
Could anyone provide help on how I would develop the inference script or what value I should pass as the target value?