search

Kyubyong / transformer

A TensorFlow Implementation of the Transformer: Attention Is All You Need
Apache License 2.0
4.25k stars 1.29k forks source link

train.py, small logic problems #143

Open 915288938lx opened 4 years ago

915288938lx commented 4 years ago

in the eval_dataset, the shuffle param is set to False, so every time we run(evel_init_op), we just get the first batch in the eval_dataset, so get_hypotheses just get a ramdom sample of the first batch of the eval_dataset

lidongxing commented 4 years ago

Also, the training loss should be last batch mean loss of every epoch, why is epoch loss?