What do the Metrics and Losses mean?

[kuboko-jp]

Hi. Thank you for publishing this wonderful program. I have a question about Metrics and loss. I have checked the training of the 1b_lyric model in train.py through Tensorboard. In Tensorboard, I can monitor things like loss, but I couldn't figure out what each value indicates.

For example, in prior/autoregressive.py, you will find the code shown below, and I understand that it is a cross entropy loss, but I did not understand what exactly these values indicate. prime_loss = F.cross_entropy(x_prime, x_t[:, :self.prime_len].reshape(-1)) / np.log(2.) gen_loss = F.cross_entropy(x_gen, x_t[:, self.prime_len:].reshape(-1)) / np.log(2.)

Could you please tell me the meaning of the values shown below as they are in TensorBoard_training.pdf ?

• bpd
• gen_loss
• gn
• lg_loss_scale
• loss