hi, i do experiment basts and have some question about valid loss.
in experiment, valid_loss decrease for 100 epochs. and i saw below code.
loss = run_epoch_([val_batches[0]], model,
MultiGPULossCompute(model.generator, criterion, devices=devices, opt=model_opt))
i think opt=model_opt make model train at valid dataset either.
so when i experiments with opt=None , valid loss increased in some epoch ( between 10~20 epoch ).
my question is that..
What method does the performance of paper follow?
hi, i do experiment basts and have some question about valid loss.
in experiment, valid_loss decrease for 100 epochs. and i saw below code.
loss = run_epoch_([val_batches[0]], model,
MultiGPULossCompute(model.generator, criterion, devices=devices, opt=model_opt))
i think opt=model_opt make model train at valid dataset either.
so when i experiments with opt=None , valid loss increased in some epoch ( between 10~20 epoch ).
my question is that..
What method does the performance of paper follow?