地址：https://parl.ai/docs/tutorial_fast.html
We’ll start with an example training command, which trains a transformer/generator on ConvAI2 for 1 epoch,
using a batchsize of 64, with a roughly 20M parameter model. We’ll train using ADAM optimizer, with a learning rate of 1e-3.
We’ll build the dictionary ahead of time to ensure it’s kept the same. This will be our baseline.
parlai build_dict --task convai2 --dict-file dictfile parlai train --dict-file dictfile --model transformer/generator --task --task convai2 --num-epochs 1.0 --batchsize 64 --n-layers 8 --embedding-size 250 --ffn-size 1000 --optimizer adam --learningrate 1e-3
原文：https://www.cnblogs.com/yanghh/p/14784681.html