在rllib的example中,有一个custom model的例子,链接,但是运行这个就会发现一个问题,没有训练的日志输出,只有这个状态信息,训练过程中的training iter, episode reward mean等信息都不输出。
== Status ==
Memory usage on this node: 1.2/9.3 GiB
Using FIFO scheduling algorithm.
Resources requested: 0/1 CPUs, 0/0 GPUs, 0.0/5.22 GiB heap, 0.0/2.61 GiB objects
Result logdir: /home/yan/ray_results/IMPALA
Number of trials: 1/1 (1 PENDING)
== Status ==
Memory usage on this node: 1.2/9.3 GiB
Using FIFO scheduling algorithm.
Resources requested: 0/1 CPUs, 0/0 GPUs, 0.0/5.22 GiB heap, 0.0/2.61 GiB objects
Result logdir: /home/yan/ray_results/IMPALA
Number of trials: 1/1 (1 PENDING)
解决途径:
调用tune.run()时候,将verbose等级设为2级或者3级。
tune.run(config=config, verbose=3)
也可以自定义logger的行为,详见custom logger
参考:
https://discuss.ray.io/t/no-logs-using-custom-model/2063