qa.md

Q & A

1. How to train my customized network using the framework of APS?

   There are two options. The first one is straightforward: modify the source code of the APS following the steps:

   • Give your implementation of network structure if needed in aps.asr, aps.sse or aps.xxx (may coming in the future) and decorate it using @ApsRegisters.xxx.register(...).

   • Give your implementation of Task if needed in aps.task and also decorate it using @ApsRegisters.task.register(...).

   • Give your implementation of dataloader if needed in aps.loader and also decorate with @ApsRegisters.loader.register(...).

   • Prepare your training & validation & test data & configuration files and train the models using the scripts scripts/*.sh.

     If new python files are added, remember to update the ApsModules class in aps/libs.py to make sure your implementation can be imported correctly.

   Another way only requires us to modify the training configurations. Assuming we have my_nnet.py, my_task.py under /path/to/my_code, the .yaml configuration like

   nnet: /path/to/my_code/my_nnet.py:MyNnet
   nnet_conf:
       # put parameters here
       ...
   task: /path/to/my_code/my_task.py:MyTask
   task_conf:
       # put parameters here
       ...
   # other configurations
   ...

   could be used. In this case, please remember to make sure there is no import errors in your python code.