bug in log level

[amcadmus]

I want to discuss the determination of the log_level by argparse, the code is in the following https://github.com/deepmodeling/deepmd-kit/blob/18ba5b19d8d2d237e600c6478ba8c8ba59e48d7b/source/train/main.py#L42-L50

Surprisingly, the action of "count" is to ADD the number of flags to the default value of log_level. Then if one sets -vvv, he/she will have a log_level of 5 rather than 3. To reproduce this behavoior,

import argparse
parser = argparse.ArgumentParser()
parser.add_argument("-v", default=2, action="count", dest="level")
args = parser.parse_args(['-vvv'])
print(args.level)

[amcadmus]

@marian-code Please take a look.

The unittest for parser_log does not find the issue which means that the unittest should be improved.

[marian-code]

OK, that behaviour is a really unexpected and unintuitive, I will take a look at it. EDIT: It makes sence when one looks at the code but this is not really mentioned in the docs. I will have to make some other system of passing in log level although I really liked this one as it is quite common in linux utilities and thus should be quite familiar to lot of people.

[marian-code]

I have realized the current solution is a complete mess. For instance there is no way of setting log level to 0 because if you do not pass any -v arguments it just defaults to 2.

I see here few possible reasonalbe solutions.

1. keep the curent system but set default=0 and nothing has to be changed. But downside is that by default user would not see any output unless he passes at least one -v argument. You would need to pass -vv to see logs that are now considered "default" - the INFO level. This is unfortuantelly not really intuitive.
2. The other option is to use someting of this style:

   parser.add_argument("-l", "--log-level", choices=['DEBUG', 'INFO', 'WARNING', 'ERROR'], default="INFO")

   Which is quite clear and consise but lacks the familiarity of -v style setting as is custom for linux tools.

3. The third solution is a bit hack-y but preserves the current behaviour with one minor change (the -v count is shifted so you need one more v for the same level as before):

   parser.add_argument("-v", default=1, action="count", dest="level", help="1(-v)=ERROR, 2(-vv)=WARNING, 3(-vvv)=INFO " 
                       "and 4(-vvvv)=DEBUG.  Default is INFO")
   args = parser.parse_args()
   # levels: 2=ERROR, 3=WARNING, 4=INFO, 5=DEBUG
   # adjust if `-v` was not passed in
   if args.level == 1:
       args.level = 4
   
   args.level -= 2  # this gives results excatly as they are expected in current state i.e. levels from 0 to 3

Please let me now what you think is the best solution.

[amcadmus]

1 is very good but we may want a default behavior of log_level == 2 I do not like the logic used by solution 3 Thus I prefer the second way, it may be improved as

parser.add_argument("-l", "--log-level", choices=['DEBUG', 3, 'INFO', 2, 'WARNING', 1, 'ERROR', 0], default="INFO")

so one can quickly set the log level to for example ERROR by

dp train -l 0 input.json