带网格搜索的火车

Question

I would like to test trains usage during grid search and it not clear how to do so.

from trains import Task 
Task.init(project_name="project name", task_name='name')

creates an experiment in the demo server and logs all but you can't call init twice no matter the 'task_name' and

from trains import Task 
Task.create(project_name="project name", task_name='name')

can be called with different 'task_name' but thus not log any data into the server and creates only 'Draft'.

here is a sample code:

 epochs=[160,300]
 for epoch in epochs:
    model = define_model_run(epoch)
    model.fit(x_train,y_train)
    score = model.score(...)

my final try was:

 epochs=[160,300]
 task=Task.init(project_name="demo", task_name='search')
 for epoch in epochs:
    task.create(project_name="demo", task_name=f'search_{epoch}')
    model = define_model_run(epoch)
    model.fit(x_train,y_train)
    score = model.score(...)

which logs all information under the experiments tab and none under the 'Draft'. I tried the last two hour the read the few documentations provided and reading the source code, but no luck.

any help?

Answer 1

Solved by looking on closed bugs in their GitHub repository: link

 epochs=[160,300]
 for epoch in epochs:
    task=Task.init(project_name="demo", task_name=f'search_{epoch}')
    model = define_model_run(epoch)
    model.fit(x_train,y_train)
    score = model.score(...)
    task.close()

then each task is a different experiment in project "demo".

Answer 2

Declaimer: I'm a member of TRAINS team

Yes, that's exactly the answer. The idea is that you always have one main Task, in order to create a new one you need to close the running Task, and re-initialize with a new name. Kudos on solving it so quickly :)

BTW: You can see examples here / and here , showing how to send accuracy logs so it is easier to compare the experiments, especially when running hyper-parameter search.