[REQUEST] Save model less frequently than metrics

[pstansell]

[REQUEST] Save model less frequently than metrics

[takuseno]

@pstansell Hi, sorry for being late. Now eval_interval argument is available.
0d91df6

[pstansell]

@takuseno, thanks for adding eval_interval, but I can only see it available in fit_batch_online(), not in fit_online(), which is where I wanted to use it.

[takuseno]

Ah, sorry. I've added save_interval to both methods.

[pstansell]

After further experimentation, I think v0.51 already had the argument I wanted to control the rate at which model files were written, that is,

  save_interval – interval to save parameters.

I hadn't realised what it did. Sometimes I struggle to understand exactly what the fuction arguments do based on their names and descriptions. But, anyway, the problem is solved now so I will close the issue. You make changes so quickly, it's hard for me to keep up :-). Thanks very much!

[pstansell]

@takuseno, it seems that fit_online() does not allow argument eval_interval.

[takuseno]

No, fit_online does not have that option. But you can control save frequency with save_interval.

[pstansell]

But doesn't that mean that I have to save models at the same rate as metrics? Put another way, the number of lines in metric files, for example time_step.csv, will be the same as the number of output files named model_*.pt. My initial desire was to be able to save model files at a lower rate than metrics. Given the current input arguments, I'm not sure if I can do that.

[takuseno]

The saved model file name and the epoch number in metrics will match. So, if you give save_interval=2, the files and metrics will be like below.

metrics

0,xxx,yyy
1,xxx,yyy
2,xxx,yyy
3,xxx,yyy
4,xxx,yyy
.
.

model files

model_0.pt
model_2.pt
model_4.pt
.
.

[takuseno]

Oh, wait. I found my bad. I'll fix this.

d3rlpy/d3rlpy/online/iterators.py

Line 258 in 0d91df6

logger.commit(epoch, total_step)

[takuseno]

@pstansell I've fixed this in the last commit.
b6c9b69

[pstansell]

I tested it and confirm that it's fixed now. Thanks very much.