Cleaner RunningAverage syntax

The current way to add multiple running averages computed over the output is a bit too verbose imo, and it can be made simpler.

DCGAN Example:

    def step(engine, batch):

       (...)

        return {
            'errD': errD.item(),
            'errG': errG.item(),
            'D_x': D_x,
            'D_G_z1': D_G_z1,
            'D_G_z2': D_G_z2
        }

Current API

To compute running averages over all of these, I would have to create an object per output and call the attach method:

    avg_err_d = RunningAverage(output_transform=lambda x: x['errD'])
    avg_err_d.attach(trainer, 'errD')
    avg_err_g = RunningAverage(output_transform=lambda x: x['errG'])
    avg_err_g.attach(trainer, 'errG')
    avg_dx = RunningAverage(output_transform=lambda x: x['D_x'])
    avg_dx.attach(trainer, 'D_x')
    avg_dg_z1 = RunningAverage(output_transform=lambda x: x['D_G_z1'])
    avg_dg_z1.attach(trainer, 'D_G_z1')
    avg_dg_z2 = RunningAverage(output_transform=lambda x: x['D_G_z2'])
    avg_dg_z2.attach(trainer, 'D_G_z2')

Proposed API

Initialize one RunningAverage object and pass the output_transform function to the attach method

    running_avgs = RunningAverage()
    running_avgs.attach(trainer, 'errD', output_transform=lambda x: x['errD'])
    running_avgs.attach(trainer, 'errG', output_transform=lambda x: x['errG'])
    running_avgs.attach(trainer, 'D_x', output_transform=lambda x: x['D_x'])
    running_avgs.attach(trainer, 'D_G_z1', output_transform=lambda x: x['D_G_z1'])
    running_avgs.attach(trainer, 'D_G_z2', output_transform=lambda x: x['D_G_z2'])

cc @vfdev-5

[miguelvr]

i’m going to close this, as the condensed syntax proposed by @jasonkriss seems okay to me

[jasonkriss]

You’re right. There is no compatibility. Not sure what i was thinking 😃.

Yea i mean in general, but since the create_supervised_* functions are the only explicit place where we define process functions, that’s where it would show up.