Feature/sg 573 Integrate new EMA decay schedules

[BloodAxe]

This PR brings option to use additional EMA decay schedules:

• constant - Fixed EMA decay value (Former config: exp_activation: False)
• exp - EMA decay warmup using exp rule decay * (1 - math.exp(-x * self.beta)) (Former config: exp_activation: True)
• threshold - New schedule used in PP-Yolo training min(decay, (1 + step) / (10 + step))

Examples:

ema: True
ema_params:
  decay_type: constant
  decay: 0.9999

ema: True
ema_params:
  decay: 0.9999
  decay_type: exp
  beta: 15

ema: True
ema_params:
  decay_type: threshold
  decay: 0.9999

Note: We still keep backward compatibility with old params and emit a deprecation warning showing what should be updated:

    def test_train_with_old_ema_params(self):
        self._init_model()
        self._train({"decay": 0.9999, "exp_activation": True, "beta": 10})

Emits the following warning message:

[2023-01-24 10:33:37] WARNING - sg_trainer.py - Parameter `exp_activation` is deprecated for EMA model. Please update your config to use decay_type: str (constant|exp|threshold) instead:
ema: True
ema_params: 
  decay: 0.9999
  decay_type: exp
  beta: 15

ema: True
ema_params: 
  decay: 0.9999
  decay_type: constant

ema: True
ema_params: 
  decay: 0.9999
  decay_type: threshold
In the next major release of SG this warning will become an error.

[dagshub]

Join the discussion on DagsHub!

[ofrimasad]

looks great.
I think we should use deprecate library for this deprecation. see my comments inline

[ofrimasad]

LGTM