CODEWITHSUNDEEP
tensorflowtensorboard
Nicolas Bruno
Nicolas Bruno

Reputation: 69

Download smoothed tensorboard values

Is it possible to downloaded the smooth values generated from Tensorboard or at least get the smoothing function to be able to generate the same graphics as in Tensorboard.

Upvotes: 1

Views: 1367

Answers (3)

Charlie Parker
Charlie Parker

Reputation: 5201

they use some type of exp weighted avg. Try:

def my_tb_smooth(scalars: list[float], weight: float) -> list[float]:  # Weight between 0 and 1
    """

    ref: https://stackoverflow.com/questions/42011419/is-it-possible-to-call-tensorboard-smooth-function-manually

    :param scalars:
    :param weight:
    :return:
    """
    last = scalars[0]  # First value in the plot (first timestep)
    smoothed: list = []
    for point in scalars:
        smoothed_val = last * weight + (1 - weight) * point  # Calculate smoothed value
        smoothed.append(smoothed_val)                        # Save it
        last = smoothed_val                                  # Anchor the last smoothed value
    return smoothed

credit: Is it possible to call tensorboard smooth function manually?

Upvotes: 0

Yaoshiang
Yaoshiang

Reputation: 1941

rate = 0.1 # range of 0.0 for no smoothing, 1.0 for overly perfect smoothing.

def ema(old:float, new:float, rate:float)->float: return old * rate + new * (1.0 - rate)

ema(10, 9, 0.1)

Upvotes: 0

Allen Lavoie
Allen Lavoie

Reputation: 5808

It has changed a bit recently, but currently TensorBoard does exponential averaging for its smoothing. Should be quite easy to re-implement.

Upvotes: 1

Related Questions