Sklearn MLPRegressor to FluidMLPRegressor (json)

tedmoore · February 12, 2021, 1:10pm

Ciao,

I was doing some training in sklearn and realized that it may be possible to just package the parameters of the sklearn MLPRegressor to write a Fluid-style json for the MLPRegressor. Here’s the code. It seems to be working well. Maybe it’s helpful for others.

Cheers,
T

github.com

tedmoore/FluCoMa-stuff/blob/master/sklearn_mlp_to_fluid_mlp.py

from sklearn.neural_network import MLPRegressor

def mlpToFluidJsonDict(mlp):
    '''
    "layers":{ // layers are in feedforward order
        "activation": // int from FluidMLPRegressor //,
        "biases": biases for the layer (array of "cols" length)
        "cols": int num of neurons in this layer ,
        "rows": num inputs to this layer,
        "weights: array of "rows" arrays with "cols" items in each,
    }
    '''
    weights = mlp.coefs_
    biases = mlp.intercepts_
    json_dict = {"layers":[{} for _ in range(mlp.n_layers_ - 1)]}
    activation = 0
    
    if mlp.activation == 'identity':
        activation = 0
    elif mlp.activation == 'logistic':

This file has been truncated. show original

a.harker · February 12, 2021, 3:47pm

What’s the advantage or reason (from your perspective) of doing your training in sklearn, if you don’t mind me asking?

tedmoore · February 12, 2021, 3:59pm

Partially it’s because I had some existing code there that loaded in my csv files and parsed them how I wanted, etc.

But also, there is some stuff that is easier to do in Python than in other places, mainly because of all of the sklearn packages, like some data manipulation utilities that aren’t implemented in Fluid stuff yet (randomly splitting a dataset into train and test sets and other numpy-like stuff). Also, the server client management in SC can be annoying when trying to train many MLPs for example–lots of callback functions, server syncing, buffer filling, and the like, where in Python it’s more straight ahead sequential code.

However, the MLP training is faster in FluCoMa than it is in sklearn, which makes sense. I don’t think sklearn is better just different. In the long run I’m planning to be primarily in SC, but the two do feel different, those are some thoughts.

weefuzzy · February 13, 2021, 10:30am

A particular thing might be that there are many more tools for doing validation of learnt models available in sklearn, and python more generally.

Also, if people start wanting to train some much more chunky models and just use FluCoMa stuff at runtime, it may be practical to lease processing time from google for training or whoever and use a python library for that.

tremblap · February 13, 2021, 11:49am

Espectially when someone says that ‘you can even train your own dataset in 2 to 4 days’ here:

jamesbradbury · February 13, 2021, 12:46pm

You would think in today’s current news climate you would not succumb to such Murdoch style editorial sniper shots of my words