A simple pytorch based MNL lib

Fit your Multinomial Logistic Regression with Pytorch

Install

pip install pytorch_mnl

How to use

import the lib

import pandas as pd
from pytorch_mnl.core import *

load data

data = pd.read_csv("./data/Iris.csv").drop("Id", axis=1)

choose x, y cols:

x_cols=['SepalLengthCm', 'SepalWidthCm', 'PetalLengthCm', 'PetalWidthCm']
target_col = 'Species'

the number of classes to predict:

n_targets = len(data[target_col].unique())
n_targets

3

X, y = prepare_data(data, x_cols=x_cols, target_col=target_col)

we get pytorch tensors ready to use!

type(X), type(y)

(torch.Tensor, torch.Tensor)

let's split in train/valid choosing a percenage as holdout, and choose a batch size to fit our model

dls = DataLoaders.from_Xy(X, y, pct=0.2, batch_size=8)

as our model has 4 variables, we will fit a 4 MNL, with 3 targets.

model = LinearMNL(len(x_cols), n_targets)

learn = Learner(dls, model)

learn.fit(25)

epoch =   0, val_loss = 2.072, accuracy = 0.53
epoch =   1, val_loss = 1.908, accuracy = 0.53
epoch =   2, val_loss = 1.770, accuracy = 0.80
epoch =   3, val_loss = 1.657, accuracy = 0.80
epoch =   4, val_loss = 1.564, accuracy = 0.80
epoch =   5, val_loss = 1.487, accuracy = 0.80
epoch =   6, val_loss = 1.422, accuracy = 0.80
epoch =   7, val_loss = 1.368, accuracy = 0.80
epoch =   8, val_loss = 1.321, accuracy = 0.80
epoch =   9, val_loss = 1.282, accuracy = 0.83
epoch =  10, val_loss = 1.247, accuracy = 0.83
epoch =  11, val_loss = 1.217, accuracy = 0.83
epoch =  12, val_loss = 1.190, accuracy = 0.83
epoch =  13, val_loss = 1.166, accuracy = 0.83
epoch =  14, val_loss = 1.144, accuracy = 0.87
epoch =  15, val_loss = 1.125, accuracy = 0.87
epoch =  16, val_loss = 1.107, accuracy = 0.90
epoch =  17, val_loss = 1.091, accuracy = 0.90
epoch =  18, val_loss = 1.076, accuracy = 0.90
epoch =  19, val_loss = 1.063, accuracy = 0.90
epoch =  20, val_loss = 1.050, accuracy = 0.90
epoch =  21, val_loss = 1.038, accuracy = 0.90
epoch =  22, val_loss = 1.027, accuracy = 0.90
epoch =  23, val_loss = 1.016, accuracy = 0.90
epoch =  24, val_loss = 1.007, accuracy = 0.90