Python version: >3.6
External Packages: pandas, matplotlib, sklearn, numpy
Last update: 31st Jan 2018

About this project: This is an academic project, submitted as a part of semester credits in "Machine Learning by Prof. Micheli " at Università di Pisa.

Demonstration of Perceptron Model

In this notebook, the perceptron model is demonstrated and tested. The model is functional perceptron model is developed from the scratch. The aim was to implement our understanidng of perceptron model and to test the effect of various hyper-parameters. It is a very flexible model with support of classification as well as regression tasks. This model supports hyper parameters such as learning rate, momentum, regularization, weight initialization types etc. Almost every parameter can be passed through grid search and can be tuned.

Grid search with cross validation is also developed in addition, to try different combinations of hyperparameters in order to do classification/regression tasks. Grid search returns the log(if log=True) and the list of best hyperparameters (length depends upon 'topn' function parameter). The hyper parameters are chosen according to either least trainingLoss or least validationLoss or the mean of two.

Experiments on Monk 1 dataset
Experiments on Monk 2 dataset
Experiments on Monk 3 dataset
Regression Problem | ML Cup

import pdb, time
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
from sklearn.model_selection import train_test_split
%matplotlib inline
import sklearn.preprocessing as preprocessing
from itertools import product

from source.perceptron import perceptron
from source.abGridSearchCV import abGridSearchCV
from source.loadData import loadMonk

0. Monk Datasets (to top)

Attributes with granularity: a1: 1, 2, 3 a2: 1, 2, 3 a3: 1, 2 a4: 1, 2, 3 a5: 1, 2, 3, 4 a6: 1, 2 target: 0, 1

In Monk 1 Dataset, the targets are 1 if (a1==a2) or (a5==1)

In Monk 2 Dataset, the targets are 1 if exactly two of {a1 = 1, a2 = 1, a3 = 1, a4 = 1, a5 = 1, a6 = 1} are correct.

In Monk 3 Dataset, targets are 1 if (a5 = 3 and a4 = 1) or (a5 /= 4 and a2 /= 3).

Carnegie Mellon University Pittsburgh, PA 15213, USA E-mail: thrun '@' cs.cmu. edu.

1. Experiments on Monk 1 (to top)

trainData, trainLabels=loadMonk(1, 'train', encodeLabel=False)

A. Grid Search on Monk1 (to top)

ETA is learning rate, LAMBDA is regularization parameter, ALPHA is momentum,

defaultParameters={
    'hiddenUnits':3,
    'randomSeed':0,
    'activation':'sigm',
    'epochs':500,
    'ETA':0.4,
    'LAMBDA':0.0,
    'loss':"MEE",
    'ALPHA':0.9,
    'weightInitialization':'xav',
    'regression':False,
    'earlyStopping':True,
    'tolerance':1e-3,
    'patience':20
}

parameterGridForModelSelection={
#     'randomSeed':[0, 20],
    'hiddenUnits':[4, 16],
    'activation':['sigm', 'relu', 'tanh'],
    'ETA':[0.1, 0.2, 0.3, 0.5, 0.7,0.9],
    'LAMBDA':[0.001,0.01],
    'ALPHA':[0.2, 0.5, 0.7, 0.9],
#     'weightInitialization':['xav', 'he', 'type1'],
#     'epochs':[300, 500]
}

top5BestParams=abGridSearchCV(defaultParameters, parameterGridForModelSelection, trainData, trainLabels, winnerCriteria="meanLosses", validationSplit=0.3, log=False, topn=5)

iteration 21/500        287/288

top5BestParams

[{'meanLosses': 0.080383052475397454,
  'meanTrainingLoss': 0.074816536086853846,
  'meanValidationLoss': 0.085949568863941075,
  'params': {'ALPHA': 0.5,
   'ETA': 0.3,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 413,
   'hiddenUnits': 4}},
 {'meanLosses': 0.081376470456553995,
  'meanTrainingLoss': 0.07491194201533527,
  'meanValidationLoss': 0.087840998897772707,
  'params': {'ALPHA': 0.9,
   'ETA': 0.2,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 477,
   'hiddenUnits': 4}},
 {'meanLosses': 0.082419646642824682,
  'meanTrainingLoss': 0.07471128658272097,
  'meanValidationLoss': 0.090128006702928393,
  'params': {'ALPHA': 0.7,
   'ETA': 0.5,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 352,
   'hiddenUnits': 4}},
 {'meanLosses': 0.084242516160344783,
  'meanTrainingLoss': 0.072312472974062944,
  'meanValidationLoss': 0.096172559346626621,
  'params': {'ALPHA': 0.9,
   'ETA': 0.7,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 299,
   'hiddenUnits': 4}},
 {'meanLosses': 0.084373773845286978,
  'meanTrainingLoss': 0.079003026519703604,
  'meanValidationLoss': 0.089744521170870353,
  'params': {'ALPHA': 0.7,
   'ETA': 0.2,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 476,
   'hiddenUnits': 4}}]

bestParams=top5BestParams[0]['params']
defaultParameters['earlyStopping']=False

bestParams

{'ALPHA': 0.5,
 'ETA': 0.3,
 'LAMBDA': 0.001,
 'activation': 'sigm',
 'epochs': 413,
 'hiddenUnits': 4}

B. Training Monk 1 with best parameters obtained (to top)

For fun.. testing a neural network initialized with zero weights

x=perceptron(**defaultParameters)
x.set_params(**bestParams)
x.set_params(weightInitialization='zero')
x.fit(trainData, trainLabels)

iteration 413/413

plt.plot(x.losses)

[<matplotlib.lines.Line2D at 0x7f725519ffd0>]

plt.plot(x.accuracies)

[<matplotlib.lines.Line2D at 0x7f7255100080>]

testData, testLabels=loadMonk(1, 'test', encodeLabel=False)
testResults, testAccuracy=x.predict(testData, testLabels, acc_=True)

testAccuracy

0.5

fun's over!!!; now let's train our network with best params

x=perceptron(**defaultParameters)
x.set_params(**bestParams)
x.fit(trainData, trainLabels)

iteration 413/413

plt.plot(x.losses)

[<matplotlib.lines.Line2D at 0x7f72550d83c8>]

plt.plot(x.accuracies)

[<matplotlib.lines.Line2D at 0x7f725502dcf8>]

B. Testing Monk 1 trained model(to top)

testData, testLabels=loadMonk(1, 'test', encodeLabel=False)

testResults, testAccuracy=x.predict(testData, testLabels, acc_=True)

testAccuracy

1.0

C. Training Monk 1 with best parameters obtained, with cross validation (to top)

trainData, trainLabels=loadMonk(1, 'train', encodeLabel=False)
trainData, validationData, trainLabels, validationLabels=train_test_split(trainData, trainLabels, test_size=0.1)

x=perceptron(**defaultParameters)
x.set_params(**bestParams)
x.fit(trainData, trainLabels, validationFeatures=validationData, validationLabels=validationLabels)

iteration 413/413

plt.plot(x.accuracies)
plt.plot(x.validationAccuracies)

[<matplotlib.lines.Line2D at 0x7f7255078400>]

plt.plot(x.losses)
plt.plot(x.validationLosses)

[<matplotlib.lines.Line2D at 0x7f71bac6a748>]

2. Monk 2 (to top)

trainData, trainLabels=loadMonk(2, 'train', encodeLabel=False)

A. Grid Search on Monk 2 (to top)

defaultParameters['earlyStopping']=True
top5BestParams=abGridSearchCV(defaultParameters, parameterGridForModelSelection, trainData, trainLabels, winnerCriteria="meanLosses", validationSplit=0.3, log=False, topn=5)

iteration 21/500        287/288

top5BestParams

[{'meanLosses': 0.073474774082772629,
  'meanTrainingLoss': 0.071237613323866827,
  'meanValidationLoss': 0.07571193484167843,
  'params': {'ALPHA': 0.9,
   'ETA': 0.3,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 372,
   'hiddenUnits': 4}},
 {'meanLosses': 0.077636512762921425,
  'meanTrainingLoss': 0.074975553679050108,
  'meanValidationLoss': 0.080297471846792742,
  'params': {'ALPHA': 0.7,
   'ETA': 0.3,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 376,
   'hiddenUnits': 4}},
 {'meanLosses': 0.085343996795136656,
  'meanTrainingLoss': 0.082409373163706043,
  'meanValidationLoss': 0.08827862042656727,
  'params': {'ALPHA': 0.5,
   'ETA': 0.3,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 384,
   'hiddenUnits': 4}},
 {'meanLosses': 0.087062271878194264,
  'meanTrainingLoss': 0.083728924138565419,
  'meanValidationLoss': 0.090395619617823095,
  'params': {'ALPHA': 0.9,
   'ETA': 0.5,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 315,
   'hiddenUnits': 4}},
 {'meanLosses': 0.089620261402815996,
  'meanTrainingLoss': 0.085772966902846179,
  'meanValidationLoss': 0.093467555902785812,
  'params': {'ALPHA': 0.7,
   'ETA': 0.5,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 321,
   'hiddenUnits': 4}}]

defaultParameters['earlyStopping']=False
bestParams=top5BestParams[0]['params']

B. Training Monk 2 with the best parameters obtained(to top)

x=perceptron(**defaultParameters)
x.set_params(**bestParams)
x.fit(trainData, trainLabels)

iteration 372/372

plt.plot(x.accuracies)

[<matplotlib.lines.Line2D at 0x7fe23a8dfcc0>]

plt.plot(x.losses)

[<matplotlib.lines.Line2D at 0x7fe2379cf0b8>]

C. Testing Monk 2(to top)

# to see if a model trained on Monk2 can classify Monk 1
testData, testLabels=loadMonk(1, 'test', encodeLabel=False)
testResults, testAccuracy=x.predict(testData, testLabels, acc_=True)

testAccuracy #of Monk1 on model trained on monk2

0.43981481481481483

#getting serious now ;D
testData, testLabels=loadMonk(2, 'test', encodeLabel=False)
testResults, testAccuracy=x.predict(testData, testLabels, acc_=True)

testAccuracy #yayy!

1.0

D. Training Monk 2 with the best parameters and with cross validation(to top)

trainData, trainLabels=loadMonk(2, 'train', encodeLabel=False)
trainData, validationData, trainLabels, validationLabels=train_test_split(trainData, trainLabels, test_size=0.1)
x=perceptron(**defaultParameters)
x.set_params(**bestParams)
x.fit(trainData, trainLabels, validationFeatures=validationData, validationLabels=validationLabels)

iteration 372/372

plt.plot(x.accuracies)
plt.plot(x.validationAccuracies)

[<matplotlib.lines.Line2D at 0x7fe23a8c4c50>]

plt.plot(x.losses)
plt.plot(x.validationLosses)

[<matplotlib.lines.Line2D at 0x7fe2379679e8>]

3. Monk 3 (to top)

trainData, trainLabels=loadMonk(3, 'train', encodeLabel=False)

A. Grid Search on Monk 3 (to top)

defaultParameters['earlyStopping']=True
top5BestParams=abGridSearchCV(defaultParameters, parameterGridForModelSelection, trainData, trainLabels, winnerCriteria="meanLosses", validationSplit=0.3, log=False, topn=5)
defaultParameters['earlyStopping']=False

iteration 21/500        287/288

top5BestParams

[{'meanLosses': 0.073088388910850718,
  'meanTrainingLoss': 0.05514765608540876,
  'meanValidationLoss': 0.09102912173629267,
  'params': {'ALPHA': 0.9,
   'ETA': 0.7,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 363,
   'hiddenUnits': 4}},
 {'meanLosses': 0.074405958835380953,
  'meanTrainingLoss': 0.06467883672931346,
  'meanValidationLoss': 0.084133080941448432,
  'params': {'ALPHA': 0.9,
   'ETA': 0.3,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 382,
   'hiddenUnits': 4}},
 {'meanLosses': 0.076441115686924674,
  'meanTrainingLoss': 0.065130088221405716,
  'meanValidationLoss': 0.087752143152443632,
  'params': {'ALPHA': 0.7,
   'ETA': 0.3,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 396,
   'hiddenUnits': 4}},
 {'meanLosses': 0.082981430081672025,
  'meanTrainingLoss': 0.069712231732486349,
  'meanValidationLoss': 0.096250628430857715,
  'params': {'ALPHA': 0.5,
   'ETA': 0.3,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 375,
   'hiddenUnits': 4}},
 {'meanLosses': 0.083468995709044772,
  'meanTrainingLoss': 0.073056830286287025,
  'meanValidationLoss': 0.093881161131802518,
  'params': {'ALPHA': 0.2,
   'ETA': 0.9,
   'LAMBDA': 0.001,
   'activation': 'sigm',
   'epochs': 363,
   'hiddenUnits': 4}}]

bestParams=top5BestParams[0]['params']

B. Training Monk 3 with the best parameters obtained(to top)

x=perceptron(**defaultParameters)
x.set_params(**bestParams)
x.fit(trainData, trainLabels)

iteration 363/363

plt.plot(x.accuracies)

[<matplotlib.lines.Line2D at 0x7fe23789e080>]

plt.plot(x.losses)

[<matplotlib.lines.Line2D at 0x7fe233163668>]

C. Testing Monk 3 trained Model(to top)

testData, testLabels=loadMonk(3, 'test', encodeLabel=False)
testResults, testAccuracy=x.predict(testData, testLabels, acc_=True)

testAccuracy

0.94444444444444442

D. Training monk 3 with the best parameters and with cross validation(to top)

trainData, trainLabels=loadMonk(3, 'train', encodeLabel=False)
trainData, validationData, trainLabels, validationLabels=train_test_split(trainData, trainLabels, test_size=0.2)
x=perceptron(**defaultParameters)
x.set_params(**bestParams)
x.fit(trainData, trainLabels, validationFeatures=validationData, validationLabels=validationLabels)

iteration 363/363

plt.plot(x.accuracies)
plt.plot(x.validationAccuracies)

[<matplotlib.lines.Line2D at 0x7fe23783ef98>]

plt.plot(x.losses)
plt.plot(x.validationLosses)

[<matplotlib.lines.Line2D at 0x7fe23315a518>]

4. Regression Problem | ML Cup(to top)

The dataset (ML Cup) contains 10 features and 2 labels. Both features and lables are continuous values. The dataset contains 1015 rows for training. The blind test data contains 315 unlabled rows with 10 columns.

Dataset provided by Prof. Alessio Micheli, Università di Pisa.

Data preparation

cupTrain=pd.read_csv('data/mlCup/ML-CUP18-TR.csv', skiprows=9, usecols=range(1,13))
colums=["i{}".format(i) for i in range(1,11)]+['classX', 'classY']
cupTrain.columns=colums
len(cupTrain)

features=cupTrain.drop(['classX','classY'],axis = 1).values
labels=cupTrain[['classX', 'classY']].values

cupTest=pd.read_csv('data/mlCup/ML-CUP18-TS.csv', header=None)
colums=['index']+["i{}".format(i) for i in range(1,11)]
cupTest.columns=colums
len(cupTest)

cupTest=cupTest.drop(['index'], axis=1)
blindFeatures=cupTest.values

trainDataCup, validationDataCup, trainLabelsCup, validationLabelsCup=train_test_split(features,labels, test_size=0.2, shuffle=True)

minMaxScaler = preprocessing.MinMaxScaler()
stdScaler=preprocessing.StandardScaler()

scaling

trainData=stdScaler.fit_transform(features)
trainLabels=stdScaler.fit_transform(labels)
trainDataCup =stdScaler.fit_transform(trainDataCup)
validationDataCup=stdScaler.fit_transform(validationDataCup)
blindFeatures=stdScaler.fit_transform(blindFeatures)

A. Grid Search on regression problem (to top)

defaultParameters={
    'hiddenUnits':3,
    'randomSeed':0,
    'activation':'sigm',
    'epochs':1000,
    'ETA':0.4,
    'LAMBDA':0.0,
    'loss':"MSE",
    'ALPHA':0.9,
    'weightInitialization':'xav',
    'regression':True,
    'earlyStopping':True,
    'tolerance':1e-3,
    'patience':20
    
}

parameterGridForModelSelection={
#     'randomSeed':[0, 20],
    'hiddenUnits':[10, 50, 80],
    'activation':['sigm', 'relu', 'tanh'],
    'ETA':[0.01, 0.2, 0.4, 0.9],
    'LAMBDA':[0.1,0.01, 0.001],
    'ALPHA':[0.4, 0.1, 0.7, 0.01],
    'weightInitialization':['xav', 'he'],
    'epochs':[800] #before that, it'll stop automatically due to earlystopping so need to add more options
}

top5BestParams=abGridSearchCV(defaultParameters, parameterGridForModelSelection, trainData, trainLabels, winnerCriteria="meanLosses", validationSplit=0.3, log=False, topn=5)

iteration 800/800       863/864

defaultParameters['earlyStopping']=False #we don't need it to stop early during only training as we've already chosen our best params
top5BestParams

[{'meanLosses': 0.072376807920726011,
  'meanTrainingLoss': 0.072214817972934775,
  'meanValidationLoss': 0.07253879786851726,
  'params': {'ALPHA': 0.7,
   'ETA': 0.2,
   'LAMBDA': 0.001,
   'activation': 'relu',
   'epochs': 800,
   'hiddenUnits': 80,
   'weightInitialization': 'xav'}},
 {'meanLosses': 0.073865404590804068,
  'meanTrainingLoss': 0.074877710251905982,
  'meanValidationLoss': 0.072853098929702154,
  'params': {'ALPHA': 0.7,
   'ETA': 0.2,
   'LAMBDA': 0.01,
   'activation': 'relu',
   'epochs': 800,
   'hiddenUnits': 80,
   'weightInitialization': 'xav'}},
 {'meanLosses': 0.074071496601288839,
  'meanTrainingLoss': 0.075142642736041237,
  'meanValidationLoss': 0.073000350466536454,
  'params': {'ALPHA': 0.7,
   'ETA': 0.2,
   'LAMBDA': 0.001,
   'activation': 'relu',
   'epochs': 800,
   'hiddenUnits': 50,
   'weightInitialization': 'xav'}},
 {'meanLosses': 0.074848918692942454,
  'meanTrainingLoss': 0.076864281759301281,
  'meanValidationLoss': 0.072833555626583626,
  'params': {'ALPHA': 0.7,
   'ETA': 0.2,
   'LAMBDA': 0.01,
   'activation': 'relu',
   'epochs': 800,
   'hiddenUnits': 50,
   'weightInitialization': 'xav'}},
 {'meanLosses': 0.075499881161355337,
  'meanTrainingLoss': 0.075738744094420871,
  'meanValidationLoss': 0.075261018228289789,
  'params': {'ALPHA': 0.7,
   'ETA': 0.2,
   'LAMBDA': 0.001,
   'activation': 'relu',
   'epochs': 800,
   'hiddenUnits': 80,
   'weightInitialization': 'he'}}]

bestParams=top5BestParams[0]['params']
bestParams

{'ALPHA': 0.7,
 'ETA': 0.2,
 'LAMBDA': 0.001,
 'activation': 'relu',
 'epochs': 800,
 'hiddenUnits': 80,
 'weightInitialization': 'xav'}

B. Training with best parameters obtained (to top)

x=perceptron(**defaultParameters)
x.set_params(**bestParams)
x.fit(trainData, trainLabels, realTimePlotting=False)

iteration 800/800

plt.plot(x.losses)

[<matplotlib.lines.Line2D at 0x7fb4afab0588>]

def cutIt(array):
    x=[]
    y=[]
    for i in array:
        x.append(i[0])
        y.append(i[1])
    return x,y

predictions=x.predict(trainData)

xx, yy=cutIt(predictions)
xxx,yyy=cutIt(trainLabels)

plt.scatter(xx,yy, alpha=0.9, s=10)
plt.scatter(xxx,yyy, alpha= 0.5, s = 2, marker='^')

<matplotlib.collections.PathCollection at 0x7fb4afaa5080>

C. Training with best parameters obtained, with validation (to top)

x=perceptron(**defaultParameters)
x.set_params(**bestParams)
x.fit(trainDataCup, trainLabelsCup, validationDataCup, validationLabelsCup, realTimePlotting=False)

iteration 800/800

plt.plot(x.losses)
plt.plot(x.validationLosses)

[<matplotlib.lines.Line2D at 0x7fb4afb06748>]

D. Blind Test (to top)

predicting on model trained with cross validation

blindPredictions=x.predict(blindFeatures)

xx, yy=cutIt(blindPredictions)
xxx,yyy=cutIt(trainLabelsCup)
plt.scatter(xx,yy,alpha= 0.9, s = 10, marker='^')
plt.scatter(xxx,yyy, alpha= 0.5, s = 2)

<matplotlib.collections.PathCollection at 0x7fb4ab2b8c50>

regressionModel=perceptron(**defaultParameters)
regressionModel.set_params(**bestParams)
regressionModel.fit(trainDataCup, trainLabelsCup, realTimePlotting=False)
blindPredictions=regressionModel.predict(blindFeatures)

iteration 800/800

plt.plot(regressionModel.losses)

[<matplotlib.lines.Line2D at 0x7fb4ab21f5f8>]

blindPredictions=regressionModel.predict(blindFeatures)

xx, yy=cutIt(blindPredictions)
xxx,yyy=cutIt(trainLabelsCup)
plt.scatter(xx,yy,alpha= 0.9, s = 10, marker='^')
plt.scatter(xxx,yyy, alpha= 0.5, s = 2)

<matplotlib.collections.PathCollection at 0x7fb4123ee908>

iabd / perceptron-model Goto Github PK

perceptron-model's Introduction

Demonstration of Perceptron Model

Table of Contents