MNIST 朴素贝叶斯分类器

MNIST朴素贝叶斯分类器

import glob
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns; sns.set()
from sklearn.naive_bayes import GaussianNB
from sklearn.metrics import accuracy_score
from sklearn.metrics import confusion_matrix
from skimage.transform import pyramid_gaussian
from sklearn.model_selection import train_test_split

print(glob.glob("input\*"))

%matplotlib inline
%precision 2

['input\\test.csv', 'input\\train.csv']





'%.2f'

def draw_confusionmatrix(ytest, yhat):
    plt.figure(figsize=(10,7))
    cm = confusion_matrix(ytest, yhat)
    ax = sns.heatmap(cm, annot=True, fmt="d")
    plt.ylabel('True label')
    plt.xlabel('Predicted label')

    acc = accuracy_score(ytest, yhat)
    print(f"Sum Axis-1 as Classification accuracy: {acc}")

train_df = pd.read_csv('input\\train.csv')
test_df = pd.read_csv('input\\test.csv')

X = train_df.iloc[:,1:]
y = train_df.label
label = sorted(y.unique())

plt.figure(figsize=(9,9))
for i in label:
    plt.subplot(1,10, i+1)
    img = np.array( X[y==i][1:2] ).reshape(28,28)
    plt.imshow(img)

Draw Average Point of every class

m = y.unique().shape[0]
n = X.shape[1]

mu = np.zeros((m,n))
si = np.zeros((m,n))
for i in label:
    mu[i] = X[y==i].mean()
    si[i] = X[y==i].std()

plt.figure(figsize=(9,9))
for i in label:
    plt.subplot(3,4, i+1)
    img = np.array( mu[i] ).reshape(28,28)
    plt.imshow(img)

Naive Bayse Classifier

model = GaussianNB()

X_train, X_valid, y_train, y_valid = train_test_split(X, y, random_state=0)
model.fit(X_train, y_train)
yhat = model.predict(X_valid)

draw_confusionmatrix(y_valid, yhat)

Sum Axis-1 as Classification accuracy: 0.5652380952380952

Naive Bayes Classification on Histogram as Feature

Xtmp = np.array(X).reshape(X.shape[0], 28, 28)
Xaxis1 = Xtmp.sum(axis=1)
Xaxis2 = Xtmp.sum(axis=2)

1. Axis 1 As Feature

Xtrain, Xvalid, ytrain, yvalid = train_test_split(Xaxis1, y, random_state=0)
model.fit(Xtrain, ytrain)
yhat = model.predict(Xvalid)

draw_confusionmatrix(yvalid, yhat)

Sum Axis-1 as Classification accuracy: 0.31438095238095237

2. Axis 2 As Feature

Xtrain, Xvalid, ytrain, yvalid = train_test_split(Xaxis2, y)
model.fit(Xtrain, ytrain)
yhat = model.predict(Xvalid)

draw_confusionmatrix(yvalid, yhat)

Sum Axis-1 as Classification accuracy: 0.6517142857142857

3. Axis 1 and 2 As Feature

Xaxises = np.hstack((Xaxis1, Xaxis2))

Xtrain, Xvalid, ytrain, yvalid = train_test_split(Xaxises, y)
model.fit(Xtrain, ytrain)
yhat = model.predict(Xvalid)

draw_confusionmatrix(yvalid, yhat)

Sum Axis-1 as Classification accuracy: 0.5941904761904762

Image Filters

1. Sobel Filter

from scipy import ndimage
Xtmp = np.array(X).reshape(X.shape[0], 28, 28)
result = np.zeros_like(Xtmp)
for i in range(Xtmp.shape[0]):
    result[i] = ndimage.sobel(Xtmp[i])
    
result = result.reshape(result.shape[0], 28*28)

Xtrain, Xvalid, ytrain, yvalid = train_test_split(result, y)

model = GaussianNB()
model.fit(Xtrain, ytrain)
yhat = model.predict(Xvalid)
draw_confusionmatrix(yvalid, yhat)

Sum Axis-1 as Classification accuracy: 0.5404761904761904

2. Convolution

k = np.array([[1,1,1], [1,5,1], [1,1,1]])
images = np.zeros_like(Xtmp)
for i in range(Xtmp.shape[0]):
    images[i] = ndimage.convolve(Xtmp[i], k, mode='constant')

images = images.reshape(images.shape[0], 28*28)

Xtrain, Xvalid, ytrain, yvalid = train_test_split(result, y)

model = GaussianNB()
model.fit(Xtrain, ytrain)
yhat = model.predict(Xvalid)
draw_confusionmatrix(yvalid, yhat)

Sum Axis-1 as Classification accuracy: 0.540952380952381