Hand Written Recognition with Gradient Boosting

Instruction

Make the following three folders. The notebook should be in the same path as these three folders

• train images: Containing few subfolders. Each subfolder contains images of a class. In this example, data has two subfolders: 1 (containing number 1 images) and 2 (containing number 2 images).

• test images: Containing only images for testing. There is NO subfolder within test images

• The notebook should be in the same folder as these two above folders.

Import some packages

import numpy as np
from sklearn.ensemble import RandomForestClassifier
import matplotlib.pyplot as plt
import seaborn as sns; sns.set()
import matplotlib
 
# import the necessary packages
from sklearn.model_selection import train_test_split
import os
import tkinter
from os import listdir
from PIL import Image

# Set image resolution
dim = 200

# Set the gradient boosting
number_of_trees = 300
learning_rate = .05
depth_of_each_tree = 1 # if depth_of_each_tree=1, we have a stump!

train_folder = 'train images'
test_folder = 'test images'
lab = os.listdir(train_folder)

Xtrain = np.empty((0,3*dim*dim), float)
ytrain = np.empty((0,), int)

for i in range(0, len(lab)): 
    path = train_folder + '/' + lab[i]
    names = [f for f in os.listdir(path)]
    for n in names[0:]:
        img = Image.open(path +'/'+ n)
        image = img.convert('RGB')
        image = image.resize((dim, dim))
        image = np.array(image)
        image = image.ravel().reshape(1,-1)
        Xtrain = np.append(Xtrain, image,0)
        ytrain = np.append(ytrain, i)  

Training

from sklearn.ensemble import GradientBoostingClassifier

model = GradientBoostingClassifier(n_estimators=number_of_trees, 
                                   learning_rate=learning_rate, 
                                   max_depth=depth_of_each_tree)
model.fit(Xtrain, ytrain)

GradientBoostingClassifier(learning_rate=0.05, max_depth=1, n_estimators=300)

Accuracy on Training Data

ypred=model.predict(Xtrain)
from sklearn.metrics import confusion_matrix
mat = confusion_matrix(ypred, ytrain)
mat

array([[20,  0,  0],
       [ 0, 29,  0],
       [ 0,  0, 23]], dtype=int64)

Predicting New Images

In the next codes, we will use our fitted model to classify new images of dogs and cats. The model will label as Dog or Cat. These labeled images is then save in the predict folder.

onlyfiles = [f for f in listdir(test_folder)]

import matplotlib.pyplot as plt

plt.figure(figsize=(10, 10))

for i, images in enumerate(onlyfiles):
    
    ax = plt.subplot(int(1+len(onlyfiles)/3), 3, i + 1)
    img = Image.open(test_folder+'/'+images)
    image = img.convert('RGB')
    image = image.resize((dim, dim))
    image_print = image.resize((dim, dim))
    image = np.array(image)
    image = image.ravel().reshape(1,-1)

    predictions = model.predict(image)
    
    confidence = (100*model.predict_proba(image).max()).round(1)
    
    plt.imshow(image_print)
    plt.title(str(confidence)+'% '+lab[predictions[0]])
    plt.axis("off")