<?php
import cv2
import pandas as pd
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
from sklearn.neural_network import MLPClassifier
from sklearn.metrics import classification_report
scaler = StandardScaler()
surf = cv2.xfeatures2d.SURF_create(300,upright=True, extended = True)
def read_train_image_by_id(id):
return cv2.imread('./train/'+str(id)+'.png')
def read_test_image_by_id(id):
return cv2.imread('/test/'+str(id)+'.png')
def read_train_data(df):
#should append descriptors and keypoints to the dataframe
descriptors = []
key_points = []
for x in df.iterrows():
img = read_train_image_by_id(x[1]['id'])
kp, des = surf.detectAndCompute(img,None)
descriptors.append(des)
key_points.append(kp)
df['descriptors'] = pd.Series(descriptors)
df['keyPoints'] = pd.Series(key_points)
return
#initial values
neurons_per_layer = (128,)
hidden_layer_increment = 1
currrent_accuracy = 0.
min_accepted_accuracy = 0.9
learning_rate = 1e-5
print('neurons per layer: '+str(neurons_per_layer))
print('hidden layers increment: '+str(hidden_layer_increment))
print('learning rate:'+str(learning_rate))
#load training data, it contains the ID and correct label
data = pd.read_csv('./trainLabels.csv')
#convert label field from string to number so it works with neural network
data['label'], labelMap = pd.factorize(data['label'])
print('got: '+str(len(labelMap))+' object labels')
print('reading train data')
#fill dataframe with the actual image data
read_train_data(data)
data = data[(data['descriptors'].notnull())]
print('done reading data, splitting them to train/test sets')
#split data to train and test
train, test = train_test_split(data, test_size=0.2)
##following part isn't working, some why fit doesn't work, it throws an exception.
#scaler.fit(train['descriptors'])
#train['descriptors'] = scaler.transform(train['descriptors'])
#test['descriptors'] = scaler.transform(test['descriptors'])
#X = train['descriptors'] is the input for training network
#Y = train['label'] is the correct labels for training network.
iteration = 0
while currrent_accuracy < 0.9:
#keep increasing hidden layer count until reaching acceptable accuracy
iteration+=1
print('iteration: '+str(iteration))
clf = MLPClassifier(
hidden_layer_sizes=neurons_per_layer*(iteration*hidden_layer_increment),
activation='tanh',
solver='lbfgs',
alpha=learning_rate,
random_state=1337
)
#an exception is being thrown here: ValueError: setting an array element with a sequence.
clf.fit(train['descriptors'], train['label'])
predictions = clf.predict(test['descriptors'])
result = classification_report(test['label'],predictions)
print(result)
#check result and determine what is the accuracy, then updated it.
input('click to start next iteration...')
