sieci neuronowe - litery

This commit is contained in:
AgataWojciech 2021-06-21 01:43:32 +02:00
parent aa3e0b59c1
commit 5930cb6d3c
6 changed files with 102 additions and 46 deletions

BIN
src/litery/1.png Normal file

Binary file not shown.

After

Width:  |  Height:  |  Size: 383 B

BIN
src/litery/2.png Normal file

Binary file not shown.

After

Width:  |  Height:  |  Size: 402 B

BIN
src/litery/3.png Normal file

Binary file not shown.

After

Width:  |  Height:  |  Size: 403 B

BIN
src/litery/4.png Normal file

Binary file not shown.

After

Width:  |  Height:  |  Size: 454 B

BIN
src/litery/5.png Normal file

Binary file not shown.

After

Width:  |  Height:  |  Size: 376 B

View File

@ -1,91 +1,78 @@
from emnist import list_datasets
from emnist import extract_test_samples
from emnist import extract_training_samples
import numpy as np
import torch
from torch import nn
from torch import optim
import scipy.special
from matplotlib.pyplot import imshow
import glob
import imageio
""" pobranie obrazów cyfr i liter z biblioteki """
dig_train_images, dig_train_labels = extract_training_samples('digits')
dig_test_images, dig_test_labels = extract_test_samples('digits')
let_train_images, let_train_labels = extract_training_samples('letters')
let_test_images, let_test_labels = extract_test_samples('letters')
#print(dig_train_images[0])
""" przekształcenie tablic """
dig_train_images = dig_train_images.reshape(len(dig_train_images),28*28)
dig_test_images = dig_test_images.reshape(len(dig_test_images),28*28)
#print(d_test.shape)
print(d_labelstest)
#print(dig_train_images[0])
#print(dig_train_images.shape)
let_train_images = let_train_images.reshape(len(let_train_images),28*28)
let_test_images = let_test_images.reshape(len(let_test_images),28*28)
class NeuralNetwork:
""" inicjalizacja sieci neuronowej """
def __init__(self, inputNodes, hiddenNodes, outputNodes, learningGrade, fileWeight, fileHidden):
self.inodes = inputNodes
self.hnodes = hiddenNodes
self.onodes = outputNodes
"""te pierwsze dwa użyj przy nauce, potem zostaw cały czas te 2"""
#self.weights = (np.random.rand(self.hnodes, self.inodes) - 0.5)
#self.hidden = (np.random.rand(self.onodes, self.hnodes) - 0.5)
self.weights = np.load(fileWeight)
self.hidden = np.load(fileHidden)
#print( 'Matrix1 \n', self.weights)
#print( 'Matrix2 \n', self.hidden)
""" używane przy uczeniu sieci """
self.weights = (np.random.rand(self.hnodes, self.inodes) - 0.5)
self.hidden = (np.random.rand(self.onodes, self.hnodes) - 0.5)
""" używane przy pobieraniu danych o nauczonej sieci, z pliku """
# self.weights = np.load(fileWeight)
# self.hidden = np.load(fileHidden)
self.lr = learningGrade
""" funkcja aktywacji """
self.activationFunction = lambda x: scipy.special.expit(x)
pass
"""trening sieci neuronowej"""
def train(self, inputsList, targetsList):
""" konwersja list na tablice 2d """
inputs = np.array(inputsList,ndmin=2).T
targets = np.array(targetsList,ndmin=2).T
#forward pass
hiddenInputs = np.dot(self.weights, inputs)
""" forward pass """
hiddenInputs = np.dot(self.weights, inputs) # input -> hidden layer
hiddenOutputs = self.activationFunction(hiddenInputs)
finalInputs = np.dot(self.hidden, hiddenOutputs)
finalOutputs = self.activationFunction(finalInputs)
""" backward pass """
outputErrors = targets - finalOutputs
#print(outputErrors.shape)
x =self.weights.T
#print(x.shape)
hiddenErrors = np.dot(self.hidden.T, outputErrors)
#print('OutputErrors', outputErrors.shape)
#print('finalOutputs',finalOutputs.shape)
#print(x.shape)
self.hidden += self.lr * np.dot((outputErrors * finalOutputs * (1.0 - finalOutputs)) , np.transpose(hiddenOutputs))
self.weights += self.lr * np.dot((hiddenErrors * hiddenOutputs * (1.0 - hiddenOutputs)) , np.transpose(inputs))
pass
""" zapisywanie wytrenowanej sieci do pliku """
def saveTraining(self, fileWeight, fileHidden):
np.save(fileWeight, self.weights)
np.save(fileHidden, self.hidden)
""" wykorzystanie sieci """
def query(self, inputsList):
""" konwersja listy na tablicę 2d """
inputs = np.array(inputsList, ndmin=2).T
hiddenInputs = np.dot(self.weights, inputs)
hiddenOutputs = self.activationFunction(hiddenInputs)
@ -95,25 +82,26 @@ class NeuralNetwork:
return finalOutputs
""" dodaj tablicę literek"""
#n = NeuralNetwork(inputNodes=3, hiddenNodes=5, outputNodes=2, learningGrade=0.2)
""" tablice sieci neuronowych """
digitNetwork = NeuralNetwork(inputNodes=784, hiddenNodes=200, outputNodes=10, learningGrade=0.1, fileWeight="Dweights.npy", fileHidden="Dhidden.npy")
letterNetwork = NeuralNetwork(inputNodes=784, hiddenNodes=200, outputNodes=27, learningGrade=0.1, fileWeight="Lweights.npy", fileHidden="Lhidden.npy")
# trainNetwork(digitNetwork, "Dweights_test.npy", "Dhidden_test.npy", let_train_images, let_train_labels)
def trainNetwork(n, fWeight, fHidden, trainingSamples, trainingLabels):
epochs = 10
outputNodes = 10
outputNodes = 27
for e in range(epochs):
m=0
print('Epoch', e+1)
for record in trainingSamples:
""" zmiana wartości przedziału z [0,255] na [0,1] """
inputs = (np.asfarray(record[0:])/255 * 0.99) + 0.01
#print(inputs.shape)
targets = np.zeros(outputNodes) + 0.01
targets[trainingLabels[m]] = 0.99
#print(targets)
n.train(inputs,targets)
m+=1
@ -122,14 +110,82 @@ def trainNetwork(n, fWeight, fHidden, trainingSamples, trainingLabels):
n.saveTraining(fileWeight=fWeight, fileHidden=fHidden)
def testing(n, testingSamples, testingLabels):
scorecard = []
k = 0
for record in testingSamples:
inputs = (np.asfarray(record[0:])/255 * 0.99) + 0.01
correctLabels = testingLabels[k]
##################################### ODPALANIE TRAINING
#trainNetwork(digitNetwork, "Dweights.npy", "Dhidden.npy", dig_train_images, dig_train_labels)
outputs = n.query(inputs)
label = np.argmax(outputs)
#record = d_test[0]
#print('Label', d_labelstest[0])
#inputs = np.asfarray(record[0:])/ 255 * 0.99 + 0.01
#print(n.query(inputs))
if(label == correctLabels):
scorecard.append(1)
else:
scorecard.append(0)
k+=1
scorecardArray = np.asfarray(scorecard)
print('Performance', scorecardArray.sum() / scorecardArray.size)
testing(digitNetwork,dig_test_images,dig_test_labels)
#testing
li = []
ourOwnDataset = []
record_cache = None
for imageFileName in glob.glob('litery/?.png'):
label = int(imageFileName[-5:-4])
print('loading...', imageFileName)
imgArray = imageio.imread(imageFileName, as_gray=True)
#print(' imgArray: ', imgArray)
imgData = 255 - imgArray.reshape(784)
#print('imgData1: ',imgData)
imgData = (imgData/255 * 0.99) + 0.01
#print('imgData2: ',imgData)
#print(np.min(imgData))
#print(np.max(imgData))
record = np.append(label,imgData)
#print('Record: ',record)
ourOwnDataset.append(record)
if record_cache is None:
record_cache = record
#print(ood[0])
li.append(label)
pass
assert record_cache.shape == ourOwnDataset[0].shape
labelInput = np.asfarray(li)
#print(labelInput)
word = ""
for item in range(0,5):
correctLabels = labelInput[item]
outputs = letterNetwork.query(ourOwnDataset[item][1:])
print(outputs)
label = np.argmax(outputs)
print('label: ',label)
#print('Network says: ', label)
#labelString = np.array_str(label)
letters=['','a','b','c']
word = word + str(label)
print('slowo: ', word)
print('yep')
##################################### URUCHOMIENIE TRENINGU
trainNetwork(letterNetwork, "Lweights_test.npy", "Lhidden_test.npy", let_train_images, let_train_labels)
# trainNetwork(digitNetwork, "Dweights_test.npy", "Dhidden_test.npy", let_train_images, let_train_labels)