sidSimpleNN 0.2.8

Multilayered Perceptron NN light weight, make, train, test models

pip install sidSimpleNN

you can also visit my website for testing manual Digit and Letters Recognition https://siddhantofficialsidsimpl-868fa.web.app you can also train your model through this lib and upload your weights and bias json file to this site to test your figures(letter, shape etc) recognition

In data folder Can be downloaded from http://yann.lecun.com/exdb/mnist/

net.chosenLoss='CE' ## for classification use CE and for regression use MSE net.chosenActivation='relu' ## can choose from relu, tanh, sigmoid net.applySoftmax=True ## for classification make this true, gives output as probability net.applyRegularization=False ## apply L2 regularization net.calcLoss=False ## should calculate loss net.otherImplementation=False ## for non batch wise training net.saveWeightsBiasJSON() ##save WB file for using it in my website net.save() ## to save model net.load() ## to load the model you need to have same name as saved model net.showModel() ## details of model net.changeActivation() ## to apply choosen activation net.lossGraph() ## to display loss vs iter graph net.feedforward(input) ## to get the last layer output

from sidSimpleNN import mainTrain

mainTrain.run()

if data,mnist folder dont get made automatically then manualy download from MNIST dataset and place in datafolder

if above code doesnt work then test with this

import numpy as np import mnist import sidSimpleNN.myNN as myNN

def run():

# load data
num_classes = 10
train_images = mnist.train_images() #[60000, 28, 28]
train_labels = mnist.train_labels()
test_images = mnist.test_images()
test_labels = mnist.test_labels()

# print("Training...")

# # data processing
X_train = train_images.reshape(train_images.shape[0], train_images.shape[1]*train_images.shape[2]).astype('float32') #flatten 28x28 to 784x1 vectors, [60000, 784]
x_train = X_train / 255 #normalization
y_train = np.eye(num_classes)[train_labels] #convert label to one-hot

X_test = test_images.reshape(test_images.shape[0], test_images.shape[1]*test_images.shape[2]).astype('float32') #flatten 28x28 to 784x1 vectors, [60000, 784]
x_test = X_test / 255 #normalization
y_test = test_labels

np.random.seed(1)

net = myNN.Network(
             num_nodes_in_layers = [784, 10,20, 10], 
             batch = 1,
             epochs = 6,
             learning_rate = 0.001, 
             weights_file=None,
             chosenActivation='tanh'
             # name='tempnet3'
             # weightsAndBias_file = 'digitRecog',
             # includeWeightsBias=True

         )


print('before training , testing with test dataset')
# net.chosenLoss='CE'
# net.applySoftmax=True
# net.applyRegularization=False
# net.regularizationConst=0.01
# net.calcLoss=False
# net.saveWeightsBiasJSON()
message=net.test(x_test, y_test)


net.train(x_train, y_train)
net.save()
net.lossGraph()
print('before training , testing with test dataset')
print(message)

print("after training")
net.test(x_test, y_test)
net.showModel()

if name=='main': run()