Saturday, April 6, 2019

Batch Normalization Adam Optimizer and L2 reg and Dropout

Batch Normalization with Adam and L2 Reg and Dropout

Batch Normalization Adam Optimizer and L2 reg and Dropout

Training Deep Neural Networks is complicated by the fact that the distribution of each layer's inputs changes during training, as the parameters of the previous layers change. This slows down the training by requiring lower learning rates and careful parameter initialization, and makes it notoriously hard to train models with saturating nonlinearities. We refer to this phenomenon as internal covariate shift, and address the problem by normalizing layer inputs.we would use MNIST data set to test our algorithm.
Article written Soumil Shah

Step 1: Import the Necessary Library

In [ ]:
try:
    import os
    import sys
    import cv2
    import numpy as np
    from sklearn.utils import shuffle
    import matplotlib.pyplot as plt
    import time
    import datetime
    import math
    import cmath

except:
    print("Library not found ")
Lets us create a list for number of epoch and loss so that we can append our data in list and plot later on using matplotlib and also we want to know execution time taken for that we create a start time.
In [ ]:
now_time = datetime.datetime.now()

x_data_epoch = []
y_data_error = []


train = np.empty((1000,28,28), dtype='float64')
trainY = np.zeros((1000,10))

test = np.empty((10000, 28, 28), dtype='float64')
testY = np.zeros((10000, 10))

Step 2 : Load Train Image

In [ ]:
i = 0
for filename in os.listdir('/Users/soumilshah/IdeaProjects/Deep Learning/MNIST /Training1000'):
    y = int(filename[0])

    trainY[i, y] = 1.0

    train[i] = cv2.imread('/Users/soumilshah/IdeaProjects/Deep Learning/MNIST /Training1000/{0}'.format(filename), 0)/255.0
    i = i+1

Step 3 : Load Test Image

In [ ]:
i = 0
# read test data
for filename in os.listdir('/Users/soumilshah/IdeaProjects/Deep Learning/MNIST /Test10000'):
    y = int(filename[0])

    testY[i, y] = 1.0

    test[i] = cv2.imread('/Users/soumilshah/IdeaProjects/Deep Learning/MNIST /Test10000/{0}'.format(filename), 0)/255.0
    i = i+1

Step 4 : Declare the Parameters

In [ ]:
trainX = train.reshape(train.shape[0], train.shape[1]*train.shape[2])
testX = test.reshape(test.shape[0], test.shape[1] * test.shape[2])
In [ ]:
numNeuronsLayer1 = 50                  # Number of Neurons in Layer 1
numNeuronsLayer2 = 10                   # Number of Output Neurons
numEpochs = 20                          # NUmber of Epoch
learningRate = 0.2                     # Learning  Rate
mini_batch_size = 25                   # Batch size

Step 5: Define Weight Matrix and Bias Matrix

In [ ]:
w1 = np.random.uniform(low=-0.1,high=0.1,size=(numNeuronsLayer1,784))
b1 = np.random.uniform(low=-1,high=1, size=(numNeuronsLayer1))
w2 = np.random.uniform(low=- 0.1,high=0.1, size=(numNeuronsLayer2, numNeuronsLayer1))
b2 = np.random.uniform(low=-0.1,high=0.1, size=(numNeuronsLayer2))

gamma = np.random.uniform(low=-0.1,high=0.1, size=(numNeuronsLayer1))
beta = np.random.uniform(low=-0.1,high=0.1, size=(numNeuronsLayer1))

Step 6: Define Optimization Parameters

In [ ]:
l2_reg = 1e-5
zero_out = np.random.binomial(n=1, p=.8, size= (numNeuronsLayer1))/0.8

Step 7: Define the Adam Parameters

In [ ]:
beta_momentum = 0.9
beta_rms = 0.99
l2_reg = 1e-5
epsilon = 1e-8
EPS = 1e-8


runningsigma2 = 0
runningmu = 0

sdw1 = np.zeros_like(w1)
sdw2 = np.zeros_like(w2)
sdb1 = np.zeros_like(b1)
sdb2 = np.zeros_like(b2)

vdw1 = np.zeros_like(w1)
vdw2 = np.zeros_like(w2)
vdb1 = np.zeros_like(b1)
vdb2 = np.zeros_like(b2)
Now we write our First for loop that is for the number of Epoch
In [ ]:
for n in range(0, numEpochs):

    loss = 0
    trainX, trainY = shuffle(trainX, trainY)

    gradw2 = np.zeros((numNeuronsLayer2, numNeuronsLayer1))
    gradw1 = np.zeros((numNeuronsLayer1, 784))
    gradb1 = np.zeros((numNeuronsLayer1))
    gradb2 = np.zeros((numNeuronsLayer2))

    dgamma = np.zeros((numNeuronsLayer1))
    dbeta = np.zeros((numNeuronsLayer1))
For second For loop that is Mini Batch size

Step 8: Foreward Pass

In [ ]:
    for i in range(0, int(trainX.shape[0]), mini_batch_size):        # This loop Iterate for 20 Times
        X_train_mini = trainX[i:i + mini_batch_size]
        y_train_mini = trainY[i:i + mini_batch_size]
        s1 = np.dot(X_train_mini, w1.T) + b1
Add Bacth Normalization parameters
In [ ]:
        UB = np.mean(s1, axis=0)
        sigma2 = np.var(s1,axis=0)
        Shat = (s1 - UB)/np.sqrt(sigma2 + epsilon)
        Sb = Shat * gamma + beta

        runningmu = 0.9 * runningmu + (1 - 0.9) * UB
        runningsigma2 = 0.9 * runningsigma2 + (1 - 0.9) * sigma2

        a1 = 1 / (1 + np.exp(-1 * Sb))
Add L2 Reg and drop out
In [ ]:
        a1 = np.multiply(a1, zero_out)

        s2 = np.dot(a1, w2.T) + b2
        a2 = 1 / (1 + np.exp(-1 * s2))
Calculate Loss
In [ ]:
        W = np.sum((w1)) + np.sum((w2))
        reg_cost = 0.5 * l2_reg * W * W
        loss = loss + (0.5 * np.multiply((y_train_mini - a2), (y_train_mini - a2))).sum() + reg_cost

Step 9: Back Propogate

In [ ]:
        derivative_error_2 = - (y_train_mini - a2)
        derv_act_2 = np.multiply(a2, (1 - a2))
        delta2 = np.multiply(derivative_error_2, derv_act_2)

        derv_act_1 = np.multiply(a1, (1 - a1))
        derv_act_1 = np.multiply(derv_act_1, zero_out)

        error_2 = np.dot(delta2, w2)
        delta1 = np.multiply(error_2, derv_act_1)
Calculate Delta BN
In [ ]:
        deltabn = (delta1 *gamma)/(mini_batch_size * np.sqrt(sigma2 + epsilon)) * (mini_batch_size - 1 - (Shat *Shat))
Calculate Grad
In [ ]:
        gradw2 = np.dot(delta2.T, a1)
        gradw1 = np.dot(deltabn.T, X_train_mini)

        gradb2 = np.sum(delta2, axis=0)
        gradb1 = np.sum(deltabn, axis=0)
In [ ]:
        dbeta = np.sum(delta1, axis=0)
        dgamma =np.sum(delta1*Shat,axis=0)
Momentum
In [ ]:
        # ---------------- MOMENTUM ------------------------------
        vdw2 = beta_momentum * vdw2 + ((1 - beta_momentum) * gradw2)

        vdw1 = beta_momentum * vdw1 + ((1 - beta_momentum) * gradw1)

        vdb1 = beta_momentum * vdb1 + ((1 - beta_momentum) * gradb1)
        vdb2 = beta_momentum * vdb2 + ((1 - beta_momentum) * gradb2)
RMS PROP
In [ ]:
        # ---------------- MOMENTUM ------------------------------
        vdw2 = beta_momentum * vdw2 + ((1 - beta_momentum) * gradw2)

        vdw1 = beta_momentum * vdw1 + ((1 - beta_momentum) * gradw1)

        vdb1 = beta_momentum * vdb1 + ((1 - beta_momentum) * gradb1)
        vdb2 = beta_momentum * vdb2 + ((1 - beta_momentum) * gradb2)

        # RMS ------------------------------------------------------
        sdw2 = beta_rms * sdw2 + ((1 - beta_rms) * (gradw2*gradw2))
        sdw1 = beta_rms * sdw1 + ((1 - beta_rms) * (gradw1 * gradw1))

        sdb1 = beta_rms * sdb1 + ((1 - beta_rms) * (gradb1*gradb1))
        sdb2 = beta_rms * sdb2 + ((1 - beta_rms) * (gradb2 * gradb2))
Add Bias Correction
In [ ]:
        VDW1_CORRECTED = vdw1 / (1 - beta_momentum)
        VDW2_CORRECTED = vdw2 / (1 - beta_momentum)

        VDB1_CORRECTED = vdb1 / (1 - beta_momentum)
        VDB2_CORRECTED = vdb2 / (1 - beta_momentum)

        # ------------------------------------------

        SDW1_CORRECTED = sdw1 / (1 - beta_rms)
        SDW2_CORRECTED = sdw2 / (1 - beta_rms)

        SDB1_CORRECTED = sdb1 / (1 - beta_rms)
        SDB2_CORRECTED = sdb2 / (1 - beta_rms)

Step 10: Update Weight and Bias and other Parameter

In [ ]:
        w2 = w2 - learningRate * (1/mini_batch_size) \
             * (VDW2_CORRECTED) * (1/(np.sqrt(SDW2_CORRECTED)+EPS)) \
             - (1/mini_batch_size) * learningRate * l2_reg * W

        w1 = w1 - learningRate *(1/mini_batch_size) * (1/mini_batch_size) * \
             (VDW1_CORRECTED) * (1/(np.sqrt(SDW1_CORRECTED)+EPS)) \
             - (1/mini_batch_size) * learningRate * l2_reg * W

        b1 = b1 - learningRate * (1/mini_batch_size) * \
             (VDB1_CORRECTED) * (1/(np.sqrt(SDB1_CORRECTED)+EPS)) \
             - (1/mini_batch_size) * learningRate * l2_reg * W

        b2 = b2 - learningRate * (1/mini_batch_size) * \
             (VDB2_CORRECTED) * (1/(np.sqrt(SDB2_CORRECTED)+EPS)) \
             - (1/mini_batch_size) * learningRate * l2_reg * W

        beta = beta - learningRate * dbeta
        gamma = gamma -learningRate * dgamma
Print Loss
In [ ]:
    print("epoch = " + str(n) + " loss = " + (str(loss)))
In [ ]:
print("done training , starting testing..")
accuracyCount = 0

Step 11: Test Image

In [ ]:
for i in range(testY.shape[0]):
    s1 = np.dot(w1, testX[i]) + b1
    mu = runningmu
    sigma2 = runningsigma2

    Shat = (s1 - mu) / np.sqrt(sigma2 + EPS)
    sb = Shat * gamma + beta

    a1 = 1 / (1 + np.exp(-1 * sb))  # np.exp operates on the array

    s2 = np.dot(w2, a1) + b2
    a2 = 1 / (1 + np.exp(-1 * s2))

    a2index = a2.argmax(axis=0)             # Calculate Maximum of these 10 Neuron

    if testY[i, a2index] == 1:
        accuracyCount = accuracyCount + 1
        print("Accuracy count = " + str(accuracyCount/10000.0))

Putting all Together

In [3]:
try:
    import os
    import sys
    import cv2
    import numpy as np
    from sklearn.utils import shuffle
    import matplotlib.pyplot as plt
    import time
    import datetime
    import math
    import cmath

except:
    print("Library not found ")

now_time = datetime.datetime.now()

x_data_epoch = []
y_data_error = []


train = np.empty((1000,28,28), dtype='float64')
trainY = np.zeros((1000,10))

test = np.empty((10000, 28, 28), dtype='float64')
testY = np.zeros((10000, 10))

# --------------------------------------------------------Load in Image--------------------------------------
i = 0
for filename in os.listdir('/Users/soumilshah/IdeaProjects/Deep Learning/MNIST /Training1000'):
    y = int(filename[0])

    trainY[i, y] = 1.0

    train[i] = cv2.imread('/Users/soumilshah/IdeaProjects/Deep Learning/MNIST /Training1000/{0}'.format(filename), 0)/255.0
    i = i+1

# -------------------------------------------------LOAD TEST IMAGE ------------------------------------------------

i = 0
# read test data
for filename in os.listdir('/Users/soumilshah/IdeaProjects/Deep Learning/MNIST /Test10000'):
    y = int(filename[0])

    testY[i, y] = 1.0

    test[i] = cv2.imread('/Users/soumilshah/IdeaProjects/Deep Learning/MNIST /Test10000/{0}'.format(filename), 0)/255.0
    i = i+1
# ---------------------------------------------------------------------------------------------------------------------

trainX = train.reshape(train.shape[0], train.shape[1]*train.shape[2])
testX = test.reshape(test.shape[0], test.shape[1] * test.shape[2])

# ---------------------------------  Neural Network Parameter  ------------------------------------

numNeuronsLayer1 = 50                  # Number of Neurons in Layer 1
numNeuronsLayer2 = 10                   # Number of Output Neurons
numEpochs = 20                          # NUmber of Epoch
learningRate = 0.2                     # Learning  Rate
mini_batch_size = 25                   # Batch size

w1 = np.random.uniform(low=-0.1,high=0.1,size=(numNeuronsLayer1,784))
b1 = np.random.uniform(low=-1,high=1, size=(numNeuronsLayer1))
w2 = np.random.uniform(low=- 0.1,high=0.1, size=(numNeuronsLayer2, numNeuronsLayer1))
b2 = np.random.uniform(low=-0.1,high=0.1, size=(numNeuronsLayer2))

gamma = np.random.uniform(low=-0.1,high=0.1, size=(numNeuronsLayer1))
beta = np.random.uniform(low=-0.1,high=0.1, size=(numNeuronsLayer1))

zero_out = np.random.binomial(n=1, p=0.8, size=(numNeuronsLayer1)) / 0.8

beta_momentum = 0.9
beta_rms = 0.99
l2_reg = 1e-5
epsilon = 1e-8
EPS = 1e-8


runningsigma2 = 0
runningmu = 0

sdw1 = np.zeros_like(w1)
sdw2 = np.zeros_like(w2)
sdb1 = np.zeros_like(b1)
sdb2 = np.zeros_like(b2)

vdw1 = np.zeros_like(w1)
vdw2 = np.zeros_like(w2)
vdb1 = np.zeros_like(b1)
vdb2 = np.zeros_like(b2)

# =======================================================================================================
for n in range(0, numEpochs):

    loss = 0
    trainX, trainY = shuffle(trainX, trainY)

    gradw2 = np.zeros((numNeuronsLayer2, numNeuronsLayer1))
    gradw1 = np.zeros((numNeuronsLayer1, 784))
    gradb1 = np.zeros((numNeuronsLayer1))
    gradb2 = np.zeros((numNeuronsLayer2))

    dgamma = np.zeros((numNeuronsLayer1))
    dbeta = np.zeros((numNeuronsLayer1))

    for i in range(0, int(trainX.shape[0]), mini_batch_size):        # This loop Iterate for 20 Times

        X_train_mini = trainX[i:i + mini_batch_size]
        y_train_mini = trainY[i:i + mini_batch_size]
        s1 = np.dot(X_train_mini, w1.T) + b1

        # -------------------------------

        UB = np.mean(s1, axis=0)
        sigma2 = np.var(s1,axis=0)
        Shat = (s1 - UB)/np.sqrt(sigma2 + epsilon)
        Sb = Shat * gamma + beta

        runningmu = 0.9 * runningmu + (1 - 0.9) * UB
        runningsigma2 = 0.9 * runningsigma2 + (1 - 0.9) * sigma2

        a1 = 1 / (1 + np.exp(-1 * Sb))
        a1 = np.multiply(a1, zero_out)

        s2 = np.dot(a1, w2.T) + b2
        a2 = 1 / (1 + np.exp(-1 * s2))

        W = np.sum((w1)) + np.sum((w2))
        reg_cost = 0.5 * l2_reg * W * W
        loss = loss + (0.5 * np.multiply((y_train_mini - a2), (y_train_mini - a2))).sum() + reg_cost



        # =======BACK PROPOGATION ==================================

        derivative_error_2 = - (y_train_mini - a2)
        derv_act_2 = np.multiply(a2, (1 - a2))
        delta2 = np.multiply(derivative_error_2, derv_act_2)

        derv_act_1 = np.multiply(a1, (1 - a1))
        derv_act_1 = np.multiply(derv_act_1, zero_out)

        error_2 = np.dot(delta2, w2)
        delta1 = np.multiply(error_2, derv_act_1)

        deltabn = (delta1 *gamma)/(mini_batch_size * np.sqrt(sigma2 + epsilon)) * (mini_batch_size - 1 - (Shat *Shat))

        gradw2 = np.dot(delta2.T, a1)
        gradw1 = np.dot(deltabn.T, X_train_mini)

        gradb2 = np.sum(delta2, axis=0)
        gradb1 = np.sum(deltabn, axis=0)

        dbeta = np.sum(delta1, axis=0)
        dgamma =np.sum(delta1*Shat,axis=0)

        # ---------------- MOMENTUM ------------------------------
        vdw2 = beta_momentum * vdw2 + ((1 - beta_momentum) * gradw2)

        vdw1 = beta_momentum * vdw1 + ((1 - beta_momentum) * gradw1)

        vdb1 = beta_momentum * vdb1 + ((1 - beta_momentum) * gradb1)
        vdb2 = beta_momentum * vdb2 + ((1 - beta_momentum) * gradb2)

        # RMS ------------------------------------------------------
        sdw2 = beta_rms * sdw2 + ((1 - beta_rms) * (gradw2*gradw2))
        sdw1 = beta_rms * sdw1 + ((1 - beta_rms) * (gradw1 * gradw1))

        sdb1 = beta_rms * sdb1 + ((1 - beta_rms) * (gradb1*gradb1))
        sdb2 = beta_rms * sdb2 + ((1 - beta_rms) * (gradb2 * gradb2))

        VDW1_CORRECTED = vdw1 / (1 - beta_momentum)
        VDW2_CORRECTED = vdw2 / (1 - beta_momentum)

        VDB1_CORRECTED = vdb1 / (1 - beta_momentum)
        VDB2_CORRECTED = vdb2 / (1 - beta_momentum)

        # ------------------------------------------

        SDW1_CORRECTED = sdw1 / (1 - beta_rms)
        SDW2_CORRECTED = sdw2 / (1 - beta_rms)

        SDB1_CORRECTED = sdb1 / (1 - beta_rms)
        SDB2_CORRECTED = sdb2 / (1 - beta_rms)


    # ======================After BATCH set is done Upgrade WEIGHT===============================

        w2 = w2 - learningRate * (1/mini_batch_size) \
             * (VDW2_CORRECTED) * (1/(np.sqrt(SDW2_CORRECTED)+EPS)) \
             - (1/mini_batch_size) * learningRate * l2_reg * W

        w1 = w1 - learningRate *(1/mini_batch_size) * (1/mini_batch_size) * \
             (VDW1_CORRECTED) * (1/(np.sqrt(SDW1_CORRECTED)+EPS)) \
             - (1/mini_batch_size) * learningRate * l2_reg * W

        b1 = b1 - learningRate * (1/mini_batch_size) * \
             (VDB1_CORRECTED) * (1/(np.sqrt(SDB1_CORRECTED)+EPS)) \
             - (1/mini_batch_size) * learningRate * l2_reg * W

        b2 = b2 - learningRate * (1/mini_batch_size) * \
             (VDB2_CORRECTED) * (1/(np.sqrt(SDB2_CORRECTED)+EPS)) \
             - (1/mini_batch_size) * learningRate * l2_reg * W

        beta = beta - learningRate * dbeta
        gamma = gamma -learningRate * dgamma

        # ====================================

    print("epoch = " + str(n) + " loss = " + (str(loss)))

print("done training , starting testing..")
accuracyCount = 0


# --------------------------------------  Testing Neural Network ------------------------------------------------------

for i in range(testY.shape[0]):
    s1 = np.dot(w1, testX[i]) + b1
    mu = runningmu
    sigma2 = runningsigma2

    Shat = (s1 - mu) / np.sqrt(sigma2 + EPS)
    sb = Shat * gamma + beta

    a1 = 1 / (1 + np.exp(-1 * sb))  # np.exp operates on the array

    s2 = np.dot(w2, a1) + b2
    a2 = 1 / (1 + np.exp(-1 * s2))

    a2index = a2.argmax(axis=0)             # Calculate Maximum of these 10 Neuron

    if testY[i, a2index] == 1:
        accuracyCount = accuracyCount + 1
print("Accuracy count = " + str(accuracyCount/10000.0))

epoch = 0 loss = 532.1756796162031
epoch = 1 loss = 480.7396258331274
epoch = 2 loss = 455.4973894594472
epoch = 3 loss = 448.2418784303272
epoch = 4 loss = 439.37509369610217
epoch = 5 loss = 406.42522590455377
epoch = 6 loss = 343.63774610253245
epoch = 7 loss = 280.13154499851544
epoch = 8 loss = 238.71309487232813
epoch = 9 loss = 209.146081391587
epoch = 10 loss = 186.92705706084968
epoch = 11 loss = 171.581760438389
epoch = 12 loss = 154.38057383771329
epoch = 13 loss = 145.30534933940683
epoch = 14 loss = 133.96395146575767
epoch = 15 loss = 126.2422918427009
epoch = 16 loss = 118.70752047102664
epoch = 17 loss = 110.32233689559683
epoch = 18 loss = 106.05267423794864
epoch = 19 loss = 99.76414970373104
done training , starting testing..
Accuracy count = 0.8727
In [ ]:
at

No comments:

Post a Comment

Subscribe to: Post Comments (Atom)

How to Use Publish-Audit-Merge Workflow in Apache Iceberg: A Beginner’s Guide

publish How to Use Publish-Audit-Merge Workflow in Apache Iceberg: A Beginner’s Guide ¶ In [24]: from ...