def nesterov_momentum(learning_rate):
X, Y = transform_data()
X, Y = shuffle(X, Y)
N = len(X) // 2
Xtrain = X[:N]
Ytrain = Y[:N]
Ttrain = generate_T(Ytrain)
Xtest = X[N:]
Ytest = Y[N:]
Ttest = generate_T(Ytest)
N, D = Xtrain.shape
M = 100
K = len(set(Y))
iterations = 50
batch_N = 250
batches = N // batch_N
v = np.random.randn(M, K) / np.sqrt(M + K)
b_1 = np.random.randn(K) / np.sqrt(K)
w = np.random.randn(D, M) / np.sqrt(D + M)
b_0 = np.random.randn(M) / np.sqrt(M)
mu = .9
dv = 0
db_1 = 0
dw = 0
db_0 = 0
nesterov_cost = []
nesterov_cr = []
best_nesterov = 0
best_iteration = 0
for i in range(iterations):
for b in range(batches):
X = Xtrain[b * batch_N:(b + 1) * batch_N, :]
T = Ttrain[b * batch_N:(b + 1) * batch_N, :]
Y, Z = generate_Y('tanh', X, w, b_0, v, b_1)
Y_test, _ = generate_Y('tanh', Xtest, w, b_0, v, b_1)
P_test = np.argmax(Y_test, axis=1)
if b % batches == 0:
nesterov_cost.append(cross_entropy(Y_test, Ttest))
cr = classification_rate(P_test, Ytest)
nesterov_cr.append(cr)
if cr > best_nesterov:
best_nesterov = cr
best_iteration = i
dv = mu * dv - learning_rate * derivative_v('tanh', Z, Y, T)
db_1 = mu * db_1 - learning_rate * derivative_b1('tanh', Y, T)
dw = mu * dw - learning_rate * derivative_w('tanh', X, Y, Z, T, v)
db_0 = mu * db_0 - learning_rate * derivative_b0(
'tanh', Y, Z, T, v)
v += mu * dv - learning_rate * derivative_v('tanh', Z, Y, T)
b_1 += mu * db_1 - learning_rate * derivative_b1('tanh', Y, T)
w += mu * dw + learning_rate * derivative_w('tanh', X, Y, Z, T, v)
b_0 += mu * db_0 - learning_rate * derivative_b0(
'tanh', Y, Z, T, v)
if i % 100 == 0:
print('Nesterov Cost: ', nesterov_cost[i],
'Nesterov Classification: ', nesterov_cr[i])
return nesterov_cost, nesterov_cr, best_nesterov, best_iteration
def exp_decay(learning_rate):
X, Y = transform_data()
X, Y = shuffle(X, Y)
N = len(X) // 2
Xtrain = X[:N]
Ytrain = Y[:N]
Ttrain = generate_T(Ytrain)
Xtest = X[N:]
Ytest = Y[N:]
Ttest = generate_T(Ytest)
N, D = Xtrain.shape
M = 100
K = len(set(Y))
iterations = 50
batch_N = 250
batches = N // batch_N
dv = 0
d_b1 = 0
dw = 0
d_b0 = 0
mu = .9
v = np.random.randn(M, K) / np.sqrt(M + K)
b_1 = np.random.randn(K) / np.sqrt(K)
w = np.random.randn(D, M) / np.sqrt(D + M)
b_0 = np.random.randn(M) / np.sqrt(M)
learning_rate = learning_rate
exp_cost = []
exp_cr = []
exp_lr = []
best_exp = 0
best_iteration = 0
for i in range(iterations):
learning_rate = learning_rate * np.exp(-K * i)
exp_lr.append(learning_rate)
for b in range(batches):
X = Xtrain[b * batches:(b + 1) * batches, :]
T = Ttrain[b * batches:(b + 1) * batches, :]
Y, Z = generate_Y('tanh', X, w, b_0, v, b_1)
Y_test, _ = generate_Y('tanh', Xtest, w, b_0, v, b_1)
P_test = np.argmax(Y_test, axis=1)
if b % batches == 0:
exp_cost.append(cross_entropy(Y_test, Ttest))
cr = classification_rate(P_test, Ytest)
exp_cr.append(cr)
if cr > best_exp:
best_exp = cr
best_iteration = i
dv = mu * dv - learning_rate * derivative_v('tanh', Z, Y, T)
d_b1 = mu * d_b1 - learning_rate * derivative_b1('tanh', Y, T)
dw = mu * dw - learning_rate * derivative_w('tanh', X, Y, Z, T, v)
d_b0 = mu * d_b0 - learning_rate * derivative_b0(
'tanh', Y, Z, T, v)
v += dv
b_1 += d_b1
w += dw
b_0 += d_b0
if i % 10 == 0:
print('Exp Cost: ', exp_cost[i], 'Exp Classification: ', exp_cr[i])
return exp_cost, exp_cr, exp_lr, best_exp, best_iteration
def grid():
X,Y = transform_data()
X,Y = shuffle(X,Y)
N = len(X)//2
Xtrain = X[:N]
Ytrain = Y[:N]
Ttrain = generate_T(Ytrain)
Xtest = X[N:]
Ytest = Y[N:]
Ttest = generate_T(Ytest)
N,D = Xtrain.shape
K = len(set(Y))
w0 = np.random.randn(D,K)/np.sqrt(D+K)
b0 = np.random.randn(K)/np.sqrt(K)
learning_rates = [10**i for i in range(-7,-3,1)]
momentums = [1-10**i for i in sorted(list(range(-4,0)),reverse=True)]
iterations = 2000
best_lr = 0
best_momentum = 0
best_cr = 0
cost = {}
cr = {}
for lr in learning_rates:
learning_rate = lr
for mu in momentums:
dw = 0
db = 0
cost[(lr,mu)] = list()
cr[(lr,mu)] = list()
for i in range(iterations):
if i == 0:
A_train = relu(Xtrain.dot(w0) + b0)
A_test = relu(Xtest.dot(w0) + b0)
else:
A_train = relu(Xtrain.dot(w) + b0)
A_test = relu(Xtest.dot(w) + b0)
Y_train = np.exp(A_train)/np.exp(A_train).sum(axis=1,keepdims=True)
Y_test = np.exp(A_test)/np.exp(A_test).sum(axis=1,keepdims=True)
P_test = np.argmax(Y_test,axis=1)
cost[(lr,mu)].append(cross_entropy(Y_test,Ttest))
current_cr = classification_rate(P_test,Ytest)
cr[(lr,mu)].append(current_cr)
if current_cr > best_cr:
best_cr = current_cr
best_lr = lr
best_mu = mu
dw = mu*dw - (1-mu)*learning_rate*derivative_w(Xtrain,Y_train,Ttrain)
db = mu*db - (1-mu)*learning_rate*derivative_b(Y_train,Ttrain)
if i == 0:
w = w0 + dw
b = b0 + db
else:
w += dw
b += db
if i % 100 == 0:
print('Learning Rate: ',lr,'Momentum: ',mu,'Cost: ',cost[(lr,mu)][i],'Classification Rate: ',cr[(lr,mu)][i])
if i == (iterations - 1):
print('')
return cost,cr,best_lr,best_mu,best_cr
def batch(learning_rate):
X, Y = transform_data()
X, Y = shuffle(X, Y)
N = len(X) // 2
Xtrain = X[:N]
Ytrain = Y[:N]
Ttrain = generate_T(Ytrain)
Xtest = X[N:]
Ytest = Y[N:]
Ttest = generate_T(Ytest)
N, D = Xtrain.shape
M = 100
K = len(set(Y))
iterations = 50
batch_N = 250
batches = len(X) // batch_N
v = np.random.randn(M, K) / np.sqrt(M + K)
b_1 = np.random.randn(K) / np.sqrt(K)
w = np.random.randn(D, M) / np.sqrt(D + M)
b_0 = np.random.randn(M) / np.sqrt(M)
batch_cost = []
batch_cr = []
best_batch = 0
best_iteration = 0
for i in range(iterations):
for b in range(batches):
X = Xtrain[b * batch_N:(b + 1) * batch_N, :]
T = Ttrain[b * batch_N:(b + 1) * batch_N, :]
Y, Z = generate_Y('tanh', X, w, b_0, v, b_1)
Y_test, _ = generate_Y('tanh', Xtest, w, b_0, v, b_1)
P_test = np.argmax(Y_test, axis=1)
if b % batches == 0:
batch_cost.append(cross_entropy(Y_test, Ttest))
cr = classification_rate(P_test, Ytest)
batch_cr.append(cr)
if cr > best_batch:
best_batch = cr
best_iteration = i
v -= learning_rate * derivative_v('tanh', Z, Y, T)
b_1 -= learning_rate * derivative_b1('tanh', Y, T)
w -= learning_rate * derivative_w('tanh', X, Y, Z, T, v)
b_0 -= learning_rate * derivative_b0('tanh', Y, Z, T, v)
if i % 100 == 0:
print('Batch Cost: ', batch_cost[i], 'Batch Classification: ',
batch_cr[i])
return batch_cost, batch_cr, best_batch, best_iteration
def rmsprop(learning_rate):
X, Y = transform_data()
X, Y = shuffle(X, Y)
N = len(X) // 2
Xtrain = X[:N]
Ytrain = Y[:N]
Ttrain = generate_T(Ytrain)
Xtest = X[N:]
Ytest = Y[N:]
Ttest = generate_T(Ytest)
Ttest = generate_T(Ytest)
N, D = Xtrain.shape
M = 100
K = len(set(Y))
iterations = 50
batch_N = 250
batches = N // batch_N
dv = 0
d_b1 = 0
dw = 0
d_b0 = 0
mu = .9
v = np.random.randn(M, K) / np.sqrt(M + K)
b_1 = np.random.randn(K) / np.sqrt(K)
w = np.random.randn(D, M) / np.sqrt(D + M)
b_0 = np.random.randn(M) / np.sqrt(M)
cache_v = np.ones((M, K))
cache_b1 = np.ones(K)
cache_w = np.ones((D, M))
cache_b0 = np.ones(M)
epsilon = 10e-10
decay = .9
rmsprop_cost = []
rmsprop_cr = []
best_rms = 0
best_iteration = 0
for i in range(iterations):
for b in range(batches):
X = Xtrain[b * batches:(b + 1) * batches, :]
T = Ttrain[b * batches:(b + 1) * batches, :]
Y, Z = generate_Y('tanh', X, w, b_0, v, b_1)
Y_test, _ = generate_Y('tanh', Xtest, w, b_0, v, b_1)
P_test = np.argmax(Y_test, axis=1)
if b % batches == 0:
rmsprop_cost.append(cross_entropy(Y_test, Ttest))
cr = classification_rate(P_test, Ytest)
rmsprop_cr.append(cr)
if cr > best_rms:
best_rms = cr
best_iteration = i
cache_v = decay * cache_v + (1 - decay) * derivative_v(
'tanh', Z, Y, T)**2
cache_b1 = decay * cache_b1 + (1 - decay) * derivative_b1(
'tanh', Y, T)**2
cache_w = decay * cache_w + (1 - decay) * derivative_w(
'tanh', X, Y, Z, T, v)**2
cache_b0 = decay * cache_b0 + (1 - decay) * derivative_b0(
'tanh', Y, Z, T, v)**2
dv = mu * dv - learning_rate * derivative_v(
'tanh', Z, Y, T) / (np.sqrt(cache_v + epsilon))
d_b1 = mu * d_b1 - learning_rate * derivative_b1(
'tanh', Y, T) / (np.sqrt(cache_b1 + epsilon))
dw = mu * dw - learning_rate * derivative_w(
'tanh', X, Y, Z, T, v) / (np.sqrt(cache_w + epsilon))
d_b0 = mu * d_b0 - learning_rate * derivative_b0(
'tanh', Y, Z, T, v) / (np.sqrt(cache_b0 + epsilon))
v += dv
b_1 += d_b1
w += dw
b_0 += d_b0
if i % 10 == 0:
print('RMSProp Cost: ', rmsprop_cost[i],
'RMSProp Classification: ', rmsprop_cr[i])
return rmsprop_cost, rmsprop_cr, best_rms, best_iteration
def full(self):
for i in range(self.iterations):
Y_train, Z = generate_Y(self.activation, self.Xtrain, self.w,
self.b_0, self.v, self.b_1)
P_train = np.argmax(Y_train, axis=1)
Y_test, _ = generate_Y(self.activation, self.Xtest, self.w,
self.b_0, self.v, self.b_1)
P_test = np.argmax(Y_test, axis=1)
self.train_cost.append(cross_entropy(Y_train, self.Ttrain))
self.test_cost.append(cross_entropy(Y_test, self.Ttest))
train_cr = classification_rate(P_train, self.Ytrain)
self.train_cr.append(train_cr)
test_cr = classification_rate(P_test, self.Ytest)
self.test_cr.append(test_cr)
if train_cr > self.best_train:
self.best_train = train_cr
self.train_iteration = i
if test_cr > self.best_test:
self.best_test = test_cr
self.test_iteration = i
self.m_v = self.decay_0 * self.m_v + (
1 - self.decay_0) * derivative_v(self.activation, Z, Y_train,
self.Ttrain)
self.dm_v = self.m_v / (1 - self.decay_0**(i + 1))
self.v_v = self.decay_1 * self.v_v + (
1 - self.decay_1) * derivative_v(self.activation, Z, Y_train,
self.Ttrain)**2
self.dv_v = self.v_v / (1 - self.decay_1**(i + 1))
self.m_b1 = self.decay_0 * self.m_b1 + (
1 - self.decay_0) * derivative_b1(self.activation, Y_train,
self.Ttrain)
self.dm_b1 = self.m_b1 / (1 - self.decay_0**(i + 1))
self.v_b1 = self.decay_1 * self.v_b1 + (
1 - self.decay_1) * derivative_b1(self.activation, Y_train,
self.Ttrain)**2
self.dv_b1 = self.v_b1 / (1 - self.decay_1**(i + 1))
self.m_w = self.decay_0 * self.m_w + (
1 - self.decay_0) * derivative_w(self.activation, self.Xtrain,
Y_train, Z, self.Ttrain,
self.v)
self.dm_w = self.m_w / (1 - self.decay_0**(i + 1))
self.v_w = self.decay_1 * self.v_w + (
1 - self.decay_1) * derivative_w(self.activation, self.Xtrain,
Y_train, Z, self.Ttrain,
self.v)**2
self.dv_w = self.v_w / (1 - self.decay_1**(i + 1))
self.m_b0 = self.decay_0 * self.m_b0 + (
1 - self.decay_0) * derivative_b0(self.activation, Y_train, Z,
self.Ttrain, self.v)
self.dm_b0 = self.m_b0 / (1 - self.decay_0**(i + 1))
self.v_b0 = self.decay_1 * self.v_b0 + (
1 - self.decay_1) * derivative_b0(self.activation, Y_train, Z,
self.Ttrain, self.v)**2
self.dv_b0 = self.v_b0 / (1 - self.decay_1**(i + 1))
self.v -= self.learning_rate * self.dm_v / (np.sqrt(self.dv_v +
self.epsilon))
self.b_1 -= self.learning_rate * self.dm_b1 / (
np.sqrt(self.dv_b1 + self.epsilon))
self.w -= self.learning_rate * self.dm_w / (np.sqrt(self.dv_w +
self.epsilon))
self.b_0 -= self.learning_rate * self.dm_b0 / (
np.sqrt(self.dv_b0 + self.epsilon))
if i % 100 == 0:
print(i, 'Train Cost: ', self.train_cost[i],
'Train Classification Rate: ', self.train_cr[i])
def stochastic(self, samples):
for i in range(self.iterations):
current_X, current_T = shuffle(self.Xtrain, self.Ttrain)
for s in range(samples):
X = current_X[s, :].reshape(1, current_X.shape[1])
T = current_T[s, :].reshape(1, current_T.shape[1])
Y, Z = generate_Y(self.activation, X, self.w, self.b_0, self.v,
self.b_1)
Y_train, _ = generate_Y(self.activation, self.Xtrain, self.w,
self.b_0, self.v, self.b_1)
P_train = np.argmax(Y_train, axis=1)
Y_test, _ = generate_Y(self.activation, self.Xtest, self.w,
self.b_0, self.v, self.b_1)
P_test = np.argmax(Y_test, axis=1)
self.train_cost.append(cross_entropy(Y_train, self.Ttrain))
self.test_cost.append(cross_entropy(Y_test, self.Ttest))
train_cr = classification_rate(P_train, self.Ytrain)
self.train_cr.append(train_cr)
test_cr = classification_rate(P_test, self.Ytest)
self.test_cr.append(test_cr)
if train_cr > self.best_train:
self.best_train = train_cr
self.train_iteration = i
if test_cr > self.best_test:
self.best_test = test_cr
self.test_iteration = i
self.m_v = self.decay_0 * self.m_v + (
1 - self.decay_0) * derivative_v(self.activation, Z, Y, T)
self.dm_v = self.m_v / (1 - self.decay_0**(i + 1))
self.v_v = self.decay_1 * self.v_v + (
1 - self.decay_1) * derivative_v(self.activation, Z, Y,
T)**2
self.dv_v = self.v_v / (1 - self.decay_1**(i + 1))
self.m_b1 = self.decay_0 * self.m_b1 + (
1 - self.decay_0) * derivative_b1(self.activation, Y, T)
self.dm_b1 = self.m_b1 / (1 - self.decay_0**(i + 1))
self.v_b1 = self.decay_1 * self.v_b1 + (
1 - self.decay_1) * derivative_b1(self.activation, Y, T)**2
self.dv_b1 = self.v_b1 / (1 - self.decay_1**(i + 1))
self.m_w = self.decay_0 * self.m_w + (
1 - self.decay_0) * derivative_w(self.activation, X, Y, Z,
T, self.v)
self.dm_w = self.m_w / (1 - self.decay_0**(i + 1))
self.v_w = self.decay_1 * self.v_w + (
1 - self.decay_1) * derivative_w(self.activation, X, Y, Z,
T, self.v)**2
self.dv_w = self.v_w / (1 - self.decay_1**(i + 1))
self.m_b0 = self.decay_0 * self.m_b0 + (
1 - self.decay_0) * derivative_b0(self.activation, Y, Z, T,
self.v)
self.dm_b0 = self.m_b0 / (1 - self.decay_0**(i + 1))
self.v_b0 = self.decay_1 * self.v_b0 + (
1 - self.decay_1) * derivative_b0(self.activation, Y, Z, T,
self.v)**2
self.dv_b0 = self.v_b0 / (1 - self.decay_1**(i + 1))
self.v -= self.learning_rate * self.dm_v / (
np.sqrt(self.dv_v + self.epsilon))
self.b_1 -= self.learning_rate * self.dm_b1 / (
np.sqrt(self.dv_b1 + self.epsilon))
self.w -= self.learning_rate * self.dm_w / (
np.sqrt(self.dv_w + self.epsilon))
self.b_0 -= self.learning_rate * self.dm_b0 / (
np.sqrt(self.dv_b0 + self.epsilon))
if i % 100 == 0:
print(i, 'Train Cost: ', self.train_cost[i],
'Train Classification Rate: ', self.train_cr[i])
