def seq_mlp_mnist():
"""test MLP with MNIST data and Sequential
"""
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('./tmp/data', one_hot=True)
training_data = np.array([image.flatten() for image in mnist.train.images])
training_label = mnist.train.labels
valid_data = np.array(
[image.flatten() for image in mnist.validation.images])
valid_label = mnist.validation.labels
input_dim = training_data.shape[1]
label_size = training_label.shape[1]
model = Sequential()
model.add(Input(input_shape=(input_dim, )))
model.add(Dense(300, activation='relu'))
model.add(Dropout(0.5))
model.add(Softmax(label_size))
model.compile('CE', optimizer=SGD())
model.fit(training_data,
training_label,
validation_data=(valid_data, valid_label),
metric='Accuracy',
peek_type='single-cls')
def model_mlp_random_cls():
"""test MLP with random data and Model
"""
input_size = 600
input_dim = 20
label_size = 2
train_X = np.random.random((input_size, input_dim))
# train_y = np.zeros((input_size, label_size))
train_y = np.random.randint(0, label_size, (input_size, 1))
# for _ in range(input_size):
# train_y[_, np.random.randint(0, label_size)] = 1
Inputs = Input(input_shape=input_dim)
X = Dense(100, activation='relu')(Inputs)
X = Softmax(label_size)(X)
model = Model(Inputs, X)
model.compile('CE')
model.fit(train_X,
train_y,
batch_size=256,
verbose=20,
epochs=100,
metric='Accuracy',
peek_type='single-cls')
def svm():
print(10 * '#' + ' SGD SVM model ' + 10 * '#')
# build the linear model with gradient descent
# define layer
Inputs = Input(input_shape=X_train.shape[1])
# X = Linear(output_dim=64,
# regularizer=L2_Regularizer(1e-5),
# # regularizer=L1_Regularizer(1e-2),
# # regularizer=L1L2_Regularizer(l2=1),
# activation='swish')(Inputs)
# X = Linear(output_dim=128,
# regularizer=L2_Regularizer(1e-5),
# # regularizer=L1_Regularizer(1e-2),
# # regularizer=L1L2_Regularizer(l2=1),
# activation='swish')(X)
# X = Linear(output_dim=256,
# regularizer=L2_Regularizer(1e-5),
# # regularizer=L1_Regularizer(1e-2),
# # regularizer=L1L2_Regularizer(l2=1),
# activation='swish')(X)
X = Linear(
output_dim=1,
# initializer='default_weight_initializer',
initializer=default_weight_initializer,
regularizer=L2_Regularizer(1e-5),
# regularizer=L1_Regularizer(1e-2),
# regularizer=L1L2_Regularizer(l2=1),
activation=None)(Inputs)
model = Model(Inputs, X)
# model.compile('HL', optimizer=SGD(lr=0.001))
model.compile('HL', optimizer=Adam(lr=0.001))
model.fit(X_train,
y_train,
verbose=10,
epochs=100,
validation_data=(X_test, y_test),
batch_size=64,
metric='svm_binary_accuracy',
shuffle=True,
peek_type='single-svm-cls')
plt.subplot(211)
plt.plot(model.train_losses, label='train_losses')
plt.plot(model.validation_losses, label='valid_losses')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.subplot(212)
plt.plot(model.train_metrics, label='train_accuracy')
plt.plot(model.validation_metrics, label='valid_accuracy')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.show()
print(10 * '#' + ' SGD SVM model end ' + 10 * '#')
print()
def seq_cnn_mnist():
"""test CNN with MNIST data and Sequential
"""
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets('/tmp/data', one_hot=False)
training_data = np.array(
[image.reshape(28, 28, 1) for image in mnist.train.images])
training_label = mnist.train.labels
valid_data = np.array(
[image.reshape(28, 28, 1) for image in mnist.validation.images])
valid_label = mnist.validation.labels
label_size = 10
model = Sequential()
model.add(Input(batch_input_shape=(None, 28, 28, 1)))
model.add(Conv2d(3, 16, stride=1, padding=2, activation='relu'))
# model.add(MaxPooling2D(4, stride=2))
model.add(AvgPooling2D(4, stride=2))
model.add(Conv2d(2, 32, stride=1, padding=0, activation='relu'))
# model.add(MaxPooling2D(3, stride=2))
model.add(AvgPooling2D(3, stride=2))
model.add(Conv2d(1, 64, stride=1, padding=0, activation='relu'))
# model.add(MaxPooling2D(3, stride=3))
# model.add(AvgPooling2D(3, stride=3))
model.add(Flatten())
model.add(Softmax(label_size))
model.compile('CE', optimizer=Adam(lr=1e-3))
# model.fit(training_data, training_label, validation_data=(valid_data, valid_label),
# batch_size=256, verbose=1, epochs=5, metric='Accuracy', peek_type='single-cls')
# model.fit(training_data[:1000], training_label[:1000], validation_data=(valid_data[:1000], valid_label[:1000]),
# batch_size=256, verbose=1, epochs=10, metric='Accuracy', peek_type='single-cls')
model.fit(training_data[:100],
training_label[:100],
validation_data=(valid_data[:50], valid_label[:50]),
batch_size=256,
verbose=10,
epochs=100,
metric='Accuracy',
peek_type='single-cls')
plt.subplot(211)
plt.plot(model.train_losses, label='train_losses')
plt.plot(model.validation_losses, label='valid_losses')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.subplot(212)
plt.plot(model.train_metrics, label='train_accuracy')
plt.plot(model.validation_metrics, label='valid_accuracy')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.show()
def dlr():
print(10 * '#' + ' Linear model ' + 10 * '#')
# build the linear model with gradient descent
# define layer
input_dim = train_x.shape[1]
Inputs = Input(input_shape=input_dim)
X = Linear(output_dim=1, activation=None,
regularizer=L2_Regularizer(1),
initializer=ones)(Inputs)
model = Model(Inputs, X)
# lr = 0.001 for grade point prediction, use MSE is a lot better than MAE
# model.compile('MSE', optimizer=SGD(lr=0.001))
# lr = 0.01 for score prediction, use MAE is slightly better than MSE
model.compile('MAE', optimizer=SGD(lr=0.01))
# or we can use HB (huber loss) for both two
# but remember lr = 0.001 for grade point prediction
# and lr = 0.01 for score prediction
# model.compile('HB', optimizer=SGD(lr=0.01))
model.fit(train_x, train_y,
verbose=500, epochs=10000,
validation_data=(test_x, test_y),
batch_size=16, metric='mae',
peek_type='single-reg')
plt.subplot(211)
plt.plot(model.train_losses, label='train_losses')
plt.plot(model.validation_losses, label='valid_losses')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.subplot(212)
plt.plot(model.train_metrics, label='train_MAE')
plt.plot(model.validation_metrics, label='valid_MAE')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.show()
train_y_hat = model.forward(train_x)
test_y_hat = model.forward(test_x)
training_error = absolute_error(train_y, train_y_hat) / train_y.shape[0]
test_error = absolute_error(test_y, test_y_hat) / test_y.shape[0]
print('Training error: ', training_error)
print('Test error: ', test_error)
print(10 * '#' + ' Linear model end ' + 10 * '#')
print()
return X.params
def dlr():
print(10 * '#' + ' SGD Factorization model ' + 10 * '#')
# build the linear model with gradient descent
# define layer
X_train, y_train, X_val, y_val, X_test, y_test, num_user, num_item = read_data()
Inputs = Input(input_shape=2)
out = Factorization(a_dim=num_user, b_dim=num_item, k=10,
use_bias=False,
regularizer=L2_Regularizer(0.1))(Inputs)
# out = Factorization(a_dim=num_user, b_dim=num_item, k=10)(Inputs)
model = Model(Inputs, out)
# model.compile('MSE', optimizer=Adam(lr=0.001))
model.compile(HuberLoss(), optimizer=Adam(lr=0.001))
model.fit(X_train, y_train,
verbose=10, epochs=300,
validation_data=(X_val, y_val),
batch_size=256, metric='MAE',
shuffle=True,
peek_type='single-reg')
plt.plot(model.train_losses, label='$loss_{train}$')
plt.plot(model.validation_losses, label='$loss_{val}$')
plt.legend()
plt.savefig('./loss.png', dpi=300)
plt.show()
plt.plot(model.train_metrics, label='$MAE_{train}$')
plt.plot(model.validation_metrics, label='$MAE_{val}$')
plt.legend()
plt.savefig('./metric.png', dpi=300)
plt.show()
train_y_hat = model.forward(X_train)
val_y_hat = model.forward(X_val)
test_y_hat = model.forward(X_test)
training_error = absolute_error(y_train, train_y_hat) / y_train.shape[0]
val_error = absolute_error(y_val, val_y_hat) / y_val.shape[0]
test_error = absolute_error(y_test, test_y_hat) / y_test.shape[0]
print(model.best_performance(bigger=False))
print('Training error: ', training_error)
print('Val error: ', val_error)
print('Test error: ', test_error)
print(10 * '#' + ' SGD Factorization model end ' + 10 * '#')
print()
def dlr():
print(10 * '#' + ' SGD Linear model ' + 10 * '#')
# build the linear model with gradient descent
# define layer
Inputs = Input(input_shape=X_train.shape[1])
linear_out = Linear(output_dim=1, activation=None,
initializer=ones)(Inputs)
model = Model(Inputs, linear_out)
model.compile('MSE', optimizer=SGD(lr=0.01))
model.fit(
X_train,
y_train,
verbose=-1,
epochs=5000,
validation_data=(X_test, y_test),
batch_size=256,
metric='MAE',
shuffle=True,
# peek_type='single-reg'
)
plt.subplot(211)
plt.plot(model.train_losses, label='train_losses')
plt.plot(model.validation_losses, label='valid_losses')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.subplot(212)
plt.plot(model.train_metrics, label='train_metrics')
plt.plot(model.validation_metrics, label='valid_metrics')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.show()
train_y_hat = model.forward(X_train)
test_y_hat = model.forward(X_test)
training_error = mean_absolute_error(y_train,
train_y_hat) / y_train.shape[0]
test_error = mean_absolute_error(y_test, test_y_hat) / y_test.shape[0]
print('Training error: ', training_error)
print('Test error: ', test_error)
print(10 * '#' + ' SGD Linear model end ' + 10 * '#')
print()
def seq_cnn_face():
X_train, y_train, X_val, y_val, X_test, y_test = read_data(size=28)
print('train set positive class portion: %.2f (%d / %d)' %
(np.mean(y_train), int(np.sum(y_train)), y_train.shape[0]))
print('val set positive class portion: %.2f (%d / %d)' %
(np.mean(y_val), int(np.sum(y_val)), y_val.shape[0]))
print('test set positive class portion: %.2f (%d / %d)' %
(np.mean(y_test), int(np.sum(y_test)), y_test.shape[0]))
model = Sequential()
model.add(Input(batch_input_shape=(None, *X_train.shape[1:])))
model.add(Conv2d(3, 16, stride=1, padding=2, activation='swish'))
# model.add(MaxPooling2D(4, stride=2))
# model.add(AvgPooling2D(4, stride=2))
model.add(Conv2d(2, 32, stride=1, padding=0, activation='swish'))
# model.add(MaxPooling2D(3, stride=2))
# model.add(AvgPooling2D(3, stride=2))
model.add(Conv2d(1, 64, stride=1, padding=0, activation='swish'))
model.add(Flatten())
model.add(Softmax(2))
model.compile('CE', optimizer=Adam(lr=1e-3))
model.fit(X_train,
y_train,
validation_data=(X_val, y_val),
batch_size=256,
verbose=1,
epochs=100,
shuffle=True,
metric='Accuracy',
peek_type='single-cls')
plt.subplot(211)
plt.plot(model.train_losses, label='train_losses')
plt.plot(model.validation_losses, label='valid_losses')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.subplot(212)
plt.plot(model.train_metrics, label='train_accuracy')
plt.plot(model.validation_metrics, label='valid_accuracy')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.show()
def model_mlp_random_reg():
"""test MLP with random data and Sequential
"""
input_size = 600
input_dim = 20
output_dim = 1
train_X = np.random.random((input_size, input_dim))
random_weight = np.random.random((input_dim, output_dim))
random_noise = np.random.random((input_size, output_dim))
train_y = np.dot(train_X, random_weight) + random_noise
Inputs = Input(input_shape=input_dim)
X = Dense(100, activation='relu')(Inputs)
X = Dense(100, activation='relu')(X)
X = Dense(output_dim, activation=None)(X)
model = Model(Inputs, X)
model.compile('MSE', optimizer='momentum')
model.fit(
train_X,
train_y,
verbose=100,
epochs=600,
batch_size=256,
# validation_split=0.1,
metric='MAE',
peek_type='single-reg')
print(len(model.train_losses))
print(len(model.validation_losses))
print(len(model.train_metrics))
print(len(model.validation_metrics))
plt.axis([0, len(model.train_losses), 0, 5])
plt.plot(model.train_losses)
plt.plot(model.validation_losses)
# plt.plot(model.train_metrics)
# plt.plot(model.validation_metrics)
plt.show()
def dmlr():
print(10 * '#' + ' SGD Deep Linear model ' + 10 * '#')
# build the linear model with gradient descent
# define layer
Inputs = Input(input_shape=X_train.shape[1])
linear_out = Linear(output_dim=64, activation='swish')(Inputs)
linear_out = Linear(output_dim=128, activation='swish')(linear_out)
linear_out = Linear(output_dim=256, activation='swish')(linear_out)
linear_out = Linear(output_dim=1, activation=None)(linear_out)
model = Model(Inputs, linear_out)
model.compile('MSE', optimizer=Momentum(lr=0.0001))
model.fit(
X_train,
y_train,
verbose=100,
epochs=500,
validation_data=(X_test, y_test),
batch_size=256,
metric='MAE',
shuffle=True,
# peek_type='single-reg'
)
# y_pred = model.forward(X_test)
# for yp, yt in zip(y_pred, y_test):
# print(yp, yt)
plt.subplot(211)
plt.plot(model.train_losses, label='train_losses')
plt.plot(model.validation_losses, label='valid_losses')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.subplot(212)
plt.plot(model.train_metrics, label='train_metrics')
plt.plot(model.validation_metrics, label='valid_metrics')
plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left', borderaxespad=0.)
plt.show()
print(10 * '#' + ' SGD Deep Linear model end ' + 10 * '#')
print()
def seq_mlp_random_cls():
"""test MLP with random data and Sequential
"""
input_size = 600
input_dim = 20
label_size = 10
train_X = np.random.random((input_size, input_dim))
train_y = np.zeros((input_size, label_size))
for _ in range(input_size):
train_y[_, np.random.randint(0, label_size)] = 1
model = Sequential()
model.add(Input(input_shape=input_dim))
model.add(Dense(100, activation='relu'))
model.add(Softmax(label_size))
model.compile('CE')
model.fit(train_X,
train_y,
verbose=100,
epochs=5000,
validation_split=0.2,
metric='Accuracy',
peek_type='single-cls')
