def train_model(data_path, model_path, norm_path, test_size=0.05, shuffle=True, lr=0.003, minibatch_size=2048,
epochs=30, lambd=0.001, h1_nodes=20, h2_nodes=20, h3_nodes=20,
testing=True, loading=False):
"""
Description
---
Trains a normalized (min-max) three-layer feedforward neural network model, given the data from data_path.
Model will be saved to model_path. Advanced settings are set above.
Inputs
---
data_path: Path for the process data. First column should be labels
model_path: Path for the model saving.
norm_path: Path for the normalization object.
test_size: Size of testing data set
shuffle: Boolean, shuffle the data for training? Breaks time correlation of data
lr: Learning rate of the model, higher learning rate results in faster, more unstable learning.
minibatch_size: Size of batches for stochastic / minibatch gradient descent
epochs: Number of passes through the whole data
lambd: Regularization term
h1_nodes: Amount of neurons in 1st hidden layer
h2_nodes: Amount of neurons in 2nd hidden layer
h3_nodes: Amount of neurons in 3rd hidden layer
testing: Training or testing?
loading: True if you want to load an old model for further training
Returns
---
raw_data: Data used for model building
heading_names: Headings of the raw data
linear_reg: Linear regression object
"""
# Load data
raw_data = pd.read_csv(data_path)
# Get heading tags / names, then transform into NumPy array
heading_names = list(raw_data)
raw_data = raw_data.values
print('There are {} feature(s) and {} label(s) with {} examples.'.format(raw_data.shape[1] - 1, 1,
raw_data.shape[0]))
# Train / test split
train_x, test_x, train_y, test_y = train_test_split(raw_data[:, 1:], raw_data[:, 0], test_size=test_size,
shuffle=shuffle, random_state=42)
train_x = train_x.reshape(-1, raw_data.shape[1] - 1)
test_x = test_x.reshape(-1, raw_data.shape[1] - 1)
train_y = train_y.reshape(-1, 1)
test_y = test_y.reshape(-1, 1)
# Normalization
if testing:
min_max_normalization = load(norm_path)
else:
min_max_normalization = MinMaxNormalization(np.concatenate([train_y, train_x], axis=1))
training_data = min_max_normalization(np.concatenate([train_y, train_x], axis=1))
testing_data = min_max_normalization(np.concatenate([test_y, test_x], axis=1))
train_x = training_data[:, 1:].reshape(-1, raw_data.shape[1] - 1)
test_x = testing_data[:, 1:].reshape(-1, raw_data.shape[1] - 1)
train_y = training_data[:, 0].reshape(-1, 1)
test_y = testing_data[:, 0].reshape(-1, 1)
# Test cases for NaN cases
assert(not np.isnan(train_x).any())
assert(not np.isnan(test_x).any())
assert(not np.isnan(train_y).any())
assert(not np.isnan(test_y).any())
with tf.Session() as sess:
# Build the neural network object
nn = NeuralNetwork(sess, train_x, train_y, test_x, test_y, lr=lr, minibatch_size=minibatch_size,
train_size=1 - test_size, h1_nodes=h1_nodes, h2_nodes=h2_nodes, h3_nodes=h3_nodes,
epochs=epochs, lambd=lambd)
# If testing, just run the model
if testing:
# Restore model
nn.saver.restore(sess, model_path)
# Predict based on model
pred = nn.test(features=test_x, training=not testing)
# Unnormalize
pred = min_max_normalization.unnormalize_y(pred)
test_y = min_max_normalization.unnormalize_y(test_y)
# Evaluate loss
rmse, mae = nn.eval_loss(pred, test_y)
r2 = r_squared(pred, test_y)
print('Final Test Results: Test RMSE: {:2f} | Test MAE: {:2f} | R2: {:2f}'.format(rmse, mae, r2))
plt.plot(pred[:-4900], label='Predictions')
plt.plot(test_y[:-4900], label='Actual')
plt.xlabel('Time, t (min)')
plt.ylabel('Flow Rate, Q (bbl/h)')
plt.legend(loc='0', frameon=False)
# plt.savefig('08smallnn_valid.eps', format='eps', dpi=1000)
plt.show()
"""
Residual Analysis
"""
residuals = pred - test_y
plt.plot(residuals)
plt.ylim([-50, 50])
plt.show()
bad_residuals = test_y[0:-1] - test_y[1:]
plt.plot(bad_residuals)
plt.ylim([-50, 50])
plt.show()
else:
# If loading old model for continued training
if loading:
nn.saver.restore(sess, Model_path)
else:
# Initialize global variables
sess.run(nn.init)
for epoch in range(epochs):
for i in range(nn.total_batch_number):
# Minibatch gradient descent
batch_index = int(i * nn.total_batch_number)
minibatch_x = train_x[batch_index:batch_index + nn.minibatch_size, :]
minibatch_y = train_y[batch_index:batch_index + nn.minibatch_size, :]
# Optimize the machine learning model
nn.train(features=minibatch_x, labels=minibatch_y, training=not testing)
# Record loss
if i % 10 == 0:
_ = nn.loss_check(features=train_x, labels=train_y, training=not testing)
# Evaluate training and testing losses
if i % 150 == 0:
# Check for loss
current_loss = nn.loss_check(features=test_x, labels=test_y, training=not testing)
# Predict training loss
pred = nn.test(features=train_x, training=not testing)
# Unnormalize
pred = min_max_normalization.unnormalize_y(pred)
label_y = min_max_normalization.unnormalize_y(train_y)
# Evaluate training loss
train_rmse, _ = nn.eval_loss(pred, label_y)
# Predict test loss
pred = nn.test(features=test_x, training=not testing)
# Unnormalize
pred = min_max_normalization.unnormalize_y(pred)
label_y = min_max_normalization.unnormalize_y(test_y)
# Evaluate training loss
test_rmse, _ = nn.eval_loss(pred, label_y)
print('Epoch: {} | Loss: {:2f} | Train RMSE: {:2f} | Test RMSE: {:2f}'.format(epoch,
current_loss,
train_rmse,
test_rmse))
# Save the model
nn.saver.save(sess, model_path)
print("Model saved at: {}".format(model_path))
# Save normalizer
save(min_max_normalization, norm_path)
print("Normalization saved at: {}".format(norm_path))
# Final test
test_pred = nn.test(features=test_x, training=not testing)
# Unnormalize data
test_pred = min_max_normalization.unnormalize_y(test_pred)
actual_y = min_max_normalization.unnormalize_y(test_y)
test_rmse, test_mae = nn.eval_loss(test_pred, actual_y)
r2 = r_squared(test_pred, actual_y)
print('Final Test Results: Test RMSE: {:2f} | Test MAE: {:2f} | R2: {:2f}'.format(test_rmse, test_mae, r2))
return raw_data, heading_names, nn
def train_model(data_path,
model_path,
norm_path,
test_size=0.05,
shuffle=True,
lr=0.003,
minibatch_size=2048,
train_size=0.9,
epochs=30,
lambd=0.001,
testing=False,
loading=False,
num_of_const=10):
"""
Description
---
Trains a normalized (min-max) linear regression model, given the data from data_path. Model will be saved
to model_path. Advanced settings are set above.
Inputs
---
data_path: Path for the process data. First column should be labels
model_path: Path for the model saving.
norm_path: Path for the normalization object.
test_size: Size
shuffle: Boolean, shuffle the data for training? Breaks time correlation of data
lr: Learning rate of the model, higher learning rate results in faster, more unstable learning.
minibatch_size: Size of batches for stochastic / minibatch gradient descent
epochs: Number of passes through the whole data
lambd: Regularization term
testing: Training or testing?
loading: If you want to load an old model for further training
Returns
---
raw_data: Data used for model building
heading_names: Headings of the raw data
linear_reg: Linear regression object
weights_biases: Weights and biases of the model
"""
raw_data = pd.read_csv(data_path)
heading_names = list(raw_data)
raw_data = raw_data.values
print('There are {} feature(s) and {} label(s) with {} examples.'.format(
raw_data.shape[1] - 1, 1, raw_data.shape[0]))
# Train / Test split
train_x, test_x, train_y, test_y = train_test_split(raw_data[:, 1:],
raw_data[:, 0],
test_size=test_size,
shuffle=shuffle,
random_state=42)
# Reshape for TensorFlow
train_x = train_x.reshape(-1, raw_data.shape[1] - 1)
test_x = test_x.reshape(-1, raw_data.shape[1] - 1)
train_y = train_y.reshape(-1, 1)
test_y = test_y.reshape(-1, 1)
# Test cases for NaN values
assert (not np.isnan(train_x).any())
assert (not np.isnan(test_x).any())
assert (not np.isnan(train_y).any())
assert (not np.isnan(test_y).any())
with tf.Session() as sess:
# Build linear regression object
linear_reg = LinearRegression(sess,
train_x,
train_y,
test_x,
test_y,
lr=lr,
minibatch_size=minibatch_size,
train_size=train_size,
epochs=epochs,
lambd=lambd,
num_of_const=num_of_const)
# If testing, just run it
if testing:
# Restore model
linear_reg.saver.restore(sess, save_path=model_path)
# Pred testing values. First num_of_const variables are constrained, remaining are unconstrained
pred = linear_reg.test(test_x[:, :num_of_const],
test_x[:, linear_reg.nx_Cons:])
# Evaluate loss
rmse, mae = linear_reg.eval_loss(pred, test_y)
se = standard_error(pred, test_y)
r2 = r_squared(pred, test_y)
plt.plot(pred, label='Predicted')
plt.plot(test_y, label='Test Data')
plt.xlabel('Time')
plt.ylabel('Flow rate (bbl/h)')
plt.legend(loc='0', frameon=False)
plt.show()
print('Test RMSE: {:2f} | Test MAE: {:2f} | SE: {:2f} | R2: {:2f}'.
format(rmse, mae, se, r2))
cons_par, uncons_par, biases = linear_reg.weights_and_biases()
else:
# Load an old model for further training
if loading:
linear_reg.saver.restore(sess, Model_path)
else:
# Global variables initializer
sess.run(linear_reg.init)
for epoch in range(linear_reg.epochs):
for i in range(linear_reg.total_batch_number):
# Mini-batch gradient descent
batch_index = i * linear_reg.minibatch_size
minibatch_x = train_x[batch_index:batch_index +
linear_reg.minibatch_size, :]
minibatch_y = train_y[batch_index:batch_index +
linear_reg.minibatch_size, :]
# Optimize machine learning model
linear_reg.train(
const_features=minibatch_x[:, :num_of_const],
unconst_features=minibatch_x[:, linear_reg.nx_Cons:],
labels=minibatch_y)
# Record loss
if i % 10 == 0:
_ = linear_reg.loss_check(
const_features=train_x[:, :num_of_const],
unconst_features=train_x[:, linear_reg.nx_Cons:],
labels=train_y)
# Evaluate train and test losses
if i % 150 == 0:
current_loss = linear_reg.loss_check(
const_features=train_x[:, :num_of_const],
unconst_features=train_x[:, linear_reg.nx_Cons:],
labels=train_y)
train_pred = linear_reg.test(
const_features=train_x[:, :num_of_const],
unconst_features=train_x[:, linear_reg.nx_Cons:])
# Evaluate error
train_rmse, train_mae = linear_reg.eval_loss(
train_pred, train_y)
test_pred = linear_reg.test(
const_features=test_x[:, :num_of_const],
unconst_features=test_x[:, linear_reg.nx_Cons:])
test_rmse, test_mae = linear_reg.eval_loss(
test_pred, test_y)
print(
'Epoch: {} | Loss: {:2f} | Train RMSE: {:2f} | Test RMSE: {:2f}'
.format(epoch, current_loss, train_rmse,
test_rmse))
# Save model
linear_reg.saver.save(sess, model_path)
print("Model saved at: {}".format(model_path))
# Final test
test_pred = linear_reg.test(
const_features=test_x[:, :num_of_const],
unconst_features=test_x[:, linear_reg.nx_Cons:])
test_rmse, test_mae = linear_reg.eval_loss(test_pred, test_y)
se = standard_error(test_pred, test_y)
r2 = r_squared(test_pred, test_y)
plt.plot(test_pred[:], label='Predicted')
plt.plot(test_y[:], label='Test Data')
plt.xlabel('Time')
plt.ylabel('Flow rate (bbl/h)')
plt.legend(loc='0', frameon=False)
plt.show()
print(
'Final Test Results: Test RMSE: {:2f} | Test MAE: {:2f} | SE: {:2f} | R2: {:2f}'
.format(test_rmse, test_mae, se, r2))
cons_par, uncons_par, biases = linear_reg.weights_and_biases()
return raw_data, heading_names, linear_reg, cons_par, uncons_par, biases
def train_model(data_path, model_path, norm_path, test_size=0.05, shuffle=True, lr=0.003, minibatch_size=2048,
epochs=30, lambd=0.001, testing=False, loading=False, plot_start=1, plot_end=5000):
"""
Description
---
Trains a normalized (min-max) linear regression model, given the data from data_path. Model will be saved
to model_path. Advanced settings are set above.
Inputs
---
data_path: Path for the process data. First column should be labels
model_path: Path for the model saving.
norm_path: Path for the normalization object.
test_size: Size
shuffle: Boolean, shuffle the data for training? Breaks time correlation of data
lr: Learning rate of the model, higher learning rate results in faster, more unstable learning.
minibatch_size: Size of batches for stochastic / minibatch gradient descent
epochs: Number of passes through the whole data
lambd: Regularization term
testing: Training or testing?
loading: If you want to load an old model for further training
plot_start: Index for the start of the validation plot
plot_end: Index for the end of the validation plot
Returns
---
raw_data: Data used for model building
heading_names: Headings of the raw data
linear_reg: Linear regression object
weights_biases: Weights and biases of the model
"""
raw_data = pd.read_csv(data_path)
heading_names = list(raw_data)
raw_data = raw_data.values
print('There are {} feature(s) and {} label(s) with {} examples.'.format(raw_data.shape[1] - 1, 1,
raw_data.shape[0]))
# Train / Test split
train_x, test_x, train_y, test_y = train_test_split(raw_data[:, 1:], raw_data[:, 0],
test_size=test_size, shuffle=shuffle, random_state=42)
# Reshape for TensorFlow
train_x = train_x.reshape(-1, raw_data.shape[1] - 1)
test_x = test_x.reshape(-1, raw_data.shape[1] - 1)
train_y = train_y.reshape(-1, 1)
test_y = test_y.reshape(-1, 1)
# Normalization
if testing:
min_max_normalization = load(norm_path)
else:
min_max_normalization = MinMaxNormalization(np.concatenate([train_y, train_x], axis=1))
training_data = min_max_normalization(np.concatenate([train_y, train_x], axis=1))
testing_data = min_max_normalization(np.concatenate([test_y, test_x], axis=1))
# Reshape for TensorFlow
train_x = training_data[:, 1:].reshape(-1, raw_data.shape[1] - 1)
test_x = testing_data[:, 1:].reshape(-1, raw_data.shape[1] - 1)
train_y = training_data[:, 0].reshape(-1, 1)
test_y = testing_data[:, 0].reshape(-1, 1)
# Test cases for NaN values
assert(not np.isnan(train_x).any())
assert(not np.isnan(test_x).any())
assert(not np.isnan(train_y).any())
assert(not np.isnan(test_y).any())
with tf.Session() as sess:
# Build linear regression object
linear_reg = LinearRegression(sess, train_x, train_y, test_x, test_y, lr=lr, minibatch_size=minibatch_size,
train_size=(1 - test_size), epochs=epochs, lambd=lambd)
# If testing, just run it
if testing:
# Restore model
linear_reg.saver.restore(sess, save_path=model_path)
# Pred testing values
pred = linear_reg.test(test_x)
# Unnormalize
pred = min_max_normalization.unnormalize_y(pred)
test_y = min_max_normalization.unnormalize_y(test_y)
# Evaluate loss
rmse, mae = linear_reg.eval_loss(pred, test_y)
print('Test RMSE: {:2f} | Test MAE: {:2f}'.format(rmse, mae))
weights_biases = linear_reg.weights_and_biases()
# Non-scrambled data plot
seq_pred(session=sess, model=linear_reg.z, features=linear_reg.X, normalizer=min_max_normalization,
data=raw_data,
time_start=plot_start, time_end=plot_end,
adv_plot=False)
else:
# Load old model for further testing
if loading:
linear_reg.saver.restore(sess, Model_path)
else:
# Global variables initializer
sess.run(linear_reg.init)
for epoch in range(linear_reg.epochs):
for i in range(linear_reg.total_batch_number):
# Mini-batch gradient descent
batch_index = i * linear_reg.minibatch_size
minibatch_x = train_x[batch_index:batch_index + linear_reg.minibatch_size, :]
minibatch_y = train_y[batch_index:batch_index + linear_reg.minibatch_size, :]
# Optimize machine learning model
linear_reg.train(features=minibatch_x, labels=minibatch_y)
# Record loss
if i % 10 == 0:
_ = linear_reg.loss_check(features=train_x, labels=train_y)
# Evaluate train and test losses
if i % 150 == 0:
current_loss = linear_reg.loss_check(features=train_x, labels=train_y)
train_pred = linear_reg.test(features=train_x)
# Unnormalize data
train_pred = min_max_normalization.unnormalize_y(train_pred)
actual_y = min_max_normalization.unnormalize_y(train_y)
# Evaluate error
train_rmse, train_mae = linear_reg.eval_loss(train_pred, actual_y)
test_pred = linear_reg.test(features=test_x)
# Unnormalize data
test_pred = min_max_normalization.unnormalize_y(test_pred)
actual_y = min_max_normalization.unnormalize_y(test_y)
test_rmse, test_mae = linear_reg.eval_loss(test_pred, actual_y)
print('Epoch: {} | Loss: {:2f} | Train RMSE: {:2f} | Test RMSE: {:2f}'.format(epoch,
current_loss,
train_rmse,
test_rmse))
# Save model
linear_reg.saver.save(sess, model_path)
print("Model saved at: {}".format(model_path))
# Save normalizer
save(min_max_normalization, norm_path)
print("Normalization saved at: {}".format(norm_path))
# Final test
test_pred = linear_reg.test(features=test_x)
# Unnormalize data
test_pred = min_max_normalization.unnormalize_y(test_pred)
actual_y = min_max_normalization.unnormalize_y(test_y)
test_rmse, test_mae = linear_reg.eval_loss(test_pred, actual_y)
R2 = r_squared(test_pred, actual_y)
print('Final Test Results: Test RMSE: {:2f} | Test MAE: {:2f} | R2: {:2f}'.format(test_rmse, test_mae, R2))
weights_biases = linear_reg.weights_and_biases()
# Non-scrambled data plot
seq_pred(session=sess, model=linear_reg.z, features=linear_reg.X, normalizer=min_max_normalization,
data=raw_data,
time_start=plot_start, time_end=plot_end,
adv_plot=False)
return raw_data, heading_names, linear_reg, weights_biases
