def support_vector(x_train, x_test, y_train, y_test):
'''
support vector try
'''
x_train, x_test, y_train, y_test, scalar_x, scalar_y = scale_data(
x_train, x_test, y_train, y_test)
kernels = ['linear', 'poly', 'rbf', 'sigmoid']
results_r2 = []
results_r2_train = []
for kernel in kernels:
regressor_temp = SVR(kernel=kernel)
regressor_temp.fit(x_train, y_train.flatten())
preds = regressor_temp.predict(x_test)
r_squared = r2_score(y_test, preds)
results_r2.append(r_squared)
preds, actual = inverse_transform(preds, y_test, scalar_y)
preds = regressor_temp.predict(x_train)
results_r2_train.append(r2_score(y_train, preds))
kernels = np.array(kernels)
results_r2 = np.array(results_r2)
table_of_results = np.concatenate((kernels.reshape(
len(kernels), 1), results_r2.reshape(len(results_r2), 1)),
axis=1)
print(table_of_results)
def train(layers, loss_function, show_preds=False, scale_input=True):
'''
method for training FuriosaNet
Parameters
==========
`layers`:
list of specified number of neurons per layer
'''
# generate data with scaled input
x_train, x_test, y_train, y_test, _, scalar_y, dataset, _ = generate_data(
"dbs/data_2010s.csv", scale_input=scale_input)
# x_train, x_test, y_train, y_test, _, scalar_y, dataset, _ = generate_data(
# "dbs/data_2010s.csv", drop_features=['title', 'tmdb_id', 'year', 'view_count', 'like_count', 'dislike_count', 'comment_count'])
# define normalization string for specifying saved filed
normalization = ""
if scale_input:
normalization = "-norm"
# define the layers of the model, excluding the output layer
layers = [x_train.shape[1]] + layers
# call the function that will build the model
model = build_model(layers, loss_function=loss_function)
# define what weights we want to save and how we want to save them
callback = ModelCheckpoint(
filepath=
f"weights/nn{normalization}-{loss_function}-{stringify_model(layers)}-1-weights.h5",
verbose=1,
save_best_only=True,
monitor="val_" + loss_function,
save_weights_only=True)
# train the network
results = model.fit(x_train,
y_train,
batch_size=50,
epochs=100,
validation_data=(x_test, y_test),
callbacks=[callback])
# plot the history of the model based on the loss_function
plot_history(results, layers, loss_function, norm=normalization)
if show_preds:
# get the rescaled predictions
predictions, actual_values = inverse_transform(model.predict(x_test),
y_test, scalar_y)
# plot the predictions and get the r-squared value of the model
r_squared = r2_score(predictions, actual_values)
plot_predictions(predictions,
actual_values,
r_squared,
layers=layers,
norm=normalization)
def test(weights_file,
layers,
data_file="dbs/data_2010s.csv",
create_fig=True,
show_fig=True,
scale_input=True):
'''
test a model with specified `layers` architecture using the `weights_file`
Parameters
==========
`weights_file`:
path to the .h5 file containing the pretrained weights
`layers`:
list of integer values specifying the number of nodes at the each hidden layer
except the final layer
Keyword Args:
==========
`show_fig`:
default True; display graph of actual values vs predictions
Returns
==========
(predictions, actual_values, dataset, test_indices)
'''
_, x_test, _, y_test, _, scalar_y, dataset, test_indices = generate_data(
data_file, scale_input=scale_input)
# _, x_test, _, y_test, _, scalar_y, dataset, _ = generate_data(
# "dbs/data_2010s.csv", drop_features=['title', 'tmdb_id', 'year', 'view_count', 'like_count', 'dislike_count', 'comment_count'])
model = build_model(layers)
model.load_weights(weights_file)
predictions, actual_values = inverse_transform(model.predict(x_test),
y_test, scalar_y)
r_squared = r2_score(predictions, actual_values)
if create_fig:
plot_predictions(predictions,
actual_values,
r_squared,
layers=layers,
best="best-",
save_fig=True,
show_fig=show_fig)
df = create_df(predictions, dataset, test_indices)
create_interactive_plot(df, model=layers)
return predictions, actual_values, dataset, test_indices
def overview_test(weights_file):
_, x_test, _, y_test, _, scalar_y, dataset, test_indices = get_trailer_data(
"dbs/data_2010s.csv", scale_input=True)
layers = [
int(val) for val in weights_file.split("/")[-1].split("-")
if val.isdigit()
][:-1]
model = build_model(layers)
model.load_weights(weights_file)
predictions, actual_values = inverse_transform(model.predict(x_test),
y_test, scalar_y)
r_squared = r2_score(predictions, actual_values)
print(r_squared)
# plot_predictions(
# predictions, actual_values, r_squared, layers=layers, best="best-", save_fig=False, show_fig=show_fig)
return predictions, actual_values, dataset, test_indices
def evaluate(metric, weights_folder, save_table=True, create_fig=False):
'''
method that will compare all models from a specified `weights_folder`
based on `metric`
Parameters
==========
`metric`:
metric to evaluate models
`weights_folder`:
path to fodler containing the weights of model which are meant to be modeled.
'''
metric_functions = {
"r-squared": r2_score,
}
predictions = None
_, x_test, _, y_test, _, scalar_y, dataset, _ = generate_data(
"dbs/data_2010s.csv", scale_input=True)
models = dict()
weights_files = [weights for weights in os.listdir(weights_folder)]
for weights_file in weights_files:
layers = get_layers_from_file(weights_file)
model = build_model(layers)
model.load_weights(os.path.join(weights_folder, weights_file))
predictions, actual_values = inverse_transform(model.predict(x_test),
y_test, scalar_y)
r2 = metric_functions[metric](actual_values, predictions)
if create_fig:
# TODO: plot get some plottable points to be layers on one graph
print()
models[weights_file] = r2
# TODO: show points plotted by each model maybe?
if create_fig:
print()
if save_table:
pd.DataFrame.from_dict(models, orient='index',
columns=["r-squared"
]).to_csv("model-evaluation.csv")
else:
print(f"{'Model':^40s}| {metric}")
for k in models:
print(f"{k:^40s}| {models[k]:0.3f}")
return models
def overview_mlp(layers):
# generate data with scaled input
x_train, x_test, y_train, y_test, _, scalar_y, dataset, test_indices = get_trailer_data(
"dbs/data_2010s.csv", scale_input=True)
x_train, x_test = join_with_overview_bayes(x_train, x_test, dataset,
test_indices)
loss_function = "mean_squared_error"
# define the layers of the model, excluding the output layer
layers = [x_train.shape[1]] + layers
# call the function that will build the model
model = build_model(layers, loss_function=loss_function)
# define what weights we want to save and how we want to save them
callback = ModelCheckpoint(
filepath=
f"weights/prereleased-overview-{loss_function}-nn-{stringify_model(layers)}-1-weights.h5",
verbose=1,
save_best_only=True,
monitor="val_" + loss_function,
save_weights_only=True)
# train the network
results = model.fit(x_train,
y_train,
batch_size=50,
epochs=50,
validation_data=(x_test, y_test),
callbacks=[callback])
# plot the history of the model based on the loss_function
plot_history(results,
layers,
loss_function,
norm="-prerelease-overview",
show_fig=True)
# get the rescaled predictions
predictions, actual_values = inverse_transform(model.predict(x_test),
y_test, scalar_y)
r_squared = r2_score(predictions, actual_values)
print(r_squared)
def svr(save_fig=False, compare=False):
'''
main function for testing Support Vector Regression models
Keyword Args
===========
`save_fig`:
if True, save a matplotlib figure AND plotly interactive figure
of predictions
`compare`:
if true, compare values with trailer data to the baseline
'''
x_train, x_test, y_train, y_test, _, scalar_y, dataset, test_indices = generate_data(
"dbs/data_2010s.csv")
# get parameters for comparing to the "baseline"
if compare:
# line is too long
baseline_data = generate_data(
"dbs/data_2010s.csv",
drop_features=[
'title', 'tmdb_id', 'year', 'view_count',
'like_count', 'dislike_count', 'comment_count']
)
x_train_b, x_test_b, y_train_b, y_test_b, _, _, _, _ = baseline_data
# define kernels that will be compared
kernels = ['linear', 'poly', 'rbf', 'sigmoid']
#initialize lists that will be used to print results
results_r2 = []
baseline_results = []
for kernel in kernels:
# get preds and r_squared value per model
preds, r_squared = test(kernel, x_train, x_test, y_train, y_test)
# add it to the table
results_r2.append(r_squared)
if compare:
# get preds and r_squared value per model for the baseline data
preds, baseline_r2 = test(kernel, x_train_b, x_test_b, y_train_b, y_test_b)
# add it to the table
baseline_results.append(baseline_r2)
if save_fig:
# get the actual values of the revenue
# using the the inverse transform of the data using scalar_y
preds, actual = inverse_transform(preds, y_test, scalar_y)
plot_predictions(preds, actual, r_squared, model=f"SVR-{kernel}")
test_df = create_df(preds, dataset, test_indices)
create_interactive_plot(test_df, model=f"SVR-{kernel}")
# to compare without the outlier, uncomment the following line:
# remove_ghostbusters(df)
# create a table of results
kernels = np.array(kernels)
results_r2 = np.array(results_r2)
table_of_results = np.concatenate(
(kernels.reshape(len(kernels), 1), results_r2.reshape(len(results_r2), 1)),
axis=1)
print(table_of_results)
if compare:
baseline_results = np.array(baseline_results)
table_of_results = np.concatenate(
(kernels.reshape(len(kernels), 1), baseline_results.reshape(len(baseline_results), 1)),
axis=1)
print(table_of_results)