def Lasso(self, con, remove_zeros=True, alpha=1.0):
mat_builder, vec = self.Matrix(con)
print len(mat_builder.entries)
return lasso.Lasso(mat_builder,
vec,
alpha=alpha,
remove_zeros=remove_zeros)
def onpress(self, event):
#print event
if event.button != 3: return
if self.fig.canvas.widgetlock.locked(): return
if event.inaxes is None: return
#print "Get Lasso"
self.lasso = lasso.Lasso(event.inaxes, (event.xdata, event.ydata), self.select_mlt)
# acquire a lock on the widget drawing
self.fig.canvas.widgetlock(self.lasso)
import lasso
import numpy as np
import csv
Xy = []
with open("winequality-red.csv") as fp:
for row in csv.reader(fp, delimiter=";"):
Xy.append(row)
Xy = np.array(Xy[1:], dtype=np.float64)
np.random.seed(0)
np.random.shuffle(Xy)
train_X = Xy[:-1000, :-1]
train_y = Xy[:-1000, -1]
test_X = Xy[-1000:, :-1]
test_y = Xy[-1000:, -1]
# ハイパーパラメータを変えながら学習
for lambda_ in [1., 0.1, 0.01]:
model = lasso.Lasso(lambda_)
model.fit(train_X, train_y)
y = model.predict(test_X)
print("--- lambda --- {} ---".format(lambda_))
print("coefficients:")
print(model.w_)
mse = ((y - test_y)**2).mean()
print("MSE: {:.3f}".format(mse))
def main():
print("binary logistic regression")
df_train, df_test = h.load_crime()
X_train = df_train.iloc[:, 1:].values
Y_train = df_train.iloc[:, :1].values
X_test = df_test.iloc[:, 1:].values
Y_test = df_test.iloc[:, :1].values
input_indices = [
df_train.columns.get_loc(col_name) for col_name in c.input_variables
]
lamb_data = []
nonzero_data = []
w_list = []
b_list = []
lamb = h.min_lamb(X_train, Y_train)
crime_train_lasso = lasso.Lasso(X_train, Y_train)
crime_test_lasso = lasso.Lasso(X_test, Y_test)
w_zero = None
while True:
print("crime lasso lambda value: " + str(lamb))
if w_zero is None:
w_zero, b = crime_train_lasso.coord_desc(lamb)
w = w_zero
else:
w, b = crime_train_lasso.coord_desc(lamb, w=w_zero)
w_list.append(np.copy(w))
b_list.append(b)
lamb_data.append(lamb)
nz_count = np.count_nonzero(w)
nonzero_data.append(nz_count)
if lamb < c.cutoff:
break
lamb = lamb / 2
data_list = []
for index in input_indices:
data_list.append([w_list[i][index - 1] for i in range(len(w_list))])
h.plot_single("Nonzero Coefficients over Lambda", "a5_c", "Lambda",
"Nonzero Coefficients", lamb_data, nonzero_data, True)
h.plot_multiple("Nonzero Coefficients over Lambda", "a5_d", "Lambda",
"Weights", lamb_data, data_list, c.input_variables, True)
train_sqerror_list = []
test_sqerror_list = []
w_list = np.array(w_list).T
b_list = np.array(b_list)
train_sqerror_list = crime_train_lasso.get_sqerror(w_list, b_list)
test_sqerror_list = crime_test_lasso.get_sqerror(w_list, b_list)
h.plot_function("Squared Error over Lambda",
"a5_e",
"Lambda",
"Squared Error",
train_sqerror_list,
test_sqerror_list,
x_data=lamb_data,
log_scale=True)
lamb = 30
if lamb == 30:
# print("crime lasso lambda value: " + str(lamb))
w30_train, b30_train = crime_train_lasso.coord_desc(lamb)
w30_test, b30_test = crime_test_lasso.coord_desc(lamb)
all_variables = df_train.columns[1:]
nz_train = {
all_variables[i]: w30_train[i]
for i in range(len(w30_train)) if w30_train[i] != 0
}
nz_test = {
all_variables[i]: w30_test[i]
for i in range(len(w30_test)) if w30_test[i] != 0
}
output = open(c.results_path + "a5e" + c.txt_exten, "w")
output.write("nonzero training weights\n")
output.write(str(tabulate(nz_train.items())))
output.write("\n\nnonzero test weights\n")
output.write(str(tabulate(nz_test.items())))
# find the min/max entries
output.write("\n\ntraining min/max\n")
min_train = min(nz_train, key=nz_train.get)
output.write(min_train + " : " + str(nz_train.get(min_train)) + "\n")
max_train = max(nz_train, key=nz_train.get)
output.write(max_train + " : " + str(nz_train.get(max_train)) + "\n")
output.write("\ntest min/max\n")
min_test = min(nz_test, key=nz_test.get)
output.write(min_test + " : " + str(nz_test.get(min_test)) + "\n")
max_test = max(nz_test, key=nz_test.get)
output.write(max_test + " : " + str(nz_test.get(max_test)) + "\n")
output.close()
