In [1]:
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.model_selection import GridSearchCV
from sktime.datasets import load_gunpoint, load_italy_power_demand
from sktime_dl.deeplearning import CNNClassifier
sns.set_style('whitegrid')
Load a dataset¶
In [2]:
X_train, y_train = load_italy_power_demand(split='train', return_X_y=True)
X_test, y_test = load_italy_power_demand(split='test', return_X_y=True)
X_train.head()
Out[2]:
| dim_0 | |
|---|---|
| 0 | 0 -0.710520 1 -1.183300 2 -1.372400 3... |
| 1 | 0 -0.993010 1 -1.426800 2 -1.579900 3... |
| 2 | 0 1.319100 1 0.569770 2 0.195130 3... |
| 3 | 0 -0.812440 1 -1.157600 2 -1.416400 3... |
| 4 | 0 -0.972840 1 -1.390500 2 -1.536700 3... |
In [3]:
i = 0
X_train.loc[i, "dim_0"]#.iloc[0]
X_train.loc[i, "dim_0"].iloc[0]
Out[3]:
-0.71052
In [4]:
def plot_data_samples(X, y, sample_numbers):
'''
Plot the time series data relating to the input list of sample numbers.
sample_numbers: list of integers
E.g. [1, 7, 22, 42]
'''
unique_labels = np.unique(y).astype(int)
num_classes = len(unique_labels)
if num_classes<=4:
class_colors = ['red', 'blue', 'green' , 'orange']
else:
class_colors = sns.color_palette(n_colors=num_classes)
fig, ax = plt.subplots()
for i in sample_numbers:
print('sample', i, 'class', str(y[i]))
color_num = y[i].astype(int) - unique_labels.min()
X_train.loc[i, "dim_0"].plot(label=str(y[i]), color=class_colors[color_num])
print('')
plt.ylim([-3.5, 3.5])
if num_classes<=2:
title = class_colors[0]+' : class '+str(unique_labels[0])
title = title + '\n'+class_colors[1]+' : class '+str(unique_labels[1])
plt.title(title)
ax.set_ylabel('Data value')
ax.set_xlabel('Data point number')
Plot some data samples¶
In [5]:
plot_data_samples(X_train, y_train, [0, 1, 2, 3])
sample 0 class 1 sample 1 class 1 sample 2 class 2 sample 3 class 2
Train a single deep neural network classifier¶
Here we choose to use the CNN (convolutional neural network) classifier. Other classifiers provided by sktime-dl include MLP, ResNet and InceptionTime.
In [6]:
network = CNNClassifier(nb_epochs=200, verbose=False)
network.fit(X_train, y_train)
network.score(X_test, y_test)
Out[6]:
0.6180758017492711
Save the model to file¶
In [7]:
network.model.save("temp_model.h5")
Grid Search¶
sktime-dl is compatible with scikit-learn and can use sklearn's GridSearchCV.
Here we search over two parameters, number of epochs and CNN kernel size.
In [8]:
param_grid = {'nb_epochs': [50, 100],
'kernel_size': [5, 7, 9] }
grid = GridSearchCV(network, param_grid=param_grid, cv=5)
grid.fit(X_train, y_train)
print("Best cross-validation accuracy: {:.2f}".format(grid.best_score_))
print("Test set score: {:.2f}".format(grid.score(X_test, y_test)))
print("Best parameters: {}".format(grid.best_params_))
Best cross-validation accuracy: 0.68
Test set score: 0.53
Best parameters: {'kernel_size': 9, 'nb_epochs': 100}