GillesVandewiele
commented
2 years ago DTW is a rather slow function (O(m**2) with m the length of time series)), and KNN will internally keep track of pairwise distances of those 20K time series (although with optimized data structures such as BallTree and KDTree). Hence, why it is rather slow.
liorLimudim
Greetings As part of testing of different machine learning models I test the KNN model together with DTW and encounter a great slowness in the predictor function of the classifier, I would be very grateful if you could explain to me why there is such a great slowness and how to calibrate the classifier in such a way that it is not slow, attach the code That I run Thank you Lior
import numpy as np from tslearn.generators import random_walk_blobs from tslearn.neighbors import KNeighborsTimeSeriesClassifier, KNeighborsTimeSeries import time
np.random.seed(0) n_ts_per_blob, sz = 20000, 100 #, 1, 1 # , d, n_blobs X, y = random_walk_blobs(n_ts_per_blob=n_ts_per_blob, sz=sz) print(y) print(X.shape) print(len(y))
start_time = time.time() knn = KNeighborsTimeSeriesClassifier(n_neighbors=1,metric="dtw",metric_params=dict({"global_constraint":"sakoe_chiba","sakoe_chiba_radius":3}),n_jobs=-1) end_time = time.time() print("Creating the Classifier took: {} seconds".format(end_time - start_time))
start_time = time.time() knn.fit(X,y) end_time = time.time() print("FIT function took: {} seconds".format(end_time - start_time))
start_time = time.time() y_predict = knn.predict(X) end_time = time.time() print("Predict function took: {} seconds".format(end_time - start_time))