def get_data_g_forces_v1(model_id, driver_id, repeat, test=False, min_df=1, ngram_range=(1,10), digitize=0):
def process(ride, digitize):
g_forces = util.get_g_forces(ride)
if digitize:
g_forces = np.digitize(g_forces, range(0, 800, digitize))
return util.get_list_string(g_forces)
seed = random.Random(x=driver_id+model_id)
da = DataAccess()
if test:
set1 = list(da.get_rides(driver_id)) # first half of the train set
set2 = list(da.get_random_rides(settings.BIG_CHUNK_TEST * repeat, driver_id, seed=seed)) # second half of the train set
else:
driver_train, driver_test = da.get_rides_split(driver_id, settings.BIG_CHUNK)
other_train = list(da.get_random_rides(settings.BIG_CHUNK * repeat, driver_id, seed=seed))
other_test = list(da.get_random_rides(settings.SMALL_CHUNK, driver_id))
set1 = driver_train + other_train # used for training
set2 = driver_test + other_test # used for testing
set1 = [process(ride, digitize) for ride in set1]
set2 = [process(ride, digitize) for ride in set2]
vectorizer = CountVectorizer(min_df=min_df, ngram_range=ngram_range)
set1 = vectorizer.fit_transform(set1)
set2 = vectorizer.transform(set2)
return set1, set2
def segment_driver(driver_id):
''' this generated the segments in settings.SEGMENTS_FOLDER[1] '''
da = DataAccess()
for ride_id_minus_1, ride in enumerate(da.get_rides(driver_id)):
ride_id = ride_id_minus_1 + 1
if da.skip_segment(driver_id, ride_id):
continue
# apply the Ramer-Douglas-Peucker algorithm
ride = [p + [i] for i, p in enumerate(smoothen(ride))] # enrich with timestamp
ride = rdp(ride, epsilon=10)
lengths = [util.euclidian_distance(ride[i-1], ride[i]) for i in xrange(1, len(ride))]
times = [ride[i][2] - ride[i-1][2] for i in xrange(1, len(ride))]
angles = [util.get_angle(ride[i-2], ride[i-1], ride[i]) for i in xrange(2, len(ride))]
# bucket the values
lengths = util.bucket(np.log(lengths), 25, [2.2,8]) # [int(l) for l in lengths]
times = util.bucket(np.log(times), 20, [1,5.5]) # [int(t) for t in times]
angles = util.bucket(angles, 30, [0,180]) # [int(a) for a in angles]
# write results
da.write_ride_segments(driver_id, ride_id, lengths, times, angles)
logging.info('finished segmenting driver %s' % driver_id)
def get_data_movements_accel(model_id, driver_id, repeat, test=False, step=3, tf=False, extra=((1,15),2), version=1):
seed = random.Random(x=driver_id+model_id)
da = DataAccess()
ngram_range, min_df = extra
if test:
set1 = list(da.get_rides(driver_id))
set2 = list(da.get_random_rides(settings.BIG_CHUNK_TEST * repeat, driver_id, segments=False, seed=seed))
else:
driver_train, driver_test = da.get_rides_split(driver_id, settings.BIG_CHUNK, segments=False)
other_train = list(da.get_random_rides(settings.BIG_CHUNK * repeat, driver_id, segments=False, seed=seed))
other_test = list(da.get_random_rides(settings.SMALL_CHUNK, driver_id, segments=False))
set1 = driver_train + other_train
set2 = driver_test + other_test
set1 = [util.build_features4(r, step=step, version=version) for r in set1]
set2 = [util.build_features4(r, step=step, version=version) for r in set2]
if tf:
vectorizer = TfidfVectorizer(min_df=min_df, ngram_range=ngram_range)
else:
vectorizer = CountVectorizer(min_df=min_df, ngram_range=ngram_range)
set1 = vectorizer.fit_transform(set1)
set2 = vectorizer.transform(set2)
return set1, set2
def get_data_movements_v1(model_id,
driver_id,
repeat,
test=False,
step=5,
tf=False,
version=1,
extra=((1, 5), 2)):
seed = random.Random(x=driver_id + model_id)
da = DataAccess()
ngram_range, min_df = extra
if test:
set1 = list(da.get_rides(driver_id))
set2 = list(
da.get_random_rides(settings.BIG_CHUNK_TEST * repeat,
driver_id,
segments=False,
seed=seed))
else:
driver_train, driver_test = da.get_rides_split(driver_id,
settings.BIG_CHUNK,
segments=False)
other_train = list(
da.get_random_rides(settings.BIG_CHUNK * repeat,
driver_id,
segments=False,
seed=seed))
other_test = list(
da.get_random_rides(settings.SMALL_CHUNK,
driver_id,
segments=False))
set1 = driver_train + other_train
set2 = driver_test + other_test
# keep only lengths and convert to text
set1 = [util.build_features3(r, step=step, version=version) for r in set1]
set2 = [util.build_features3(r, step=step, version=version) for r in set2]
if tf:
vectorizer = TfidfVectorizer(min_df=min_df, ngram_range=ngram_range)
else:
vectorizer = CountVectorizer(min_df=min_df, ngram_range=ngram_range)
set1 = vectorizer.fit_transform(set1)
set2 = vectorizer.transform(set2)
return set1, set2
def test_model_heading(model_id, driver_id, Model, get_data, repeat):
seed = random.Random(x=driver_id+model_id)
da = DataAccess()
set1 = list(da.get_rides(driver_id)) # first half of the train set
set2 = list(da.get_random_rides(settings.BIG_CHUNK_TEST * repeat, driver_id, seed=seed)) # second half of the train set
moving_average_window = 6 if get_data == get_data_heading_v2 else 3
set1 = [heading.get_ride_heading(ride, variations=True, \
moving_average_window=moving_average_window) for ride in set1]
set2 = [heading.get_ride_heading(ride, variations=True, \
moving_average_window=moving_average_window) for ride in set2]
set1 = [[util.get_list_string(r) for r in four_pack] for four_pack in set1]
set2 = [[util.get_list_string(r) for r in four_pack] for four_pack in set2]
vectorizer = CountVectorizer(min_df=2, ngram_range=(1,15), max_df=1000000)
vectorizer.fit([r[0] for r in set1])
rides = [[vectorizer.transform([r])[0] for r in four_pack] for four_pack in set1]
other_rides = [[vectorizer.transform([r])[0] for r in four_pack] for four_pack in set2]
other_rides = list(itertools.chain(*other_rides))
rides = np.array(rides)
trainY = [1] * settings.BIG_CHUNK_TEST * 4 * repeat + [0] * settings.BIG_CHUNK_TEST * 4 * repeat
kf = KFold(200, n_folds=settings.FOLDS, shuffle=True, random_state=driver_id)
predictions = ['bug'] * 200
for train_fold, test_fold in kf:
trainX = rides[train_fold]
trainX = scipy.sparse.vstack(
list(itertools.chain(*trainX)) * repeat + \
other_rides
)
testX = scipy.sparse.vstack([r[0] for r in rides[test_fold]])
assert(trainX.shape[0] == len(trainY))
assert(testX.shape[0] == settings.SMALL_CHUNK_TEST)
model = Model(trainX, trainY, driver_id)
fold_predictions = model.predict(testX)
for i, v in enumerate(test_fold):
predictions[v] = fold_predictions[i]
predictions = np.array(predictions)
if settings.ENABLE_CACHE:
util.cache_results(Model, get_data, driver_id, True, predictions, repeat)
return driver_id, predictions
def get_data_basic_accel(model_id, driver_id, repeat, test=False, version=1):
seed = random.Random(x=driver_id+model_id)
da = DataAccess()
if test:
set1 = list(da.get_rides(driver_id))
set2 = list(da.get_random_rides(settings.BIG_CHUNK_TEST * repeat, driver_id, seed=seed))
else:
driver_train, driver_test = da.get_rides_split(driver_id, settings.BIG_CHUNK)
other_train = list(da.get_random_rides(settings.BIG_CHUNK * repeat, driver_id, seed=seed))
other_test = list(da.get_random_rides(settings.SMALL_CHUNK, driver_id))
set1 = driver_train + other_train
set2 = driver_test + other_test
set1 = [util.build_features_acc(ride, version=version) for ride in set1]
set2 = [util.build_features_acc(ride, version=version) for ride in set2]
return np.array(set1), np.array(set2)
def get_data_g_forces_v1(model_id,
driver_id,
repeat,
test=False,
min_df=1,
ngram_range=(1, 10),
digitize=0):
def process(ride, digitize):
g_forces = util.get_g_forces(ride)
if digitize:
g_forces = np.digitize(g_forces, range(0, 800, digitize))
return util.get_list_string(g_forces)
seed = random.Random(x=driver_id + model_id)
da = DataAccess()
if test:
set1 = list(da.get_rides(driver_id)) # first half of the train set
set2 = list(
da.get_random_rides(settings.BIG_CHUNK_TEST * repeat,
driver_id,
seed=seed)) # second half of the train set
else:
driver_train, driver_test = da.get_rides_split(driver_id,
settings.BIG_CHUNK)
other_train = list(
da.get_random_rides(settings.BIG_CHUNK * repeat,
driver_id,
seed=seed))
other_test = list(da.get_random_rides(settings.SMALL_CHUNK, driver_id))
set1 = driver_train + other_train # used for training
set2 = driver_test + other_test # used for testing
set1 = [process(ride, digitize) for ride in set1]
set2 = [process(ride, digitize) for ride in set2]
vectorizer = CountVectorizer(min_df=min_df, ngram_range=ngram_range)
set1 = vectorizer.fit_transform(set1)
set2 = vectorizer.transform(set2)
return set1, set2
def get_data_dist_acc(model_id, driver_id, repeat, test=False):
seed = random.Random(x=driver_id+model_id)
da = DataAccess()
if test:
set1 = list(da.get_rides(driver_id))
set2 = list(da.get_random_rides(settings.BIG_CHUNK_TEST * repeat, driver_id, seed=seed))
else:
driver_train, driver_test = da.get_rides_split(driver_id, settings.BIG_CHUNK)
other_train = list(da.get_random_rides(settings.BIG_CHUNK * repeat, driver_id, seed=seed))
other_test = list(da.get_random_rides(settings.SMALL_CHUNK, driver_id))
set1 = driver_train + other_train
set2 = driver_test + other_test
set1 = [util.get_distance_acc_words(ride, step=3) for ride in set1]
set2 = [util.get_distance_acc_words(ride, step=3) for ride in set2]
vectorizer = CountVectorizer(min_df=1, ngram_range=(1,15))
set1 = vectorizer.fit_transform(set1)
set2 = vectorizer.transform(set2)
return set1, set2
def get_data_g_forces_v6(model_id, driver_id, repeat, test=False, version=1):
seed = random.Random(x=driver_id+model_id)
da = DataAccess()
if test:
set1 = list(da.get_rides(driver_id)) # first half of the train set
set2 = list(da.get_random_rides(settings.BIG_CHUNK_TEST * repeat, driver_id, seed=seed)) # second half of the train set
else:
driver_train, driver_test = da.get_rides_split(driver_id, settings.BIG_CHUNK)
other_train = list(da.get_random_rides(settings.BIG_CHUNK * repeat, driver_id, seed=seed))
other_test = list(da.get_random_rides(settings.SMALL_CHUNK, driver_id))
set1 = driver_train + other_train # used for training
set2 = driver_test + other_test # used for testing
set1 = [util.get_g_forces_v4(ride, version=version) for ride in set1]
set2 = [util.get_g_forces_v4(ride, version=version) for ride in set2]
vectorizer = CountVectorizer(min_df=1, ngram_range=(1,20))
set1 = vectorizer.fit_transform(set1)
set2 = vectorizer.transform(set2)
return set1, set2
def get_data_fft(model_id, driver_id, repeat, test=False, version=1):
seed = random.Random(x=driver_id+model_id)
da = DataAccess()
if test:
set1 = list(da.get_rides(driver_id))
set2 = list(da.get_random_rides(settings.BIG_CHUNK_TEST * repeat, driver_id, seed=seed))
else:
driver_train, driver_test = da.get_rides_split(driver_id, settings.BIG_CHUNK)
other_train = list(da.get_random_rides(settings.BIG_CHUNK * repeat, driver_id, seed=seed))
other_test = list(da.get_random_rides(settings.SMALL_CHUNK, driver_id))
set1 = driver_train + other_train
set2 = driver_test + other_test
if version == 1:
set1 = [util.fft(ride) for ride in set1]
set2 = [util.fft(ride) for ride in set2]
else:
set1 = [util.fft_strip(ride) for ride in set1]
set2 = [util.fft_strip(ride) for ride in set2]
return np.array(set1), np.array(set2)
def segment_driver_v2(driver_id):
''' this generated the segments in settings.SEGMENTS_FOLDER[2] '''
da = DataAccess()
for ride_id_minus_1, ride in enumerate(da.get_rides(driver_id)):
ride_id = ride_id_minus_1 + 1
if da.skip_segment(driver_id, ride_id, version=2):
continue
# apply the Ramer-Douglas-Peucker algorithm
ride = [p + [i] for i, p in enumerate(ride)] # enrich with timestamp
ride = rdp(ride, epsilon=4)
lengths = [
util.euclidian_distance(ride[i - 1], ride[i])
for i in range(1, len(ride))
]
times = [ride[i][2] - ride[i - 1][2] for i in range(1, len(ride))]
angles = [
util.get_angle(ride[i - 2], ride[i - 1], ride[i])
for i in range(2, len(ride))
]
lengths = np.histogram(lengths,
bins=list(range(0, 700, 20)) + [1000000000])[0]
times = np.histogram(times,
bins=list(range(0, 60, 4)) + [1000000000])[0]
angles = np.histogram(angles, bins=list(range(0, 181, 20)))[0]
# write results
da.write_ride_segments(driver_id,
ride_id,
lengths,
times,
angles,
version=2)
logging.info('finished segmenting driver %s' % driver_id)
def get_data_basic(model_id,
driver_id,
repeat,
test=False,
normalized=False,
version=1):
seed = random.Random(x=driver_id + model_id)
da = DataAccess()
if test:
set1 = list(da.get_rides(driver_id))
set2 = list(
da.get_random_rides(settings.BIG_CHUNK_TEST * repeat,
driver_id,
seed=seed))
else:
driver_train, driver_test = da.get_rides_split(driver_id,
settings.BIG_CHUNK)
other_train = list(
da.get_random_rides(settings.BIG_CHUNK * repeat,
driver_id,
seed=seed))
other_test = list(da.get_random_rides(settings.SMALL_CHUNK, driver_id))
set1 = driver_train + other_train
set2 = driver_test + other_test
set1 = [
util.build_features(ride, normalized=normalized, version=version)
for ride in set1
]
set2 = [
util.build_features(ride, normalized=normalized, version=version)
for ride in set2
]
return np.array(set1), np.array(set2)
def segment_driver_v2(driver_id):
''' this generated the segments in settings.SEGMENTS_FOLDER[2] '''
da = DataAccess()
for ride_id_minus_1, ride in enumerate(da.get_rides(driver_id)):
ride_id = ride_id_minus_1 + 1
if da.skip_segment(driver_id, ride_id, version=2):
continue
# apply the Ramer-Douglas-Peucker algorithm
ride = [p + [i] for i, p in enumerate(ride)] # enrich with timestamp
ride = rdp(ride, epsilon=4)
lengths = [util.euclidian_distance(ride[i-1], ride[i]) for i in xrange(1, len(ride))]
times = [ride[i][2] - ride[i-1][2] for i in xrange(1, len(ride))]
angles = [util.get_angle(ride[i-2], ride[i-1], ride[i]) for i in xrange(2, len(ride))]
lengths = np.histogram(lengths, bins=range(0, 700, 20) + [1000000000])[0]
times = np.histogram(times, bins=range(0, 60, 4) + [1000000000])[0]
angles = np.histogram(angles, bins=range(0, 181, 20))[0]
# write results
da.write_ride_segments(driver_id, ride_id, lengths, times, angles, version=2)
logging.info('finished segmenting driver %s' % driver_id)
