def preprocessing_raw_csv(PATH=".//tcdata//hy_round2_train_20200225//",
local_file_name="train.pkl"):
"""Loading and processing all train csv data."""
if PATH is None:
raise ValueError("Invalid PATH !")
file_names = sorted(os.listdir(PATH), key=lambda s: int(s.split(".")[0]))
# Loading all trajectory data.
traj_data = []
for name in file_names:
traj_data.append(pd.read_csv(PATH + name, encoding="utf-8"))
# Processing each trajectory data.
print("\n@Multi-processing RAW CSV started:")
print("-----------------------------")
with mp.Pool(processes=mp.cpu_count()) as p:
tmp = list(tqdm(p.imap(preprocessing_traj, traj_data),
total=len(traj_data)))
print("-----------------------------")
print("@Multi-processing RAW CSV ended, to the local file: {}.\n".format(
local_file_name))
traj_data = [item[0] for item in tmp]
change_record = [item[1] for item in tmp]
change_record = pd.DataFrame(change_record,
columns=["speed_change", "coord_change"])
# Saving processed data to the lcoal path with *.pkl format
file_processor = LoadSave(PATH)
file_processor.save_data(path=".//tcdata_tmp//{}".format(local_file_name),
data=traj_data)
return change_record
def traj_data_train_test_split(train_ratio=0.85):
"""Split the training data into training dataset and testing dataset, This
is for the online docker docker evaluation testing.
"""
PATH = ".//tcdata//hy_round2_train_20200225_local//"
file_names = sorted(os.listdir(PATH), key=lambda s: int(s.split(".")[0]))
print("\n@Read all raw traj data started at: {}".format(datetime.now()))
print("-----------------------------")
with mp.Pool(processes = mp.cpu_count()) as p:
traj_data_total = list(tqdm(p.imap(read_traj, file_names),
total=len(file_names)))
print("-----------------------------")
print("@End at: {}".format(datetime.now()))
# Map the Chinese labels into the numerics
str_to_label = {"刺网": 2, "围网": 1, "拖网": 0}
target = [traj["type"].unique()[0] for traj in traj_data_total]
target = np.array([str_to_label[i] for i in target])
train_index, test_index = traj_data_train_test_index_generation(
train_ratio=train_ratio, n_samples=len(traj_data_total),
target=target, method="stratified")
traj_data_train = [traj_data_total[i] for i in train_index]
traj_data_train_fnames = [file_names[i] for i in train_index]
traj_data_test = [traj_data_total[i] for i in test_index]
traj_data_test_fnames = [file_names[i] for i in test_index]
train_target_dist = [target[i] for i in train_index]
test_target_dist = [target[i] for i in test_index]
print("@Total target distributions: {}".format(
np.bincount(target)/len(target)))
print("@Train distributions: {}".format(
np.bincount(train_target_dist)/len(traj_data_train)))
print("@Test distributions: {}".format(
np.bincount(test_target_dist)/len(traj_data_test)))
TEST_TARGET_PATH = ".//tcdata_tmp//"
boat_id = [int(file_names[i].split(".")[0]) for i in test_index]
df = pd.DataFrame({"boat_id": boat_id, "target": test_target_dist})
file_processor = LoadSave()
file_processor.save_data(data=df, path=TEST_TARGET_PATH+"test_target.pkl")
TRAIN_DATA_PATH = ".//tcdata//hy_round2_train_20200225//"
file_names = os.listdir(TRAIN_DATA_PATH)
if len(file_names) != 0:
raise ValueError("The dir is not empty ! Please remove all file ~~")
for df, name in zip(traj_data_train, traj_data_train_fnames):
df.to_csv(TRAIN_DATA_PATH + name, index=False, encoding="utf-8")
TEST_PATH = ".//tcdata//hy_round2_testA_20200225//"
file_names = os.listdir(TEST_PATH)
if len(file_names) != 0:
raise ValueError("The dir is not empty ! Please remove all files ~~")
for df, name in zip(traj_data_test, traj_data_test_fnames):
df.to_csv(TEST_PATH + name, index=False, encoding="utf-8")
return traj_data_train, traj_data_test
def traj_data_signal_embedding():
"""Loading the embedding vectors."""
file_processor = LoadSave()
train_embedding = file_processor.load_data(
path=".//tcdata_tmp//train_signal_embedding.pkl")
test_embedding = file_processor.load_data(
path=".//tcdata_tmp//test_signal_embedding.pkl")
return pd.concat([train_embedding, test_embedding],
axis=0,
ignore_index=True)
def find_save_unique_visit_count_table(traj_data_df=None, bin_to_coord_df=None):
"""Find and save the unique boat visit count of each bin."""
unique_boat_count_df = traj_data_df.groupby(["no_bin"])["boat_id"].nunique().reset_index()
unique_boat_count_df.rename({"boat_id":"visit_boat_count"}, axis=1, inplace=True)
unique_boat_count_df_save = pd.merge(bin_to_coord_df, unique_boat_count_df,
on="no_bin", how="left")
file_processor = LoadSave()
file_processor.save_data(data=unique_boat_count_df_save,
path=".//tcdata_tmp//bin_unique_boat_count_frequency.pkl")
return unique_boat_count_df
def find_save_visit_count_table(traj_data_df=None, bin_to_coord_df=None):
"""Find and save the visit frequency of each bin."""
visit_count_df = traj_data_df.groupby(["no_bin"]).count().reset_index()
visit_count_df = visit_count_df[["no_bin", "x"]]
visit_count_df.rename({"x":"visit_count"}, axis=1, inplace=True)
visit_count_df_save = pd.merge(bin_to_coord_df, visit_count_df, on="no_bin", how="left")
file_processor = LoadSave()
file_processor.save_data(data=visit_count_df_save,
path=".//tcdata_tmp//bin_visit_count_frequency.pkl")
return visit_count_df
def preprocessing_mnist(n_data_list=None):
img_data = pd.read_csv("..//demo_dataset//mnist//train.csv", nrows=None)
img_cols = [name for name in img_data.columns if "pixel" in name]
img_data = img_data[img_cols].values / 255
img_data_list = img_data.tolist()
# Save the proprocessed data
file_name = [".//data//mnist_{}.pkl".format(i) for i in n_data_list]
file_processor = LoadSave()
for ind, item in enumerate(n_data_list):
tmp_img_data = img_data_list[:item]
tmp_file_name = file_name[ind]
file_processor.save_data(path=tmp_file_name, data=tmp_img_data)
def find_save_mean_stay_time_table(traj_data_df=None, bin_to_coord_df=None):
"""Find and save the mean stay time of each bin."""
mean_stay_time_df = traj_data_df.groupby(
["no_bin", "boat_id"])["time_array"].sum().reset_index()
mean_stay_time_df.rename({"time_array":"total_stay_time"}, axis=1, inplace=True)
mean_stay_time_df = mean_stay_time_df.groupby(
["no_bin"])["total_stay_time"].mean().reset_index()
mean_stay_time_df.rename(
{"total_stay_time":"mean_stay_time"}, axis=1, inplace=True)
mean_stay_time_df_save = pd.merge(bin_to_coord_df, mean_stay_time_df,
on="no_bin", how="left")
file_processor = LoadSave()
file_processor.save_data(data=mean_stay_time_df_save,
path=".//tcdata_tmp//bin_mean_stay_time.pkl")
return mean_stay_time_df
def save_ais_traj_to_csv(ais=None,
round_to_print=50000,
local_file_name="ais.pkl"):
"""Save the trajectory according to the ais record with the csv format."""
ais_id_list = ais["ais_id"].astype(int).values.tolist()
# Split the DataFrame
ais_traj_list = []
head, tail = 0, 0
print("\n@Split AIS and save the traj in *.csv format:")
print("---------------------------------------")
while (tail <= (len(ais_id_list) - 1)):
if tail % round_to_print == 0:
print("--Now tail is on {}, completed {:.2f}%.".format(
tail, (tail + 1) / len(ais_id_list) * 100))
print("--time is {}.\n".format(datetime.now()))
if ais_id_list[head] == ais_id_list[tail]:
tail += 1
elif ais_id_list[head] != ais_id_list[tail]:
ais_traj_list.append(ais.iloc[head:tail])
head = tail
ais_traj_list.append(ais.iloc[head:])
print("---------------------------------------")
# Coordinate transferring
# tmp = []
# for i in range(50):
# tmp.append(preprocessing_traj(ais_traj_list[i]))
print("\n@AIS list index resetting:")
print("---------------------------------------")
with mp.Pool(processes=mp.cpu_count()) as p:
tmp = list(
tqdm(p.imap(preprocessing_traj, ais_traj_list),
total=len(ais_traj_list)))
print("---------------------------------------")
print("@Save to the local file: {}.\n".format(local_file_name))
traj_data = [item[0] for item in tmp if len(item[0]) > 1]
change_record = [item[1] for item in tmp]
change_record = pd.DataFrame(change_record,
columns=["speed_change", "coord_change"])
# Saving processed data to the lcoal path with *.pkl format
file_processor = LoadSave()
file_processor.save_data(path=".//tcdata_tmp//{}".format(local_file_name),
data=traj_data)
return traj_data, change_record
def preparing_traj_data_corpus(bin_size=600):
"""Preparing the training corpus for the traj2vec model."""
# Loading all the data
train_data = load_data("train_semantic.pkl")
test_data = load_data("test_semantic.pkl")
ais_data = load_data("ais_semantic.pkl")
train_concat = load_data("train_semantic_concat.pkl")
test_concat = load_data("test_semantic_concat.pkl")
# Print statistics
x_min = min(train_concat["x"].min(), test_concat["x"].min())
x_max = max(train_concat["x"].max(), test_concat["x"].max())
y_min = min(train_concat["y"].min(), test_concat["y"].min())
y_max = max(train_concat["y"].max(), test_concat["y"].max())
col_bins = int((x_max - x_min) / bin_size)
row_bins = int((y_max - y_min) / bin_size)
# Start cutting the traj to bins
traj_total = train_data + test_data + ais_data
res = []
# Multi-processing for loop.
partial_work = partial(traj_to_bin,
col_bins=col_bins,
row_bins=row_bins,
x_min=x_min,
x_max=x_max,
y_min=y_min,
y_max=y_max)
with mp.Pool(processes=mp.cpu_count()) as p:
res = list(
tqdm(p.imap(partial_work, traj_total), total=len(traj_total)))
unique_words = [traj["no_bin"].nunique() for traj in res]
print("\n@Cutting results basic stat:")
print("-----------------------------")
print("@Mean uniques: {:.5f}, max: {}, median: {:.5f}, std: {:.5f}".format(
np.mean(unique_words), np.max(unique_words), np.median(unique_words),
np.std(unique_words)))
print("-----------------------------\n")
file_processor = LoadSave()
file_processor.save_data(path=".//tcdata_tmp//traj_data_corpus.pkl",
data=res)
def preprocessing_HAR(n_data_list=None):
file_processor = LoadSave()
har_dataset, har_dataset_label = file_processor.load_data(
path=
"..//demo_dataset//human_activity_recognition//human_activity_recognition.pkl"
)
har_dataset_label = np.array(har_dataset_label)
# Shuffle the dataset
ind = np.random.choice(np.arange(0, len(har_dataset_label)),
size=len(har_dataset_label),
replace=False)
har_dataset = har_dataset[ind]
har_dataset_label = har_dataset_label[ind]
for ind in range(len(n_data_list)):
if n_data_list[ind] is None:
n_data_list[ind] = len(har_dataset)
file_name = [
".//data//human_activity_recognition_{}.pkl".format(i)
for i in n_data_list
]
file_processor = LoadSave()
for ind, item in enumerate(n_data_list):
tmp_data = har_dataset[:item]
tmp_data_label = har_dataset_label[:item]
tmp_file_name = file_name[ind]
file_processor.save_data(path=tmp_file_name,
data=[tmp_data, tmp_data_label])
def preprocessing_turnout(n_data_list=None):
file_processor = LoadSave()
signal_data_list = []
# Experiment fault data
def liststr_to_listnumeric(list_str):
return list(map(float, list_str.split(",")))
signal_fault_data = pd.read_csv(
"..//demo_dataset//turnout//fault_data.csv",
nrows=None).query("error_code != 0").reset_index(drop=True)
signal_fault_data["Phase_A"] = signal_fault_data["Phase_A"].apply(
liststr_to_listnumeric)
signal_fault_data["Phase_B"] = signal_fault_data["Phase_B"].apply(
liststr_to_listnumeric)
signal_fault_data["Phase_C"] = signal_fault_data["Phase_C"].apply(
liststr_to_listnumeric)
for i in range(len(signal_fault_data)):
signal = [
signal_fault_data["Phase_A"].iloc[i],
signal_fault_data["Phase_B"].iloc[i],
signal_fault_data["Phase_C"].iloc[i]
]
signal_data_list.append(signal)
# Operation fault data
signal_data = file_processor.load_data(
path="..//demo_dataset//turnout//chengdu5_raw_table.pkl")
signal_anomaly_scores = file_processor.load_data(
path="..//demo_dataset//turnout//chengdu5_anomaly_scores.pkl")
signal_data = pd.merge(signal_data,
signal_anomaly_scores,
on=["device_id", "record_id"],
how="left")
signal_data = signal_data.sort_values(
by="if_score", ascending=False).reset_index(drop=True)
for i in range(len(signal_data)):
signal = [
signal_data["phase_a"].iloc[i], signal_data["phase_b"].iloc[i],
signal_data["phase_c"].iloc[i]
]
signal_data_list.append(signal)
# Save the proprocessed data
for ind in range(len(n_data_list)):
if n_data_list[ind] is None:
n_data_list[ind] = len(signal_data_list)
file_name = [
".//data//fault_turnout_current_{}.pkl".format(i) for i in n_data_list
]
for ind, item in enumerate(n_data_list):
tmp_signal_data = signal_data_list[:item]
tmp_file_name = file_name[ind]
file_processor.save_data(path=tmp_file_name, data=tmp_signal_data)
def traj_data_labeling_semantics():
'''
Step 1: Load all possible stop grids.
'''
traj_data_list, train_nums, test_nums = load_concat_train_test_ais()
pois = poi_classification()
'''
Step 2: Find all candiate stop points.
'''
# Label all semantic points
nn = NearestNeighbors(n_neighbors=1, radius=400)
clf = nn.fit(pois[["x", "y"]].values)
traj_data_semantic = label_traj_data_semantics(
traj_data_list, clf, pois.drop(["x", "y"], axis=1))
# Spliting the training and testing data
train_data = traj_data_semantic[:train_nums]
test_data = traj_data_semantic[train_nums:(train_nums+test_nums)]
ais_data = traj_data_semantic[(train_nums+test_nums):]
# Save all data and concat the training and testing data
print("\n@Semantic labeling results:")
print("-----------------------------")
print("#training: {}, #testing: {}, #AIS: {}.".format(
len(train_data), len(test_data), len(ais_data)))
print("-----------------------------\n")
file_processor = LoadSave()
file_processor.save_data(path=".//tcdata_tmp//train_semantic.pkl",
data=train_data)
file_processor.save_data(path=".//tcdata_tmp//test_semantic.pkl",
data=test_data)
file_processor.save_data(path=".//tcdata_tmp//ais_semantic.pkl",
data=ais_data)
'''
Step 3: Concat a list of traj data.
'''
concat_list_data(data_list=train_data,
local_file_name="train_semantic_concat.pkl")
concat_list_data(data_list=test_data,
local_file_name="test_semantic_concat.pkl")
def traj_data_cbow_embedding_generating(embedding_size=70,
iters=70,
min_count=3,
window_size=25,
num_runs=1):
traj_corpus = load_data("traj_data_corpus.pkl")
train_nums = len(
sorted(os.listdir(".//tcdata//hy_round2_train_20200225//"),
key=lambda s: int(s.split(".")[0])))
test_nums = len(
sorted(os.listdir(".//tcdata//hy_round2_testA_20200225//"),
key=lambda s: int(s.split(".")[0])))
df_list, model_list = traj_cbow_embedding(traj_corpus,
embedding_size=embedding_size,
iters=iters,
min_count=min_count,
window_size=window_size,
seed=9012,
num_runs=num_runs,
word_feat="no_bin")
train_embedding_df_list = [
df.iloc[:train_nums].reset_index(drop=True) for df in df_list
]
test_embedding_df_list = [
df.iloc[train_nums:(train_nums + test_nums)].reset_index(drop=True)
for df in df_list
]
file_processor = LoadSave()
file_processor.save_data(
path=".//tcdata_tmp//train_embedding_cbow_list.pkl",
data=train_embedding_df_list)
file_processor.save_data(
path=".//tcdata_tmp//test_embedding_cbow_list.pkl",
data=test_embedding_df_list)
tmp_df.rename(
{
"DOTTING_TIME": "DOTTING_TIME_SHIFT",
"QUEUE_ID": "QUEUE_ID_SHIFT",
feat_name: target_feat_name
},
axis=1,
inplace=True)
target_df = pd.concat([target_df, tmp_df], axis=1)
# Exclude invalid target values
target_df[target_feat_name][
target_df["QUEUE_ID"] != target_df["QUEUE_ID_SHIFT"]] = np.nan
target_df[target_feat_name][
abs(target_df["DOTTING_TIME"] -
target_df["DOTTING_TIME_SHIFT"]) > tol_time_diff] = np.nan
# Drop tmp columns
target_df.drop(["DOTTING_TIME_SHIFT", "QUEUE_ID_SHIFT"],
axis=1,
inplace=True)
# Save feat engineering results
# ----------------------------
file_processor = LoadSave()
total_results = [feat_df, target_df]
file_processor.save_data(
path=".//cached_data//{}.pkl".format("nn_dense_feat"),
data=total_results)
def load_data(name=None):
"""Load data from .//tcdata_tmp//"""
file_processor = LoadSave()
data = file_processor.load_data(path=".//tcdata_tmp//" + name)
return data
##################################################
train_feature = total_features.iloc[:train_nums].reset_index(
drop=True).copy()
test_feature = total_features.iloc[train_nums:].reset_index(
drop=True).copy()
train_feature["target"] = labels
print("\n-- Train samples: {}, testA samples: {}.".format(
len(train_feature), len(test_feature)))
print("-- Train cols: {}, test cols: {}.".format(train_feature.shape[1],
test_feature.shape[1]))
print("-- Unique train cols: {}, unique testA cols: {}.\n".format(
len(np.unique(train_feature.columns)),
len(np.unique(test_feature.columns))))
file_processor = LoadSave()
file_processor.save_data(path=".//tcdata_tmp//train_feature_xgb.pkl",
data=train_feature)
file_processor.save_data(path=".//tcdata_tmp//train_target.pkl",
data=train_feature[["boat_id", "target"]])
file_processor.save_data(path=".//tcdata_tmp//test_feature_xgb.pkl",
data=test_feature)
gc.collect()
for embedding_id in [0, 1, 2]:
xgb_res_list = xgb_clf_embedding_list_train(folds=5,
id_list=[embedding_id],
embedding_enable=True)
df = training_res_to_log(training_res=xgb_res_list[0][0],
comment="xgb_{}".format(embedding_id))
def traj_data_poi_mining(visit_count_minimum=200, visit_boat_minimum=3,
mean_stay_minutes=120, bin_size=800):
'''
Step 1: Find all possible stop grids.
'''
traj_data_list, train_nums, test_nums = load_concat_train_test_ais()
print("\n@Step 1: traj2bin:")
print("-----------------------------")
col_bins = int((14226964.881853 - 12031967.16239096) / bin_size)
row_bins = int((4689471.1780792 - 1623579.449434373) / bin_size)
partial_work = partial(traj_to_bin, col_bins=col_bins, row_bins=row_bins)
with mp.Pool(processes=mp.cpu_count()) as p:
res = list(tqdm(p.imap(partial_work, traj_data_list),
total=len(traj_data_list)))
print("-----------------------------")
traj_data_df = [traj[["x", "y", "no_bin", "lon",
"lat", "boat_id", "time_array"]] for traj in res]
traj_data_df = pd.concat(traj_data_df, axis=0, ignore_index=True)
bin_to_coord_df = traj_data_df.groupby(
["no_bin"]).median().reset_index().drop(["boat_id"], axis=1)
# DataFrame tmp for finding POIs
visit_count_df = find_save_visit_count_table(
traj_data_df, bin_to_coord_df)
unique_boat_count_df = find_save_unique_visit_count_table(
traj_data_df, bin_to_coord_df)
mean_stay_time_df = find_save_mean_stay_time_table(
traj_data_df, bin_to_coord_df)
candidate_pois = visit_count_df.query(
"visit_count >= {}".format(visit_count_minimum)).reset_index(drop=True)
candidate_pois = pd.merge(
candidate_pois, unique_boat_count_df, on="no_bin", how="left")
candidate_pois = candidate_pois.query(
"visit_boat_count >= {}".format(visit_boat_minimum)).reset_index(drop=True)
candidate_pois = pd.merge(
candidate_pois, mean_stay_time_df, on="no_bin", how="left")
candidate_pois = candidate_pois.query(
"mean_stay_time >= {}".format(mean_stay_minutes)).reset_index(drop=True)
candidate_pois = pd.merge(
candidate_pois, bin_to_coord_df, on="no_bin", how="left")
candidate_pois.drop(["time_array"], axis=1, inplace=True)
clf = DBSCAN(eps=1500, min_samples=200, n_jobs=-1, algorithm="kd_tree")
candidate_pois["label"] = clf.fit_predict(candidate_pois[["x", "y"]].values,
sample_weight=candidate_pois["visit_count"].values)
pois = candidate_pois[candidate_pois["label"] != -1]
pois.to_csv(".//tcdata_tmp//pois.csv", index=False)
# Labeling fishing ground
fishing_ground = load_fishing_ground()
fishing_ground_polygons = fishing_ground["arr"].values.tolist()
print("\n********************")
print("@AIS preprocessing start at: {}".format(datetime.now()))
print("********************")
partial_work = partial(find_fishing_ground, poly_vert_list=fishing_ground_polygons)
with mp.Pool(processes=mp.cpu_count()) as p:
tmp = list(tqdm(p.imap(partial_work, traj_data_list),
total=len(traj_data_list)))
print("\n********************")
print("@AIS preprocessing ended at: {}".format(datetime.now()))
print("********************")
traj_data_semantic = tmp
# Spliting the training and testing data
train_data = traj_data_semantic[:train_nums]
test_data = traj_data_semantic[train_nums:(train_nums+test_nums)]
ais_data = traj_data_semantic[(train_nums+test_nums):]
# Save all data and concat the training and testing data
print("\n@Semantic labeling results:")
print("-----------------------------")
print("#training: {}, #testing A: {}, #AIS: {}.".format(
len(train_data), len(test_data), len(ais_data)))
print("-----------------------------\n")
file_processor = LoadSave()
file_processor.save_data(path=".//tcdata_tmp//train_semantic_tmp.pkl",
data=train_data)
file_processor.save_data(path=".//tcdata_tmp//test_semantic_tmp.pkl",
data=test_data)
file_processor.save_data(path=".//tcdata_tmp//ais_semantic_tmp.pkl",
data=ais_data)
file_processor.save_data(path=".//tcdata_tmp//pois.pkl",
data=pois)
return candidate_pois, pois
def load_data(path_name=None):
"""Loading *.pkl from path_name, path_name is like: .//data//mnist.pkl"""
file_processor = LoadSave()
return file_processor.load_data(path=path_name)
def load_fishing_ground():
file_processor = LoadSave()
data = file_processor.load_data(".//tcdata//fishing_ground.pkl")
return data
def embedding_signal_sequence(speed_embedding=True,
dir_embedding=True,
speed_dir_embedding=True,
speed_filter_stops=False,
dir_filter_stops=True):
"""Training the signal embedding."""
train_data, test_data = load_train(), load_test()
traj_data_all = train_data + test_data
train_nums = len(train_data)
boat_id = [traj["boat_id"].unique()[0] for traj in traj_data_all]
total_embedding = pd.DataFrame(boat_id, columns=["boat_id"])
# Step 1: Construct the words
traj_data_corpus = []
for traj in traj_data_all:
traj["speed_str"] = traj["speed"].apply(lambda x: str(int(x * 100)))
traj["direction_str"] = traj["direction"].apply(str)
if speed_filter_stops:
traj["speed_str"][traj["is_stop"] != -1] = "0"
if dir_filter_stops:
traj["direction_str"][traj["is_stop"] != -1] = "0"
traj["speed_dir_str"] = traj["speed_str"] + "_" + traj["direction_str"]
traj_data_corpus.append(
traj[["boat_id", "speed_str", "direction_str", "speed_dir_str"]])
# traj_data_corpus = []
# for traj in traj_data_all:
# lon_val, lat_val = traj["lon"].values, traj["lat"].values
# angle = get_angle_from_coordinate(lat_val[1:], lon_val[1:],
# lat_val[:-1], lon_val[:-1]).tolist()
# angle = [angle[0]] + angle
#
# traj["speed_str"] = traj["speed"].apply(lambda x: str(int(x*100)))
# traj["direction"] = angle
# traj["direction_str"] = traj["direction"].apply(str)
# if speed_filter_stops:
# traj["speed_str"][traj["is_stop"] != -1] = "0"
# if dir_filter_stops:
# traj["direction_str"][traj["is_stop"] != -1] = "0"
#
# traj["speed_dir_str"] = traj["speed_str"] + "_" + traj["direction_str"]
# traj_data_corpus.append(traj[["boat_id", "speed_str",
# "direction_str", "speed_dir_str"]])
# Step 2: Training the speed information
if speed_embedding:
print("\n@Round 2 speed embedding:")
print("-----------------------------")
df_list, model_list = traj_cbow_embedding(traj_data_corpus,
embedding_size=10,
iters=40,
min_count=3,
window_size=25,
seed=9102,
num_runs=1,
word_feat="speed_str")
speed_embedding = df_list[0].reset_index(drop=True)
total_embedding = pd.merge(total_embedding,
speed_embedding,
on="boat_id",
how="left")
print("-----------------------------\n")
# Step 3: Training the direcntion embedding
if dir_embedding:
print("\n@Round 2 direction embedding:")
print("-----------------------------")
df_list, model_list = traj_cbow_embedding(traj_data_corpus,
embedding_size=8,
iters=40,
min_count=3,
window_size=25,
seed=9102,
num_runs=1,
word_feat="direction_str")
dir_embedding = df_list[0].reset_index(drop=True)
total_embedding = pd.merge(total_embedding,
dir_embedding,
on="boat_id",
how="left")
print("-----------------------------\n")
# Step 4: Training the speed-direcntion embedding
if speed_dir_embedding:
print("\n@Round 2 speed_dir embedding:")
print("-----------------------------")
df_list, model_list = traj_cbow_embedding(traj_data_corpus,
embedding_size=12,
iters=70,
min_count=3,
window_size=25,
seed=9102,
num_runs=1,
word_feat="speed_dir_str")
speed_dir_embedding = df_list[0].reset_index(drop=True)
total_embedding = pd.merge(total_embedding,
speed_dir_embedding,
on="boat_id",
how="left")
print("-----------------------------")
# Step 5: Svaing the embedding vectorss
train_embedding = total_embedding.iloc[:train_nums].reset_index(drop=True)
test_embedding = total_embedding.iloc[train_nums:].reset_index(drop=True)
file_processor = LoadSave()
file_processor.save_data(path=".//tcdata_tmp//train_signal_embedding.pkl",
data=train_embedding)
file_processor.save_data(path=".//tcdata_tmp//test_signal_embedding.pkl",
data=test_embedding)
def load_pkl(file_name=None):
"""Loading *.pkl from the path .//cached_data//"""
file_processor = LoadSave()
return file_processor.load_data(
path=".//cached_data//{}".format(file_name))
"""Load the original *.csv data."""
total_name = path_name + file_name
csv_data = pd.read_csv(total_name, nrows=nrows)
return csv_data
if __name__ == "__main__":
train_df = load_csv(file_name="train.csv",
path_name=".//data//",
nrows=None)
test_df = load_csv(path_name=".//data//",
file_name="evaluation_public.csv",
nrows=None)
total_df = pd.concat([train_df, test_df], axis=0)
# Encoding category variables
# --------------------------------
cat_list = ["STATUS", "QUEUE_TYPE", "PLATFORM", "RESOURCE_TYPE"]
for name in cat_list:
total_df[name] = total_df[name].astype("category").cat.codes
total_df.sort_values(by=["QUEUE_ID", "DOTTING_TIME"],
ascending=True,
inplace=True)
total_df.reset_index(drop=True, inplace=True)
# Save data to local path
# --------------------------------
file_processor = LoadSave()
file_processor.save_data(path=".//cached_data//total_df.pkl",
data=total_df)
def stat_feature_engineering_xgb():
train_data = load_data("train_semantic.pkl")
test_data_a = load_data("test_semantic.pkl")
train_nums = len(train_data)
total_data = train_data + test_data_a
boat_id = [traj["boat_id"].unique()[0] for traj in total_data]
labels = [traj["type"].unique()[0] for traj in train_data]
total_features = pd.DataFrame(None)
total_features["boat_id"] = boat_id
# Step 1: coordinate stat features.
with mp.Pool(processes=mp.cpu_count()) as p:
tmp = list(
tqdm(p.imap(traj_coord_stat, total_data), total=len(total_data)))
coord_features = pd.concat(tmp, axis=0, ignore_index=True)
coord_features["boat_id"] = boat_id
total_features = pd.merge(total_features,
coord_features,
on="boat_id",
how="left")
# Step 2: speed stat features.
with mp.Pool(processes=mp.cpu_count()) as p:
tmp = list(
tqdm(p.imap(traj_speed_stat, total_data), total=len(total_data)))
speed_features = pd.concat(tmp, axis=0, ignore_index=True)
speed_features["boat_id"] = boat_id
total_features = pd.merge(total_features,
speed_features,
on="boat_id",
how="left")
# Step 4: expert features.
with mp.Pool(processes=mp.cpu_count()) as p:
tmp = list(tqdm(p.imap(traj_expert, total_data),
total=len(total_data)))
expert_features = pd.concat(tmp, axis=0, ignore_index=True)
expert_features["boat_id"] = boat_id
total_features = pd.merge(total_features,
expert_features,
on="boat_id",
how="left")
# Step 5: Concat the speed_dir embedding vector
dir_embedding = traj_data_direction_embedding(total_data,
embedding_size=8,
iters=70,
window_size=20,
min_count=3)
total_features = pd.merge(total_features,
dir_embedding,
on="boat_id",
how="left")
speed_embedding = traj_data_speed_embedding(total_data,
embedding_size=10,
iters=70,
window_size=20,
min_count=3)
total_features = pd.merge(total_features,
speed_embedding,
on="boat_id",
how="left")
# # Step 7: speed tfidf
# speed_tfidf = traj_data_speed_tfidf(total_data, max_features=40)
# total_features = pd.merge(total_features, speed_tfidf, on="boat_id",
# how="left")
# # Step 8: GEO tfidf
# bin_tfidf = traj_data_bin_tfidf(total_data, max_features=70)
# total_features = pd.merge(total_features, bin_tfidf, on="boat_id",
# how="left")
##################################################
train_feature = total_features.iloc[:train_nums].reset_index(
drop=True).copy()
test_feature = total_features.iloc[train_nums:].reset_index(
drop=True).copy()
train_feature["target"] = labels
print("\n--Train samples: {}, testA samples: {}.".format(
len(train_feature), len(test_feature)))
print("--Train cols: {}, test cols: {}.".format(train_feature.shape[1],
test_feature.shape[1]))
print("--Unique train cols: {}, unique testA cols: {}.\n".format(
len(np.unique(train_feature.columns)),
len(np.unique(test_feature.columns))))
file_processor = LoadSave()
file_processor.save_data(path=".//tcdata_tmp//train_feature_xgb.pkl",
data=train_feature)
file_processor.save_data(path=".//tcdata_tmp//train_target.pkl",
data=train_feature[["boat_id", "target"]])
file_processor.save_data(path=".//tcdata_tmp//test_feature_xgb.pkl",
data=test_feature)
gc.collect()
def load_data(name=None):
"""Loading *.pkl data from .//tcdata_tmp//"""
assert name is not None, "Invalid file name!"
file_processor = LoadSave()
return file_processor.load_data(path=".//tcdata_tmp//{}".format(name))
search_res, acc_list = {}, []
print("\n[INFO] DATASET NAME: {}".format(name))
for ts_ind in tqdm(selected_ts_ind):
ts_query = data_norm[ts_ind]
search_res[ts_ind] = search_top_n_similar_ts(ts_query,
data_norm,
n=KEEP_TOP_N,
use_lb_kim=USE_LB_KIM)
if CHECK_1NN_ACC:
one_nn_label = data_label[search_res[ts_ind]
["top_n_searching_res"][0][1]]
one_nn_dist = search_res[ts_ind]["top_n_searching_res"][0][0]
true_label = data_label[ts_ind]
acc_list.append(one_nn_label == true_label)
# STEP 3: Save the SEARCH_TOP_K results in experiment_res
experiment_total_res[name] = search_res
if CHECK_1NN_ACC:
print("\n[INFO] Mean 1-NN accuracy: {:.5f}.".format(
np.mean(acc_list)))
if SAVE_EXPERIMENT_RESULTS:
file_processor = LoadSave()
new_file_path = ".//data_tmp//{}_baseline_top_{}.pkl".format(
TARGET_DATASET_NAME, KEEP_TOP_N)
print("\n")
file_processor.save_data(path=new_file_path, data=experiment_total_res)
