def run(model):
app = QtGui.QApplication(sys.argv)
app.setStyle(QtGui.QStyleFactory.create('Cleanlooks'))
sol = ModelSolver(model)
print "free:", list(sol.get_freevars())
root = get_control_for_model(model)(model, sol)
root.build(None)
root.widget.show()
sol.invoke_observers()
model.invoke_observers()
return app.exec_()
def main():
parse = argparse.ArgumentParser()
# ---------- environment setting: which gpu -------
parse.add_argument('-gpu',
'--gpu',
type=str,
default='0',
help='which gpu to use: 0 or 1')
parse.add_argument('-folder_name',
'--folder_name',
type=str,
default='datasets/citibike-data/data/')
parse.add_argument('-output_folder_name',
'--output_folder_name',
type=str,
default='output/citibike-data/data/')
# ---------- input/output settings -------
parse.add_argument('-input_steps',
'--input_steps',
type=int,
default=6,
help='number of input steps')
# ---------- model ----------
parse.add_argument('-model',
'--model',
type=str,
default='GCN',
help='model: DyST, GCN, AttGCN')
parse.add_argument('-num_layers',
'--num_layers',
type=int,
default=2,
help='number of layers in model')
parse.add_argument('-num_units',
'--num_units',
type=int,
default=64,
help='dim of hidden states')
parse.add_argument('-trained_adj_mx',
'--trained_adj_mx',
type=int,
default=0,
help='if training adjacent matrix')
parse.add_argument('-filter_type',
'--filter_type',
type=str,
default='dual_random_walk',
help='laplacian, random_walk, or dual_random_walk')
parse.add_argument('-delta',
'--delta',
type=int,
default=1e7,
help='delta to calculate rescaled weighted matrix')
parse.add_argument('-epsilon',
'--epsilon',
type=float,
default=0.8,
help='epsilon to calculate rescaled weighted matrix')
#
parse.add_argument(
'-dy_temporal',
'--dy_temporal',
type=int,
default=0,
help='whether to use temporal attention module before output layer')
parse.add_argument(
'-multi_loss',
'--multi_loss',
type=int,
default=0,
help='whether to only consider last prediction into loss function.')
parse.add_argument('-att_units',
'--att_units',
type=int,
default=64,
help='dim of hidden states')
#
parse.add_argument(
'-dy_adj',
'--dy_adj',
type=int,
default=1,
help=
'whether to use dynamic adjacent matrix for lower feature extraction layer'
)
parse.add_argument(
'-dy_filter',
'--dy_filter',
type=int,
default=0,
help='whether to use dynamic filter generate region-specific filter ')
#parse.add_argument('-att_dynamic_adj', '--att_dynamic_adj', type=int, default=1, help='whether to use dynamic adjacent matrix in attention parts')
parse.add_argument('-model_save',
'--model_save',
type=str,
default='gcn',
help='folder name to save model')
parse.add_argument('-pretrained_model',
'--pretrained_model_path',
type=str,
default=None,
help='path to the pretrained model')
# ---------- params for CNN ------------
parse.add_argument('-num_filters',
'--num_filters',
type=int,
default=32,
help='number of filters in CNN')
parse.add_argument('-pooling_units',
'--pooling_units',
type=int,
default=64,
help='number of pooling units')
parse.add_argument('-dropout_keep_prob',
'--dropout_keep_prob',
type=float,
default=0.5,
help='keep probability in dropout layer')
# ---------- training parameters --------
parse.add_argument('-n_epochs',
'--n_epochs',
type=int,
default=20,
help='number of epochs')
parse.add_argument('-batch_size',
'--batch_size',
type=int,
default=8,
help='batch size for training')
parse.add_argument('-show_batches',
'--show_batches',
type=int,
default=100,
help='show how many batches have been processed.')
parse.add_argument('-lr',
'--learning_rate',
type=float,
default=0.0002,
help='learning rate')
parse.add_argument('-update_rule',
'--update_rule',
type=str,
default='adam',
help='update rule')
# ---------- train or predict -------
parse.add_argument('-train',
'--train',
type=int,
default=1,
help='whether to train')
parse.add_argument('-test', '--test', type=int, default=0, help='if test')
#
args = parse.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
print('load train, test data...')
# train: 20140401 - 20140831
# validate: 20140901 - 20140910
# test: 20140911 - 20140930
split = [3672, 240, 480]
#split = [3912, 480]
data, train_data, val_data, test_data = load_npy_data(filename=[
args.folder_name + 'd_station.npy', args.folder_name + 'p_station.npy'
],
split=split)
# data: [num, station_num, 2]
#f_data, train_f_data, val_f_data, test_f_data = load_pkl_data(args.folder_name + 'f_data_list.pkl', split=split)
f_data, train_f_data, val_f_data, test_f_data = load_npy_data(
filename=[args.folder_name + 'citibike_flow_data.npy'], split=split)
print(len(f_data))
print('preprocess train/val/test flow data...')
#f_preprocessing = StandardScaler()
f_preprocessing = MinMaxNormalization01()
f_preprocessing.fit(train_f_data)
train_f_data = f_preprocessing.transform(train_f_data)
if val_f_data is not None:
val_f_data = f_preprocessing.transform(val_f_data)
test_f_data = f_preprocessing.transform(test_f_data)
print('preprocess train/val/test data...')
pre_process = MinMaxNormalization01()
#pre_process = StandardScaler()
pre_process.fit(train_data)
train_data = pre_process.transform(train_data)
if val_data is not None:
val_data = pre_process.transform(val_data)
test_data = pre_process.transform(test_data)
#
num_station = data.shape[1]
print('number of station: %d' % num_station)
#
train_loader = DataLoader_graph(train_data,
train_f_data,
args.input_steps,
flow_format='identity')
if val_data is not None:
val_loader = DataLoader_graph(val_data,
val_f_data,
args.input_steps,
flow_format='identity')
else:
val_loader = None
test_loader = DataLoader_graph(test_data,
test_f_data,
args.input_steps,
flow_format='identity')
# f_adj_mx = None
if os.path.isfile(args.folder_name + 'f_adj_mx.npy'):
f_adj_mx = np.load(args.folder_name + 'f_adj_mx.npy')
else:
f_adj_mx = train_loader.get_flow_adj_mx()
np.save(args.folder_name + 'f_adj_mx.npy', f_adj_mx)
#
#
if args.filter_type == 'laplacian':
w = np.load(args.folder_name + 'w.npy')
# w = np.array(w, dtype=np.float32)
W = get_rescaled_W(w, delta=args.delta, epsilon=args.epsilon)
# Calculate graph kernel
L = scaled_laplacian(W)
#
f_adj_mx = L
if args.model == 'FC_LSTM':
model = FC_LSTM(num_station,
args.input_steps,
num_layers=args.num_layers,
num_units=args.num_units,
batch_size=args.batch_size)
if args.model == 'FC_GRU':
model = FC_GRU(num_station,
args.input_steps,
num_layers=args.num_layers,
num_units=args.num_units,
batch_size=args.batch_size)
if args.model == 'GCN':
model = GCN(num_station,
args.input_steps,
num_layers=args.num_layers,
num_units=args.num_units,
dy_adj=args.dy_adj,
dy_filter=args.dy_filter,
f_adj_mx=f_adj_mx,
trained_adj_mx=args.trained_adj_mx,
filter_type=args.filter_type,
batch_size=args.batch_size)
if args.model == 'flow_GCN':
model = flow_GCN(num_station,
args.input_steps,
num_layers=args.num_layers,
num_units=args.num_units,
f_adj_mx=f_adj_mx,
trained_adj_mx=args.trained_adj_mx,
filter_type=args.filter_type,
batch_size=args.batch_size)
if args.model == 'Coupled_GCN':
model = Coupled_GCN(num_station,
args.input_steps,
num_layers=args.num_layers,
num_units=args.num_units,
f_adj_mx=f_adj_mx,
trained_adj_mx=args.trained_adj_mx,
filter_type=args.filter_type,
dy_temporal=args.dy_temporal,
att_units=args.att_units,
multi_loss=args.multi_loss,
batch_size=args.batch_size)
#
model_path = os.path.join(args.output_folder_name, 'model_save',
args.model_save)
if not os.path.exists(model_path):
os.makedirs(model_path)
#model_path = os.path.join(args.folder_name, 'model_save', args.model_save)
solver = ModelSolver(
model,
train_loader,
val_loader,
test_loader,
pre_process,
batch_size=args.batch_size,
show_batches=args.show_batches,
n_epochs=args.n_epochs,
pretrained_model=args.pretrained_model_path,
update_rule=args.update_rule,
learning_rate=args.learning_rate,
model_path=model_path,
)
results_path = os.path.join(model_path, 'results')
if not os.path.exists(results_path):
os.makedirs(results_path)
if args.train:
print('==================== begin training ======================')
test_target, test_prediction = solver.train(
os.path.join(model_path, 'out'))
np.save(os.path.join(results_path, 'test_target.npy'), test_target)
np.save(os.path.join(results_path, 'test_prediction.npy'),
test_prediction)
if args.test:
print('==================== begin test ==========================')
test_target, test_prediction = solver.test()
np.save(os.path.join(results_path, 'test_target.npy'), test_target)
np.save(os.path.join(results_path, 'test_prediction.npy'),
test_prediction)
def main():
parse = argparse.ArgumentParser()
# ---------- environment setting: which gpu -------
parse.add_argument('-gpu',
'--gpu',
type=str,
default='0',
help='which gpu to use: 0 or 1')
parse.add_argument('-folder_name',
'--folder_name',
type=str,
default='datasets/citibike-data/data/')
parse.add_argument(
'-if_minus_mean',
'--if_minus_mean',
type=int,
default=0,
help='use MinMaxNormalize01 or MinMaxNormalize01_minus_mean')
# ---------- input/output settings -------
parse.add_argument('-input_steps',
'--input_steps',
type=int,
default=6,
help='number of input steps')
parse.add_argument('-output_steps',
'--output_steps',
type=int,
default=1,
help='number of output steps')
# ---------- station embeddings --------
parse.add_argument('-pretrained_embeddings',
'--pretrained_embeddings',
type=int,
default=1,
help='whether to use pretrained embeddings')
parse.add_argument('-embedding_size',
'--embedding_size',
type=int,
default=100,
help='dim of embedding')
# ---------- model ----------
#parse.add_argument('-model', '--model', type=str, default='DyST', help='model: NN, LSTM, biLSTM, CNN')
parse.add_argument('-dynamic_context',
'--dynamic_context',
type=int,
default=1,
help='whether to add dynamic_context part')
parse.add_argument('-dynamic_spatial',
'--dynamic_spatial',
type=int,
default=1,
help='whether to add dynamic_spatial part')
parse.add_argument('-add_ext',
'--add_ext',
type=int,
default=1,
help='whether to add external factors')
parse.add_argument('-model_save',
'--model_save',
type=str,
default='',
help='folder name to save model')
parse.add_argument('-pretrained_model',
'--pretrained_model_path',
type=str,
default=None,
help='path to the pretrained model')
# ---------- params for CNN ------------
parse.add_argument('-num_filters',
'--num_filters',
type=int,
default=32,
help='number of filters in CNN')
parse.add_argument('-pooling_units',
'--pooling_units',
type=int,
default=64,
help='number of pooling units')
parse.add_argument('-dropout_keep_prob',
'--dropout_keep_prob',
type=float,
default=0.5,
help='keep probability in dropout layer')
# ---------- training parameters --------
parse.add_argument('-n_epochs',
'--n_epochs',
type=int,
default=50,
help='number of epochs')
parse.add_argument('-batch_size',
'--batch_size',
type=int,
default=8,
help='batch size for training')
parse.add_argument('-show_batches',
'--show_batches',
type=int,
default=100,
help='show how many batches have been processed.')
parse.add_argument('-lr',
'--learning_rate',
type=float,
default=0.002,
help='learning rate')
parse.add_argument('-update_rule',
'--update_rule',
type=str,
default='adam',
help='update rule')
# ------ train or predict -------
parse.add_argument('-train',
'--train',
type=int,
default=1,
help='whether to train')
parse.add_argument('-test', '--test', type=int, default=0, help='if test')
#
parse.add_argument('-pretrain',
'--pretrain',
type=int,
default=0,
help='whether to pretrain')
parse.add_argument('-partial_pretrain',
'--partial_pretrain',
type=int,
default=0,
help='whether to load pretrained vars')
args = parse.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
print('load train, test data...')
# train: 20140401 - 20140910
# test: 20140911 - 20140930
split = [3912, 480]
data, train_data, test_data, _ = load_npy_data(filename=[
args.folder_name + 'd_station.npy', args.folder_name + 'p_station.npy'
],
split=split)
# data: [num, station_num, 2]
f_data, train_f_data, test_f_data, _ = load_pkl_data(args.folder_name +
'f_data_list.pkl',
split=split)
print(len(f_data))
# e_data: [num, ext_dim]
e_data, train_e_data, test_e_data, _ = load_mat_data(args.folder_name +
'fea2.mat',
'fea',
split=split)
# e_preprocess = MinMaxNormalization01()
# e_preprocess.fit(train_e_data)
# train_e_data = e_preprocess.transform(train_e_data)
# test_e_data = e_preprocess.transform(test_e_data)
print('preprocess train/test data...')
#pre_process = MinMaxNormalization01_by_axis()
if args.if_minus_mean:
pre_process = MinMaxNormalization01_minus_mean()
pre_process.fit(train_data)
norm_mean_data = pre_process.transform(data)
train_data = norm_mean_data[:split[0]]
test_data = norm_mean_data[split[0]:]
else:
pre_process = MinMaxNormalization01()
pre_process.fit(train_data)
train_data = pre_process.transform(train_data)
test_data = pre_process.transform(test_data)
# embeddings
#id_map = load_pickle(args.folder_name+'station_map.pkl')
#num_station = len(id_map)
num_station = data.shape[1]
print('number of station: %d' % num_station)
if args.pretrained_embeddings:
print('load pretrained embeddings...')
embeddings = get_embedding_from_file(
args.folder_name + 'embeddings.txt', num_station)
else:
print('train station embeddings via Word2Vec model...')
trip_data = load_pickle(args.folder_name + 'all_trip_data.pkl')
word2vec_model = Word2Vec(sentences=trip_data,
size=args.embedding_size)
print('save Word2Vec model and embeddings...')
word2vec_model.save(args.folder_name + 'word2vec_model')
word2vec_model.wv.save_word2vec_format(args.folder_name +
'embeddings.txt',
binary=False)
del word2vec_model
embeddings = get_embedding_from_file(
args.folder_name + 'embeddings.txt', num_station)
train_loader = DataLoader(train_data, train_f_data, train_e_data,
args.input_steps, args.output_steps, num_station)
# val_loader = DataLoader(val_data, val_f_data,
# args.input_steps, args.output_steps,
# num_station, pre_process)
test_loader = DataLoader(test_data, test_f_data, test_e_data,
args.input_steps, args.output_steps, num_station)
model = DyST2(num_station,
args.input_steps,
args.output_steps,
embedding_dim=args.embedding_size,
embeddings=embeddings,
ext_dim=e_data.shape[-1],
batch_size=args.batch_size,
dynamic_context=args.dynamic_context,
dynamic_spatial=args.dynamic_spatial,
add_ext=args.add_ext)
model_path = os.path.join(args.folder_name, 'model_save', args.model_save)
solver = ModelSolver(model,
train_loader,
test_loader,
pre_process,
batch_size=args.batch_size,
show_batches=args.show_batches,
n_epochs=args.n_epochs,
pretrained_model=args.pretrained_model_path,
update_rule=args.update_rule,
learning_rate=args.learning_rate,
model_path=model_path,
partial_pretrain=args.partial_pretrain)
results_path = os.path.join(model_path, 'results')
if not os.path.exists(results_path):
os.makedirs(results_path)
if args.pretrain:
print '==================== begin pretrain ======================'
w_att_1, w_att_2, w_h_in, w_h_out = solver.pretrain(
os.path.join(model_path, 'pretrain_out'))
np.save(os.path.join(model_path, 'w_att_1.npy'), w_att_1)
np.save(os.path.join(model_path, 'w_att_2.npy'), w_att_2)
np.save(os.path.join(model_path, 'w_h_in.npy'), w_h_in)
np.save(os.path.join(model_path, 'w_h_out.npy'), w_h_out)
if args.train:
print '==================== begin training ======================'
test_target, test_prediction = solver.train(
os.path.join(model_path, 'out'))
np.save(os.path.join(results_path, 'test_target.npy'), test_target)
np.save(os.path.join(results_path, 'test_prediction.npy'),
test_prediction)
if args.test:
print '==================== begin test =========================='
test_target, test_prediction = solver.test()
np.save(os.path.join(results_path, 'test_target.npy'), test_target)
np.save(os.path.join(results_path, 'test_prediction.npy'),
test_prediction)
def main():
os.environ['CUDA_VISIBLE_DEVICES'] = FLAGS.gpu
# preprocessing class
pre_process = MinMaxNormalization01()
print('load train, validate, test data...')
split = [43824, 8760, 8760]
data, train_data, val_data, test_data = load_data(filename=['data/taxi/p_map.mat', 'data/taxi/d_map.mat'],
split=split)
# data: [num, row, col, channel]
print('preprocess train data...')
pre_process.fit(train_data)
if 'ResNet' in FLAGS.model:
pre_index = max(FLAGS.closeness * 1, FLAGS.period * 7, FLAGS.trend * 7 * 24)
all_timestamps = gen_timestamps(['2009', '2010', '2011', '2012', '2013', '2014', '2015'])
data = pre_process.transform(data)
# train_data = train_data
train_data = data[:split[0]]
val_data = data[split[0] - pre_index:split[0] + split[1]]
test_data = data[split[0] + split[1] - pre_index:split[0] + split[1] + split[2]]
del data
# get train, validate, test timestamps
train_timestamps = all_timestamps[:split[0]]
val_timestamps = all_timestamps[split[0] - pre_index:split[0] + split[1]]
test_timestamps = all_timestamps[split[0] + split[1] - pre_index:split[0] + split[1] + split[2]]
# get x, y
train_x, train_y = batch_data_cpt_ext(train_data, train_timestamps,
batch_size=FLAGS.batch_size, close=FLAGS.closeness, period=FLAGS.period,
trend=FLAGS.trend)
val_x, val_y = batch_data_cpt_ext(val_data, val_timestamps,
batch_size=FLAGS.batch_size, close=FLAGS.closeness, period=FLAGS.period,
trend=FLAGS.trend)
test_x, test_y = batch_data_cpt_ext(test_data, test_timestamps,
batch_size=FLAGS.batch_size, close=FLAGS.closeness, period=FLAGS.period,
trend=FLAGS.trend)
train = {'x': train_x, 'y': train_y}
val = {'x': val_x, 'y': val_y}
test = {'x': test_x, 'y': test_y}
nb_flow = train_data.shape[-1]
row = train_data.shape[1]
col = train_data.shape[2]
if FLAGS.model == 'AttResNet':
print('k-means to cluster...')
model_path = 'taxi-results/model_save/AttResNet/'
log_path = 'taxi-results/log/AttResNet/'
if FLAGS.pre_saved_cluster:
cluster_centroid = np.load(model_path + 'cluster_centroid.npy')
else:
vector_data = np.reshape(train_data, (train_data.shape[0], -1))
kmeans = KMeans(n_clusters=FLAGS.cluster_num, init='random', n_init=FLAGS.kmeans_run_num,
tol=0.00000001).fit(vector_data)
cluster_centroid = kmeans.cluster_centers_
cluster_centroid = np.reshape(cluster_centroid,
(-1, train_data.shape[1], train_data.shape[2], train_data.shape[3]))
if not os.path.exists(model_path):
os.makedirs(model_path)
if not os.path.exists(log_path):
os.makedirs(log_path)
np.save(model_path + 'cluster_centroid.npy', cluster_centroid)
print('build AttResNet model...')
model = AttResNet(input_conf=[[FLAGS.closeness, nb_flow, row, col], [FLAGS.period, nb_flow, row, col],
[FLAGS.trend, nb_flow, row, col], [8]],
att_inputs=cluster_centroid, att_nodes=FLAGS.att_nodes,
att_layer=['conv', 'conv'],
att_layer_param=[[[3, 3], [1, 1, 1, 1], 8], [[3, 3], [1, 1, 1, 1], 2]],
batch_size=FLAGS.batch_size,
layer=['conv', 'res_net', 'conv'],
layer_param=[[[3, 3], [1, 1, 1, 1], 64],
[3, [[[3, 3], [1, 1, 1, 1], 64], [[3, 3], [1, 1, 1, 1], 64]]],
[[3, 3], [1, 1, 1, 1], 2]]
)
else:
print('build ResNet model...')
model_path = 'taxi-results/model_save/ResNet/'
log_path = 'taxi-results/log/ResNet/'
model = ResNet(input_conf=[[FLAGS.closeness, nb_flow, row, col], [FLAGS.period, nb_flow, row, col],
[FLAGS.trend, nb_flow, row, col], [8]], batch_size=FLAGS.batch_size,
layer=['conv', 'res_net', 'conv'],
layer_param=[[[3, 3], [1, 1, 1, 1], 64],
[3, [[[3, 3], [1, 1, 1, 1], 64], [[3, 3], [1, 1, 1, 1], 64]]],
[[3, 3], [1, 1, 1, 1], 2]]
)
print('model solver...')
solver = ModelSolver(model, train, val, preprocessing=pre_process,
n_epochs=FLAGS.n_epochs,
batch_size=FLAGS.batch_size,
update_rule=FLAGS.update_rule,
learning_rate=FLAGS.lr, save_every=FLAGS.save_every,
pretrained_model=FLAGS.pretrained_model, model_path=model_path,
test_model='taxi-results/model_save/ResNet/model-' + str(FLAGS.n_epochs),
log_path=log_path,
cross_val=False, cpt_ext=True)
if FLAGS.train:
print('begin training...')
test_n = {'data': test_data, 'timestamps': test_timestamps}
_, test_prediction = solver.train(test, test_n, output_steps=FLAGS.output_steps)
# get test_target and test_prediction
i = pre_index
test_target = []
while i < len(test_data) - FLAGS.output_steps:
test_target.append(test_data[i:i + FLAGS.output_steps])
i += 1
test_target = np.asarray(test_target)
if FLAGS.test:
print('begin testing for predicting next 1 step')
solver.test(test)
print('begin testing for predicting next' + str(FLAGS.output_steps) + 'steps')
test_n = {'data': test_data, 'timestamps': test_timestamps}
solver.test_1_to_n(test_n)
else:
train_data = pre_process.transform(train_data)
train_x, train_y = batch_data(data=train_data, batch_size=FLAGS.batch_size,
input_steps=FLAGS.input_steps, output_steps=FLAGS.output_steps)
val_data = pre_process.transform(val_data)
val_x, val_y = batch_data(data=val_data, batch_size=FLAGS.batch_size,
input_steps=FLAGS.input_steps, output_steps=FLAGS.output_steps)
test_data = pre_process.transform(test_data)
test_x, test_y = batch_data(data=test_data, batch_size=FLAGS.batch_size,
input_steps=FLAGS.input_steps, output_steps=FLAGS.output_steps)
train = {'x': train_x, 'y': train_y}
val = {'x': val_x, 'y': val_y}
test = {'x': test_x, 'y': test_y}
input_dim = [train_data.shape[1], train_data.shape[2], train_data.shape[3]]
if FLAGS.model == 'ConvLSTM':
print('build ConvLSTM model...')
model = ConvLSTM(input_dim=input_dim, batch_size=FLAGS.batch_size,
layer={'encoder': ['conv', 'conv', 'conv_lstm', 'conv_lstm'],
'decoder': ['conv_lstm', 'conv_lstm', 'conv', 'conv']},
layer_param={'encoder': [[[3, 3], [1, 2, 2, 1], 8],
[[3, 3], [1, 2, 2, 1], 16],
[[16, 16], [3, 3], 64],
[[16, 16], [3, 3], 64]],
'decoder': [[[16, 16], [3, 3], 64],
[[16, 16], [3, 3], 64],
[[3, 3], [1, 2, 2, 1], 8],
[[3, 3], [1, 2, 2, 1], 2]]},
input_steps=10, output_steps=10)
print('model solver...')
solver = ModelSolver(model, train, val, preprocessing=pre_process,
n_epochs=FLAGS.n_epochs,
batch_size=FLAGS.batch_size,
update_rule=FLAGS.update_rule,
learning_rate=FLAGS.lr, save_every=FLAGS.save_every,
pretrained_model=FLAGS.pretrained_model,
model_path='taxi-results/model_save/ConvLSTM/',
test_model='taxi-results/model_save/ConvLSTM/model-' + str(FLAGS.n_epochs),
log_path='taxi-results/log/ConvLSTM/')
elif 'AttConvLSTM' in FLAGS.model:
# train_data: [num, row, col, channel]
if FLAGS.use_ae:
# auto-encoder to cluster train_data
print('auto-encoder to cluster...')
model_path = 'taxi-results/model_save/AEAttConvLSTM/'
log_path = 'taxi-results/log/AEAttConvLSTM/'
if FLAGS.pre_saved_cluster:
cluster_centroid = np.load(model_path + 'cluster_centroid.npy')
else:
ae = AutoEncoder(input_dim=input_dim, z_dim=[16, 16, 16],
layer={'encoder': ['conv', 'conv'],
'decoder': ['conv', 'conv']},
layer_param={'encoder': [[[3, 3], [1, 2, 2, 1], 8],
[[3, 3], [1, 2, 2, 1], 16]],
'decoder': [[[3, 3], [1, 2, 2, 1], 8],
[[3, 3], [1, 2, 2, 1], 2]]},
model_save_path=model_path,
batch_size=FLAGS.batch_size)
if FLAGS.ae_train:
ae.train(train_data, batch_size=FLAGS.batch_size, learning_rate=FLAGS.lr, n_epochs=20,
pretrained_model=FLAGS.ae_pretrained_model)
train_z_data = ae.get_z(train_data, pretrained_model=FLAGS.ae_pretrained_model)
print(train_z_data.shape)
# k-means to cluster train_z_data
vector_data = np.reshape(train_z_data, (train_z_data.shape[0], -1))
kmeans = KMeans(n_clusters=FLAGS.cluster_num, init='random', n_init=FLAGS.kmeans_run_num,
tol=0.00000001).fit(vector_data)
cluster_centroid = kmeans.cluster_centers_
print(np.array(cluster_centroid).shape)
# reshape to [cluster_num, row, col, channel]
cluster_centroid = np.reshape(cluster_centroid,
(-1, train_z_data.shape[1], train_z_data.shape[2],
train_z_data.shape[3]))
# decoder to original space
cluster_centroid = ae.get_y(cluster_centroid, pretrained_model=FLAGS.ae_pretrained_model)
print(cluster_centroid.shape)
np.save(model_path + 'cluster_centroid.npy', cluster_centroid)
else:
# k-means to cluster train_data
print('k-means to cluster...')
model_path = 'taxi-results/model_save/' + FLAGS.model + '/'
log_path = 'taxi-results/log/' + FLAGS.model + '/'
if not os.path.exists(model_path):
os.makedirs(model_path)
if not os.path.exists(log_path):
os.makedirs(log_path)
if FLAGS.pre_saved_cluster:
cluster_centroid = np.load(model_path + 'cluster_centroid.npy')
else:
vector_data = np.reshape(train_data, (train_data.shape[0], -1))
# init_vectors = vector_data[:FLAGS.cluster_num, :]
# cluster_centroid = init_vectors
kmeans = KMeans(n_clusters=FLAGS.cluster_num, init='random', n_init=FLAGS.kmeans_run_num,
tol=0.00000001).fit(vector_data)
cluster_centroid = kmeans.cluster_centers_
# reshape to [cluster_num, row, col, channel]
cluster_centroid = np.reshape(cluster_centroid,
(-1, train_data.shape[1], train_data.shape[2], train_data.shape[3]))
np.save(model_path + 'cluster_centroid.npy', cluster_centroid)
# build model
print('build ' + FLAGS.model + ' model...')
if FLAGS.model == 'AttConvLSTM':
model = AttConvLSTM(input_dim=input_dim,
att_inputs=cluster_centroid, att_nodes=FLAGS.att_nodes,
batch_size=FLAGS.batch_size,
layer={'encoder': ['conv', 'conv', 'conv_lstm', 'conv_lstm'],
'decoder': ['conv_lstm', 'conv_lstm', 'conv', 'conv'],
'attention': ['conv', 'conv']},
layer_param={'encoder': [[[3, 3], [1, 2, 2, 1], 8],
[[3, 3], [1, 2, 2, 1], 16],
[[16, 16], [3, 3], 64],
[[16, 16], [3, 3], 64]],
'decoder': [[[16, 16], [3, 3], 64],
[[16, 16], [3, 3], 64],
[[3, 3], [1, 2, 2, 1], 8],
[[3, 3], [1, 2, 2, 1], 2]],
'attention': [[[3, 3], [1, 2, 2, 1], 8],
[[3, 3], [1, 2, 2, 1], 16]]},
input_steps=10, output_steps=10)
elif FLAGS.model == 'MultiAttConvLSTM':
model = MultiAttConvLSTM(input_dim=input_dim,
att_inputs=cluster_centroid, att_nodes=FLAGS.att_nodes,
batch_size=FLAGS.batch_size,
layer={'encoder': ['conv', 'conv', 'conv_lstm', 'conv_lstm'],
'decoder': ['conv_lstm', 'conv_lstm', 'conv', 'conv'],
'attention': ['conv', 'conv']},
layer_param={'encoder': [[[3, 3], [1, 2, 2, 1], 8],
[[3, 3], [1, 2, 2, 1], 16],
[[16, 16], [3, 3], 64],
[[16, 16], [3, 3], 64]],
'decoder': [[[16, 16], [3, 3], 64],
[[16, 16], [3, 3], 64],
[[3, 3], [1, 2, 2, 1], 8],
[[3, 3], [1, 2, 2, 1], 2]],
'attention': [[[3, 3], [1, 2, 2, 1], 8],
[[3, 3], [1, 2, 2, 1], 16]]},
input_steps=10, output_steps=10)
print('model solver...')
solver = ModelSolver(model, train, val, preprocessing=pre_process,
n_epochs=FLAGS.n_epochs,
batch_size=FLAGS.batch_size,
update_rule=FLAGS.update_rule,
learning_rate=FLAGS.lr, save_every=FLAGS.save_every,
pretrained_model=FLAGS.pretrained_model, model_path=model_path,
test_model=model_path + 'model-' + str(FLAGS.n_epochs), log_path=log_path)
if FLAGS.train:
print('begin training...')
test_prediction, _ = solver.train(test)
test_target = np.asarray(test_y)
if FLAGS.test:
print('test trained model...')
solver.test_model = solver.model_path + FLAGS.pretrained_model
test_prediction = solver.test(test)
test_target = np.asarray(test_y)
np.save('taxi-results/results/'+FLAGS.model+'/test_target.npy', test_target)
np.save('taxi-results/results/'+FLAGS.model+'/test_prediction.npy', test_prediction)
print(test_prediction.shape)
def main():
parse = argparse.ArgumentParser()
# ---------- environment setting: which gpu -------
parse.add_argument('-gpu',
'--gpu',
type=str,
default='0',
help='which gpu to use: 0 or 1')
parse.add_argument('-folder_name',
'--folder_name',
type=str,
default='datasets/didi-data/data/')
parse.add_argument('-output_folder_name',
'--output_folder_name',
type=str,
default='output/didi-data/data/')
# ---------- input/output settings -------
parse.add_argument('-input_steps',
'--input_steps',
type=int,
default=6,
help='number of input steps')
# ---------- model ----------
parse.add_argument('-model',
'--model',
type=str,
default='GCN',
help='model: GCN, ConvLSTM, flow_ConvLSTM')
parse.add_argument('-num_layers',
'--num_layers',
type=int,
default=2,
help='number of layers in model')
parse.add_argument('-num_units',
'--num_units',
type=int,
default=64,
help='dim of hidden states')
parse.add_argument('-kernel_size',
'--kernel_size',
type=int,
default=3,
help='kernel size in convolutional operations')
#
parse.add_argument(
'-dy_adj',
'--dy_adj',
type=int,
default=1,
help=
'whether to use dynamic adjacent matrix for lower feature extraction layer'
)
parse.add_argument(
'-dy_filter',
'--dy_filter',
type=int,
default=0,
help='whether to use dynamic filter generate region-specific filter ')
parse.add_argument(
'-att_dynamic_adj',
'--att_dynamic_adj',
type=int,
default=0,
help='whether to use dynamic adjacent matrix in attention parts')
#
parse.add_argument('-model_save',
'--model_save',
type=str,
default='gcn',
help='folder name to save model')
parse.add_argument('-pretrained_model',
'--pretrained_model_path',
type=str,
default=None,
help='path to the pretrained model')
# ---------- params for CNN ------------
parse.add_argument('-num_filters',
'--num_filters',
type=int,
default=32,
help='number of filters in CNN')
parse.add_argument('-pooling_units',
'--pooling_units',
type=int,
default=64,
help='number of pooling units')
parse.add_argument('-dropout_keep_prob',
'--dropout_keep_prob',
type=float,
default=0.5,
help='keep probability in dropout layer')
# ---------- training parameters --------
parse.add_argument('-n_epochs',
'--n_epochs',
type=int,
default=20,
help='number of epochs')
parse.add_argument('-batch_size',
'--batch_size',
type=int,
default=8,
help='batch size for training')
parse.add_argument('-show_batches',
'--show_batches',
type=int,
default=100,
help='show how many batches have been processed.')
parse.add_argument('-lr',
'--learning_rate',
type=float,
default=0.0002,
help='learning rate')
parse.add_argument('-update_rule',
'--update_rule',
type=str,
default='adam',
help='update rule')
# ---------- train or predict -------
parse.add_argument('-train',
'--train',
type=int,
default=1,
help='whether to train')
parse.add_argument('-test', '--test', type=int, default=0, help='if test')
#
args = parse.parse_args()
os.environ['CUDA_VISIBLE_DEVICES'] = args.gpu
print('load train, test data...')
#
# train: 20161101 - 20161125
# validate: 20161126 - 20161127
# test: 20161128 - 20161130
split = [2400, 192, 288]
data, train_data, val_data, test_data = load_npy_data(
filename=[args.folder_name + 'cd_didi_data.npy'], split=split)
# data: [num, station_num, 2]
print(data.shape)
#
if 'GCN' in args.model or 'FC' in args.model:
dataloader = DataLoader_graph
else:
data = np.reshape(data, (-1, 20, 20, 2))
train_data = np.reshape(train_data, (-1, 20, 20, 2))
val_data = np.reshape(val_data, (-1, 20, 20, 2))
test_data = np.reshape(test_data, (-1, 20, 20, 2))
# data: [num, height, width, 2]
print(data.shape)
#
dataloader = DataLoader_map
#
map_size = data.shape[1:-1]
input_dim = data.shape[-1]
num_station = np.prod(data.shape[1:-1])
#
f_data, train_f_data, val_f_data, test_f_data = load_npy_data(
[args.folder_name + 'cd_didi_flow_in.npy'], split=split)
print(len(f_data))
print('preprocess train/val/test flow data...')
#f_preprocessing = StandardScaler()
f_preprocessing = MinMaxNormalization01()
f_preprocessing.fit(train_f_data)
train_f_data = f_preprocessing.transform(train_f_data)
val_f_data = f_preprocessing.transform(val_f_data)
test_f_data = f_preprocessing.transform(test_f_data)
print('preprocess train/val/test data...')
# pre_process = StandardScaler()
pre_process = MinMaxNormalization01()
pre_process.fit(train_data)
train_data = pre_process.transform(train_data)
val_data = pre_process.transform(val_data)
test_data = pre_process.transform(test_data)
#
print('number of station: %d' % num_station)
#
train_loader = dataloader(train_data,
train_f_data,
args.input_steps,
flow_format='identity')
val_loader = dataloader(val_data,
val_f_data,
args.input_steps,
flow_format='identity')
test_loader = dataloader(test_data,
test_f_data,
args.input_steps,
flow_format='identity')
# f_adj_mx = None
if os.path.isfile(args.folder_name + 'f_adj_mx.npy'):
f_adj_mx = np.load(args.folder_name + 'f_adj_mx.npy')
else:
f_adj_mx = train_loader.get_flow_adj_mx()
np.save(args.folder_name + 'f_adj_mx.npy', f_adj_mx)
#
# if args.model == 'FC_LSTM':
# model = FC_LSTM(num_station, args.input_steps,
# num_layers=args.num_layers, num_units=args.num_units,
# batch_size=args.batch_size)
if args.model == 'FC_GRU':
model = FC_GRU(num_station,
args.input_steps,
num_layers=args.num_layers,
num_units=args.num_units,
batch_size=args.batch_size)
if args.model == 'GCN':
model = GCN(num_station,
args.input_steps,
num_layers=args.num_layers,
num_units=args.num_units,
dy_adj=args.dy_adj,
dy_filter=args.dy_filter,
f_adj_mx=f_adj_mx,
batch_size=args.batch_size)
if args.model == 'ConvGRU':
model = ConvGRU(input_shape=[map_size[0], map_size[1], input_dim],
input_steps=args.input_steps,
num_layers=args.num_layers,
num_units=args.num_units,
kernel_shape=[args.kernel_size, args.kernel_size],
batch_size=args.batch_size)
# if args.model == 'flow_ConvGRU':
# model = flow_ConvGRU(input_shape=[20, 20, input_dim], input_steps=args.input_steps,
# num_layers=args.num_layers, num_units=args.num_units, kernel_shape=[args.kernel_size, args.kernel_size],
# f_adj_mx=f_adj_mx,
# batch_size=args.batch_size)
if args.model == 'Coupled_ConvGRU':
model = CoupledConvGRU(
input_shape=[20, 20, input_dim],
input_steps=args.input_steps,
num_layers=args.num_layers,
num_units=args.num_units,
kernel_shape=[args.kernel_size, args.kernel_size],
batch_size=args.batch_size)
##
# flow_ConvGRU_2 is Stack_ConvGRU with 2 conv layers.
if args.model == 'flow_ConvGRU_2':
model = flow_ConvGRU_2(
input_shape=[20, 20, input_dim],
input_steps=args.input_steps,
num_layers=args.num_layers,
num_units=args.num_units,
kernel_shape=[args.kernel_size, args.kernel_size],
f_adj_mx=f_adj_mx,
batch_size=args.batch_size)
if args.model == 'Stack_ConvGRU':
model = Stack_ConvGRU(
input_shape=[20, 20, input_dim],
input_steps=args.input_steps,
num_layers=args.num_layers,
num_units=args.num_units,
kernel_shape=[args.kernel_size, args.kernel_size],
f_adj_mx=f_adj_mx,
batch_size=args.batch_size)
#
model_path = os.path.join(args.output_folder_name, 'model_save',
args.model_save)
if not os.path.exists(model_path):
os.makedirs(model_path)
#model_path = os.path.join(args.folder_name, 'model_save', args.model_save)
solver = ModelSolver(
model,
train_loader,
val_loader,
test_loader,
pre_process,
batch_size=args.batch_size,
show_batches=args.show_batches,
n_epochs=args.n_epochs,
pretrained_model=args.pretrained_model_path,
update_rule=args.update_rule,
learning_rate=args.learning_rate,
model_path=model_path,
)
results_path = os.path.join(model_path, 'results')
if not os.path.exists(results_path):
os.makedirs(results_path)
if args.train:
print('==================== begin training ======================')
test_target, test_prediction = solver.train(
os.path.join(model_path, 'out'))
np.save(os.path.join(results_path, 'test_target.npy'), test_target)
np.save(os.path.join(results_path, 'test_prediction.npy'),
test_prediction)
if args.test:
print('==================== begin test ==========================')
test_target, test_prediction = solver.test()
np.save(os.path.join(results_path, 'test_target.npy'), test_target)
np.save(os.path.join(results_path, 'test_prediction.npy'),
test_prediction)
def main():
os.environ['CUDA_VISIBLE_DEVICES'] = FLAGS.gpu
# preprocessing class
pre_process = MinMaxNormalization01()
print('load train, validate, test data...')
split = [17520, 4416, 4368]
data, train_data, val_data, test_data = load_npy_data(
filename=['data/citybike/p_map.npy', 'data/citybike/d_map.npy'],
split=split)
# data: [num, row, col, channel]
print('preprocess train data...')
pre_process.fit(train_data)
train_data = pre_process.transform(train_data)
train_x, train_y = batch_data(data=train_data,
batch_size=FLAGS.batch_size,
input_steps=FLAGS.input_steps,
output_steps=FLAGS.output_steps)
val_data = pre_process.transform(val_data)
val_x, val_y = batch_data(data=val_data,
batch_size=FLAGS.batch_size,
input_steps=FLAGS.input_steps,
output_steps=FLAGS.output_steps)
test_data = pre_process.transform(test_data)
test_x, test_y = batch_data(data=test_data,
batch_size=FLAGS.batch_size,
input_steps=FLAGS.input_steps,
output_steps=FLAGS.output_steps)
train = {'x': train_x, 'y': train_y}
val = {'x': val_x, 'y': val_y}
test = {'x': test_x, 'y': test_y}
input_dim = [train_data.shape[1], train_data.shape[2], train_data.shape[3]]
if 'AttConvLSTM' in FLAGS.model:
# train_data: [num, row, col, channel]
model_path = 'citybike-results/model_save/' + FLAGS.model + '/'
log_path = 'citybike-results/log/' + FLAGS.model + '/'
if not os.path.exists(model_path):
os.makedirs(model_path)
if not os.path.exists(log_path):
os.makedirs(log_path)
if FLAGS.pre_saved_cluster:
cluster_centroid = np.load(model_path + 'cluster_centroid.npy')
else:
vector_data = np.reshape(train_data, (train_data.shape[0], -1))
cluster_centroid_1 = None
cluster_centroid_2 = None
cluster_centroid = None
if FLAGS.kmeans_cluster:
print('k-means to cluster...')
kmeans = KMeans(n_clusters=FLAGS.cluster_num,
init='random',
n_init=FLAGS.kmeans_run_num,
tol=0.00000001).fit(vector_data)
cluster_centroid_1 = kmeans.cluster_centers_
if FLAGS.average_cluster:
print('average cluster...')
if FLAGS.average_cluster == 24:
cluster_centroid_2 = average_cluster_24(vector_data)
elif FLAGS.average_cluster == 48:
cluster_centroid_2 = average_cluster_48(vector_data)
if cluster_centroid_1 is not None:
cluster_centroid = cluster_centroid_1
if cluster_centroid_2 is not None:
if cluster_centroid is not None:
cluster_centroid = np.concatenate(
(cluster_centroid_1, cluster_centroid_2), axis=0)
else:
cluster_centroid = cluster_centroid_2
# reshape to [cluster_num, row, col, channel]
cluster_centroid = np.reshape(
cluster_centroid, (-1, train_data.shape[1],
train_data.shape[2], train_data.shape[3]))
np.save(model_path + 'cluster_centroid.npy', cluster_centroid)
# build model
print 'build ' + FLAGS.model + ' model...'
if FLAGS.model == 'AttConvLSTM':
model = AttConvLSTM(input_dim=input_dim,
att_inputs=cluster_centroid,
att_nodes=FLAGS.att_nodes,
batch_size=FLAGS.batch_size,
layer={
'encoder':
['conv', 'conv', 'conv_lstm', 'conv_lstm'],
'decoder':
['conv_lstm', 'conv_lstm', 'conv', 'conv'],
'attention': ['conv', 'conv']
},
layer_param={
'encoder': [[[3, 3], [1, 1, 1, 1], 8],
[[3, 3], [1, 1, 1, 1], 16],
[[16, 16], [3, 3], 64],
[[16, 16], [3, 3], 64]],
'decoder': [[[16, 16], [3, 3], 64],
[[16, 16], [3, 3], 64],
[[3, 3], [1, 1, 1, 1], 8],
[[3, 3], [1, 1, 1, 1], 2]],
'attention': [[[3, 3], [1, 1, 1, 1], 8],
[[3, 3], [1, 1, 1, 1], 16]]
},
input_steps=10,
output_steps=10)
elif FLAGS.model == 'MultiAttConvLSTM':
model = MultiAttConvLSTM(
input_dim=input_dim,
att_inputs=cluster_centroid,
att_nodes=FLAGS.att_nodes,
batch_size=FLAGS.batch_size,
layer={
'encoder': ['conv', 'conv', 'conv_lstm', 'conv_lstm'],
'decoder': ['conv_lstm', 'conv_lstm', 'conv', 'conv'],
'attention': ['conv', 'conv']
},
layer_param={
'encoder': [[[3, 3], [1, 1, 1, 1], 8],
[[3, 3], [1, 1, 1, 1], 16],
[[16, 16], [3, 3], 64], [[16, 16], [3, 3],
64]],
'decoder': [[[16, 16], [3, 3], 64], [[16, 16], [3, 3], 64],
[[3, 3], [1, 1, 1, 1], 8],
[[3, 3], [1, 1, 1, 1], 2]],
'attention': [[[3, 3], [1, 1, 1, 1], 8],
[[3, 3], [1, 1, 1, 1], 16]]
},
input_steps=10,
output_steps=10)
print('model solver...')
solver = ModelSolver(model,
train,
val,
preprocessing=pre_process,
n_epochs=FLAGS.n_epochs,
batch_size=FLAGS.batch_size,
update_rule=FLAGS.update_rule,
learning_rate=FLAGS.lr,
save_every=FLAGS.save_every,
pretrained_model=FLAGS.pretrained_model,
model_path=model_path,
test_model=model_path + 'model-' +
str(FLAGS.n_epochs),
log_path=log_path)
if FLAGS.train:
print('begin training...')
test_prediction, _ = solver.train(test)
test_target = np.asarray(test_y)
if FLAGS.test:
print('test trained model...')
solver.test_model = solver.model_path + FLAGS.pretrained_model
test_prediction = solver.test(test)
test_target = np.asarray(test_y)
np.save('citybike-results/results/' + FLAGS.model + '/test_target.npy',
test_target)
np.save('citybike-results/results/' + FLAGS.model + '/test_prediction.npy',
test_prediction)