def main (args):
"""Main function for timeGAN experiments.
Args:
- data_name: sine, stock, or energy
- seq_len: sequence length
- Network parameters (should be optimized for different datasets)
- module: gru, lstm, or lstmLN
- hidden_dim: hidden dimensions
- num_layer: number of layers
- iteration: number of training iterations
- batch_size: the number of samples in each batch
- metric_iteration: number of iterations for metric computation
Returns:
- ori_data: original data
- generated_data: generated synthetic data
- metric_results: discriminative and predictive scores
"""
## Data loading
if args.data_name in ['stock', 'energy']:
ori_data = real_data_loading(args.data_name, args.seq_len)
elif args.data_name == 'sine':
# Set number of samples and its dimensions
no, dim = 10000, 5
ori_data = sine_data_generation(no, args.seq_len, dim)
print(args.data_name + ' dataset is ready.')
## Synthetic data generation by TimeGAN
# Set newtork parameters
parameters = dict()
parameters['module'] = args.module
parameters['hidden_dim'] = args.hidden_dim
parameters['num_layer'] = args.num_layer
parameters['iterations'] = args.iteration
parameters['batch_size'] = args.batch_size
generated_data = timegan(ori_data, parameters)
print('Finish Synthetic Data Generation')
"""
## Performance metrics
# Output initialization
metric_results = dict()
# 1. Discriminative Score
discriminative_score = list()
for _ in range(args.metric_iteration):
temp_disc = discriminative_score_metrics(ori_data, generated_data)
discriminative_score.append(temp_disc)
metric_results['discriminative'] = np.mean(discriminative_score)
# 2. Predictive score
predictive_score = list()
for tt in range(args.metric_iteration):
temp_pred = predictive_score_metrics(ori_data, generated_data)
predictive_score.append(temp_pred)
metric_results['predictive'] = np.mean(predictive_score)
# 3. Visualization (PCA and tSNE)
visualization(ori_data, generated_data, 'pca')
visualization(ori_data, generated_data, 'tsne')
## Print discriminative and predictive scores
print(metric_results)
"""
return ori_data, generated_data #, metric_results
def main(args):
"""Main function for timeGAN experiments.
Args:
- data_name: sine, stock, or energy
- seq_len: sequence length
- Network parameters (should be optimized for different datasets)
- module: gru, lstm, or lstmLN
- hidden_dim: hidden dimensions
- num_layer: number of layers
- iteration: number of training iterations
- batch_size: the number of samples in each batch
- metric_iteration: number of iterations for metric computation
Returns:
- ori_data: original data
- generated_data: generated synthetic data
- metric_results: discriminative and predictive scores
"""
# Data loading
if args.data_name in ["stock", "energy"]:
ori_data = real_data_loading(args.data_name, args.seq_len)
elif args.data_name == "sine":
# Set number of samples and its dimensions
no, dim = 10000, 5
ori_data = sine_data_generation(no, args.seq_len, dim)
elif args.data_name == "hypo":
ori_data = real_data_loading(args.data_name, args.seq_len)
print(args.data_name + " dataset is ready.")
# Synthetic data generation by TimeGAN
# Set newtork parameters
parameters = dict()
parameters["module"] = args.module
parameters["hidden_dim"] = args.hidden_dim
parameters["num_layer"] = args.num_layer
parameters["iterations"] = args.iteration
parameters["batch_size"] = args.batch_size
print(len(ori_data), ori_data[0].shape)
generated_data = timegan(ori_data, parameters)
print("Finish Synthetic Data Generation")
print(len(generated_data), generated_data[0].shape)
if len(generated_data) > len(ori_data):
generated_data_part = generated_data[: len(ori_data)]
print(
"Generated data shape mismatch with original data, "
+ "calibrating part of generated data"
)
# Performance metrics
# Output initialization
metric_results = dict()
# 1. Discriminative Score
discriminative_score = list()
for _ in range(args.metric_iteration):
temp_disc = discriminative_score_metrics(ori_data, generated_data_part)
discriminative_score.append(temp_disc)
metric_results["discriminative"] = np.mean(discriminative_score)
# 2. Predictive score
predictive_score = list()
for tt in range(args.metric_iteration):
temp_pred = predictive_score_metrics(ori_data, generated_data_part)
predictive_score.append(temp_pred)
metric_results["predictive"] = np.mean(predictive_score)
# 3. Visualization (PCA and tSNE)
visualization(ori_data, generated_data_part, "pca", args)
visualization(ori_data, generated_data_part, "tsne", args)
# Print discriminative and predictive scores
print(metric_results)
return ori_data, generated_data, metric_results
# Experiments iterations
Iteration = 2
Sub_Iteration = 10
speed = 200 # 100,200,300,400,500
feed = 6 # 6,12
#%% Data Loading
seq_length = 24
if data_name == 'google':
dataX = google_data_loading(seq_length)
elif data_name == 'sine':
No = 10000
F_No = 5
dataX = sine_data_generation(No, seq_length, F_No)
elif data_name == 'TUD':
dataX,dataXs, min, max, idx, data_true = load_real_samples(seq_length, speed, feed, False)
print(data_name + ' dataset is ready.')
#%% Newtork Parameters
parameters = dict()
parameters['hidden_dim'] = len(dataX[0][0,:]) * 4
parameters['num_layers'] = 3
parameters['iterations'] = 2
parameters['batch_size'] = 128
parameters['module_name'] = 'gru' # Other options: 'lstm' or 'lstmLN'
parameters['z_dim'] = len(dataX[0][0,:])
#%% Experiments
print(X_mb, T_mb)
train_step_embedder(X_mb, T_mb)
print(itt)
#train()
####TESTING####
from data_loading import real_data_loading, sine_data_generation
data_name = 'sine'
seq_len = 5
if data_name in ['stock', 'energy']:
ori_data = real_data_loading(data_name, seq_len)
elif data_name == 'sine':
# Set number of samples and its dimensions
no, dim = 50, 2
ori_data = sine_data_generation(no, seq_len, dim)
print(data_name + ' dataset is ready.')
## Newtork parameters
parameters = dict()
parameters['module'] = 'lstm'
parameters['hidden_dim'] = 6
parameters['num_layer'] = 3
parameters['iterations'] = 10
parameters['batch_size'] = 4
timegan(ori_data, parameters)