def dataset_proto(name=name):
if name == "linear":
dataset, _, opt_linear = sample_linear_data(num_contexts,
context_dim,
num_actions,
sigma=noise_stds)
elif name == "wheel":
dataset, opt_linear = sample_wheel_bandit_data(num_contexts, delta,
mean_v, std_v, mu_large,
std_large)
elif name == "mushroom":
mush = Mushrooms(num_contexts=num_contexts)
dataset = mush.table
opt_rewards, opt_actions = mush.opts[:, 0], mush.opts[:, 1]
opt_linear = (opt_rewards, opt_actions)
elif name == "covertype":
cov = Covertype(num_contexts=num_contexts)
dataset = cov.table
opt_rewards, opt_actions = cov.opts[:, 0], cov.opts[:, 1]
opt_linear = (opt_rewards, opt_actions)
return dataset, opt_linear
def sample_data(data_type, num_contexts=None):
"""Sample data from given 'data_type'.
Args:
data_type: Dataset from which to sample.
num_contexts: Number of contexts to sample.
Returns:
dataset: Sampled matrix with rows: (context, reward_1, ..., reward_num_act).
opt_rewards: Vector of expected optimal reward for each context.
opt_actions: Vector of optimal action for each context.
num_actions: Number of available actions.
context_dim: Dimension of each context.
"""
if data_type == 'linear':
# Create linear dataset
num_actions = 8
context_dim = 10
noise_stds = [0.01 * (i + 1) for i in range(num_actions)]
dataset, _, opt_linear = sample_linear_data(num_contexts, context_dim,
num_actions, sigma=noise_stds)
opt_rewards, opt_actions = opt_linear
elif data_type == 'sparse_linear':
# Create sparse linear dataset
num_actions = 7
context_dim = 10
noise_stds = [0.01 * (i + 1) for i in range(num_actions)]
num_nnz_dims = int(context_dim / 3.0)
dataset, _, opt_sparse_linear = sample_sparse_linear_data(
num_contexts, context_dim, num_actions, num_nnz_dims, sigma=noise_stds)
opt_rewards, opt_actions = opt_sparse_linear
elif data_type == 'mushroom':
# Create mushroom dataset
num_actions = 2
context_dim = 117
file_name = FLAGS.mushroom_data
dataset, opt_mushroom = sample_mushroom_data(file_name, num_contexts)
opt_rewards, opt_actions = opt_mushroom
elif data_type == 'financial':
num_actions = 8
context_dim = 21
num_contexts = min(3713, num_contexts)
noise_stds = [0.01 * (i + 1) for i in range(num_actions)]
file_name = FLAGS.financial_data
dataset, opt_financial = sample_stock_data(file_name, context_dim,
num_actions, num_contexts,
noise_stds, shuffle_rows=True)
opt_rewards, opt_actions = opt_financial
elif data_type == 'jester':
num_actions = 8
context_dim = 32
num_contexts = min(19181, num_contexts)
file_name = FLAGS.jester_data
dataset, opt_jester = sample_jester_data(file_name, context_dim,
num_actions, num_contexts,
shuffle_rows=True,
shuffle_cols=True)
opt_rewards, opt_actions = opt_jester
elif data_type == 'statlog':
file_name = FLAGS.statlog_data
num_actions = 7
num_contexts = min(43500, num_contexts)
sampled_vals = sample_statlog_data(file_name, num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
elif data_type == 'adult':
file_name = FLAGS.adult_data
num_actions = 14
num_contexts = min(45222, num_contexts)
sampled_vals = sample_adult_data(file_name, num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
elif data_type == 'covertype':
file_name = FLAGS.covertype_data
num_actions = 7
num_contexts = min(150000, num_contexts)
sampled_vals = sample_covertype_data(file_name, num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
elif data_type == 'census':
file_name = FLAGS.census_data
num_actions = 9
num_contexts = min(150000, num_contexts)
sampled_vals = sample_census_data(file_name, num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
elif data_type == 'wheel':
delta = 0.95
num_actions = 5
context_dim = 2
mean_v = [1.0, 1.0, 1.0, 1.0, 1.2]
std_v = [0.05, 0.05, 0.05, 0.05, 0.05]
mu_large = 50
std_large = 0.01
dataset, opt_wheel = sample_wheel_bandit_data(num_contexts, delta,
mean_v, std_v,
mu_large, std_large)
opt_rewards, opt_actions = opt_wheel
return dataset, opt_rewards, opt_actions, num_actions, context_dim
def dataset_proto():
dataset, _, opt_linear = sample_linear_data(num_contexts, context_dim,
num_actions, sigma=noise_stds)
return dataset, opt_linear
def sample_data(data_type, num_contexts=None):
"""Sample data from given 'data_type'.
Args:
data_type: Dataset from which to sample.
num_contexts: Number of contexts to sample.
Returns:
dataset: Sampled matrix with rows: (context, reward_1, ..., reward_num_act).
opt_rewards: Vector of expected optimal reward for each context.
opt_actions: Vector of optimal action for each context.
num_actions: Number of available actions.
context_dim: Dimension of each context.
"""
if data_type == '2linear':
# Create linear dataset
num_actions = 2
context_dim = 10
noise_stds = [0.01 * (i + 1) for i in range(num_actions)]
dataset, _, opt_linear = sample_linear_data(num_contexts,
context_dim,
num_actions,
sigma=noise_stds)
opt_rewards, opt_actions = opt_linear
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
if data_type == 'linear':
# Create linear dataset
num_actions = 8
context_dim = 10
noise_stds = [0.01 * (i + 1) for i in range(num_actions)]
dataset, _, opt_linear = sample_linear_data(num_contexts,
context_dim,
num_actions,
sigma=noise_stds)
opt_rewards, opt_actions = opt_linear
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'sparse_linear':
# Create sparse linear dataset
num_actions = 7
context_dim = 10
noise_stds = [0.01 * (i + 1) for i in range(num_actions)]
num_nnz_dims = int(context_dim / 3.0)
dataset, _, opt_sparse_linear = sample_sparse_linear_data(
num_contexts,
context_dim,
num_actions,
num_nnz_dims,
sigma=noise_stds)
opt_rewards, opt_actions = opt_sparse_linear
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'mushroom':
# Create mushroom dataset
num_actions = 2
context_dim = 117
file_name = FLAGS.mushroom_data
dataset, opt_mushroom = sample_mushroom_data(file_name, num_contexts)
opt_rewards, opt_actions = opt_mushroom
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'financial':
num_actions = 8
context_dim = 21
num_contexts = min(3713, num_contexts)
noise_stds = [0.01 * (i + 1) for i in range(num_actions)]
file_name = FLAGS.financial_data
dataset, opt_financial = sample_stock_data(file_name,
context_dim,
num_actions,
num_contexts,
noise_stds,
shuffle_rows=True)
opt_rewards, opt_actions = opt_financial
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'jester':
num_actions = 8
context_dim = 32
num_contexts = min(19181, num_contexts)
file_name = FLAGS.jester_data
dataset, opt_jester = sample_jester_data(file_name,
context_dim,
num_actions,
num_contexts,
shuffle_rows=True,
shuffle_cols=True)
opt_rewards, opt_actions = opt_jester
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'statlog':
file_name = FLAGS.statlog_data
num_actions = 7
num_contexts = min(43500, num_contexts)
sampled_vals = sample_statlog_data(file_name,
num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'adult':
file_name = FLAGS.adult_data
num_actions = 2
num_contexts = min(45222, num_contexts)
sampled_vals = sample_adult_data(file_name,
num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'covertype':
file_name = FLAGS.covertype_data
num_actions = 7
num_contexts = min(150000, num_contexts)
sampled_vals = sample_covertype_data(file_name,
num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1] #54
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'census':
file_name = FLAGS.census_data
num_actions = 9
num_contexts = min(150000, num_contexts)
sampled_vals = sample_census_data(file_name,
num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'wheel':
delta = 0.5
num_actions = 5
context_dim = 2
mean_v = [0.1, 0.1, 0.1, 0.1, 0.2]
std_v = [0.1, 0.1, 0.1, 0.1, 0.1]
mu_large = 0.4
std_large = 0.1
dataset, opt_wheel = sample_wheel_bandit_data(num_contexts, delta,
mean_v, std_v, mu_large,
std_large)
opt_rewards, opt_actions = opt_wheel
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'wheel2':
delta = 0.7
num_actions = 2
context_dim = 2
mean_v = [0.0, 1]
std_v = [0.1, 0.1]
mu_large = 2
std_large = 0.1
dataset, opt_wheel = sample_wheel2_bandit_data(num_contexts, delta,
mean_v, std_v, mu_large,
std_large)
opt_rewards, opt_actions = opt_wheel
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'eeg':
file_name = FLAGS.eeg_data
num_actions = 5
num_contexts = min(11500, num_contexts)
sampled_vals = sample_eeg_data(file_name,
num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'diabetic':
file_name = FLAGS.diabetic_data
num_actions = 3
num_contexts = min(100000, num_contexts)
sampled_vals = sample_diabetic_data(file_name,
num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'phone':
file_name = FLAGS.phone_data
num_actions = 6
num_contexts = min(7767, num_contexts)
sampled_vals = sample_phone_data(file_name,
num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'aps':
file_name = FLAGS.aps_data
num_actions = 2
num_contexts = min(76000, num_contexts)
sampled_vals = sample_aps_data(file_name,
num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards, opt_actions) = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
return dataset, opt_rewards, opt_actions, num_actions, context_dim, None
elif data_type == 'txt':
file_name = [FLAGS.positive_data_file, FLAGS.negative_data_file]
num_actions = 2
num_contexts = min(10000, num_contexts)
sampled_vals = sample_txt_data(file_name,
num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards,
opt_actions), vocab_processor = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
return dataset, opt_rewards, opt_actions, num_actions, context_dim, vocab_processor
elif data_type == 'amazon':
file_name = FLAGS.amazon_data_file
num_actions = 5
num_contexts = min(10000, num_contexts)
sampled_vals = sample_amazon_data(file_name,
num_contexts,
shuffle_rows=True)
contexts, rewards, (opt_rewards,
opt_actions), vocab_processor = sampled_vals
dataset = np.hstack((contexts, rewards))
context_dim = contexts.shape[1]
return dataset, opt_rewards, opt_actions, num_actions, context_dim, vocab_processor
def train(self, epochs, batch_size=128, sample_interval=50):
# Load the dataset
# (X_train, _), (_, _) = mnist.load_data()
self.dirname = os.path.dirname(__file__)
if self.file == "linear":
num_actions = 8
context_dim = 10
num_contexts = 1500
noise_stds = [0.01 * (i + 1) for i in range(num_actions)]
noise_stds = [1 for i in range(num_actions)]
X_train = sample_linear_data(num_contexts, context_dim,
num_actions, sigma=noise_stds)[0][:, :-8]
elif self.file == "wheel":
num_actions = 5
context_dim = 2
num_contexts = 1500
delta = 0.95
mean_v = [1.0, 1.0, 1.0, 1.0, 1.2]
std_v = [0.05, 0.05, 0.05, 0.05, 0.05]
mu_large = 50
std_large = 0.01
X_train = sample_wheel_bandit_data(num_contexts, delta,
mean_v, std_v,
mu_large, std_large)[0][:, :-5]
else:
_, X_train = get_labels_contexts_covertype(path=os.path.join(self.dirname, self.file))
# Adversarial ground truths
valid = -np.ones((batch_size, 1))
fake = np.ones((batch_size, 1))
for epoch in range(epochs):
for _ in range(self.n_critic):
# ---------------------
# Train Discriminator
# ---------------------
# Select a random batch of images
idx = np.random.randint(0, X_train.shape[0], batch_size)
imgs = X_train[idx]
# Sample noise as generator input
noise = np.random.normal(0, 1, (batch_size, self.n_noise))
# Generate a batch of new images
gen_imgs = self.generator.predict(noise)
# Train the critic
d_loss_real = self.critic.train_on_batch(imgs, valid)
d_loss_fake = self.critic.train_on_batch(gen_imgs, fake)
d_loss = 0.5 * np.add(d_loss_fake, d_loss_real)
# print("piche", d_loss_real, d_loss_fake)
# Clip critic weights
for l in self.critic.layers:
weights = l.get_weights()
weights = [np.clip(w, -self.clip_value, self.clip_value) for w in weights]
l.set_weights(weights)
# ---------------------
# Train Generator
# ---------------------
g_loss = self.combined.train_on_batch(noise, valid)
