def __init__(self,
model_func,
exploration_schedule,
obs_shape,
input_type,
lr=1e-4,
buffer_size=1000000,
num_actions=6,
latent_dim=32,
gamma=0.99,
knn=4,
eval_epsilon=0.01,
queue_threshold=5e-5,
batch_size=32,
tf_writer=None):
self.conn, child_conn = Pipe()
self.replay_buffer = np.empty((buffer_size, ) + obs_shape, np.float32)
self.ec_buffer = PSLearningProcess(num_actions, buffer_size,
latent_dim, obs_shape, child_conn,
gamma)
self.obs = None
self.z = None
self.h = None
self.ind = -1
self.writer = tf_writer
self.sequence = []
self.gamma = gamma
self.queue_threshold = queue_threshold
self.num_actions = num_actions
self.exploration_schedule = exploration_schedule
self.latent_dim = latent_dim
self.knn = knn
self.steps = 0
self.batch_size = batch_size
self.rmax = 100000
self.logger = logging.getLogger("ecbp")
self.eval_epsilon = eval_epsilon
self.train_step = 4
self.alpha = 1
self.burnin = 2000
self.burnout = 1000000
self.loss_type = ["contrast"]
self.hash_func, self.train_func, self.eval_func, self.norm_func = build_train_contrast(
make_obs_ph=lambda name: input_type(obs_shape, name=name),
model_func=model_func,
num_actions=num_actions,
optimizer=tf.train.AdamOptimizer(learning_rate=lr, epsilon=1e-4),
gamma=gamma,
grad_norm_clipping=10,
loss_type=self.loss_type)
self.ec_buffer.start()
def __init__(self,
encoder_func,
decoder_func,
exploration_schedule,
obs_shape,
vector_input=True,
lr=1e-4,
buffer_size=1000000,
num_actions=6,
latent_dim=32,
gamma=0.99,
knn=4,
eval_epsilon=0.1,
queue_threshold=5e-5,
batch_size=32,
density=True,
trainable=True,
num_neg=10,
tf_writer=None):
self.conn, child_conn = Pipe()
self.replay_buffer = np.empty((buffer_size + 10, ) + obs_shape,
np.float32 if vector_input else np.uint8)
self.ec_buffer = PSLearningProcess(num_actions,
buffer_size,
latent_dim * 2,
obs_shape,
child_conn,
gamma,
density=density)
self.obs = None
self.z = None
self.cur_capacity = 0
self.ind = -1
self.writer = tf_writer
self.sequence = []
self.gamma = gamma
self.queue_threshold = queue_threshold
self.num_actions = num_actions
self.exploration_schedule = exploration_schedule
self.latent_dim = latent_dim
self.knn = knn
self.steps = 0
self.batch_size = batch_size
self.rmax = 100000
self.logger = logging.getLogger("ecbp")
self.log("psmp learning agent here")
self.eval_epsilon = eval_epsilon
self.train_step = 4
self.alpha = 1
self.burnin = 2000
self.burnout = 10000000000
self.update_target_freq = 10000
self.buffer_capacity = 0
self.trainable = trainable
self.num_neg = num_neg
self.loss_type = ["attention"]
input_type = U.Float32Input if vector_input else U.Uint8Input
# input_type = U.Uint8Input
self.hash_func, self.unmask_z_func, self.train_func, self.eval_func, self.norm_func, self.attention_func, self.value_func, self.reconstruct_func, self.update_target_func = build_train_mer_bvae_attention(
input_type=input_type,
obs_shape=obs_shape,
encoder_func=encoder_func,
decoder_func=decoder_func,
num_actions=num_actions,
optimizer=tf.train.AdamOptimizer(learning_rate=lr, epsilon=1e-4),
gamma=gamma,
grad_norm_clipping=10,
latent_dim=latent_dim,
loss_type=self.loss_type,
batch_size=batch_size,
num_neg=num_neg,
c_loss_type="sqmargin",
)
self.finds = [0, 0]
self.ec_buffer.start()