def __init__(self, env_name='Hopper-v2', total_episodes=1000, learning_steps=1000, update_time=1,
episode_length=1000, total_steps=int(1e6), lr=1e-3, action_bound=1, num_samples=10, noise=0.02, best=2,
std_dev=0.03, batch_size=100, elite_percentage=0.2, mutate=0.9, crossover=0.2, hidden_size=64, seed=1,
namescope='default'):
self.env = gym.make(env_name)
np.random.seed(seed)
self.env.seed(seed)
tf.set_random_seed(seed)
self.input_size = self.env.observation_space.shape[0]
self.output_size = self.env.action_space.shape[0]
self.total_episodes = total_episodes
self.episode_length = episode_length
self.total_steps = total_steps
self.update_time = update_time
self.lr = lr
self.action_bound = action_bound
self.num_samples = num_samples
self.noise = noise
self.stddev = std_dev
self.batch_size = batch_size
self.elite_percentage = elite_percentage
self.mutate = mutate
self.crossover = crossover
self.hidden_size = hidden_size
self.normalizer = utils.Normalizer(self.input_size)
self.batch_size = batch_size
self.namescope = namescope
self.gamma = 1
self.best = best
# config = tf.ConfigProto(device_count={'GPU': gpu})
self.learning_steps = learning_steps
self.td3_agent = td3_network.TD3(self.input_size, self.output_size, 1, namescope=self.namescope,
hidden_size=hidden_size)
def __init__(self, env_name='Hopper-v2', total_episodes=1000, action_bound=1,
episode_length=1000, learning_rate=0.02, weight=0.01, learning_steps=100,
num_samples=8, noise=0.02, bc_index=[], std_dev=0.03, syn_step=10,
meta_population_size=5, seed=1, hidden_size=300):
self.env = gym.make(env_name)
np.random.seed(seed)
self.env.seed(seed)
self.action_bound = action_bound
self.input_size = self.env.observation_space.shape[0]
self.output_size = self.env.action_space.shape[0]
self.total_episodes = total_episodes
self.episode_length = episode_length
self.lr = learning_rate
self.num_samples = num_samples
self.noise = noise
self.meta_population_size = meta_population_size
self.seed = seed
self.syn_step = syn_step
self.learning_steps = learning_steps
self.bc_index = bc_index
self.weight = weight
self.normalizer = utils.Normalizer(self.env.observation_space.shape[0])
self.hidden_size = hidden_size
self.stddev = std_dev
self.td3_agent = td3_network.TD3(self.input_size, self.output_size, 1, hidden_size=self.hidden_size)
self.num_best_deltas = 4
def __init__(self,
env_name='Hopper-v2',
total_episodes=1000,
action_bound=1,
episode_length=1000,
learning_rate=0.02,
weight=0.01,
learning_steps=100,
num_samples=8,
noise=0.02,
bc_index=[],
std_dev=0.03,
syn_step=10,
num_best=4,
meta_population_size=5,
seed=1,
hidden_size=300,
coefficient=1):
self.env = gym.make(env_name)
np.random.seed(seed)
self.env.seed(seed)
self.action_bound = action_bound
self.input_size = self.env.observation_space.shape[0]
self.output_size = self.env.action_space.shape[0]
self.total_episodes = total_episodes
self.episode_length = episode_length
self.lr = learning_rate
self.num_best = num_best
self.num_samples = num_samples
self.noise = noise
self.meta_population_size = meta_population_size
self.seed = seed
self.syn_step = syn_step
self.coefficient = coefficient
self.learning_steps = learning_steps
self.bc_index = bc_index
self.weight = weight
self.normalizer = utils.Normalizer(self.env.observation_space.shape[0])
self.hidden_size = hidden_size
self.stddev = std_dev
self.intrinsic_network = IntrinsicNetwork(state_dim=self.input_size,
action_dim=self.output_size,
seed=self.seed,
namescope=str(seed),
weight=self.weight)
self.replay = utils.ReplayBuffer()
def __init__(self,
env_name='Hopper-v2',
total_episodes=1000,
namescope='default',
episode_length=1000,
total_steps=int(1e6),
lr=0.01,
action_bound=1,
num_samples=10,
noise=0.02,
std_dev=0.03,
batch_size=64,
elite_percentage=0.2,
mutate=0.9,
crossover=0.2,
hidden_size=64,
seed=1):
self.env = gym.make(env_name)
np.random.seed(seed)
self.env.seed(seed)
tf.set_random_seed(seed)
self.input_size = self.env.observation_space.shape[0]
self.output_size = self.env.action_space.shape[0]
self.total_episodes = total_episodes
self.episode_length = episode_length
self.total_steps = total_steps
self.lr = lr
self.action_bound = action_bound
self.num_samples = num_samples
self.noise = noise
self.stddev = std_dev
self.batch_size = batch_size
self.elite_percentage = elite_percentage
self.mutate = mutate
self.gamma = 1
self.crossover = crossover
self.hidden_size = hidden_size
self.normalizer = utils.Normalizer(self.input_size)
self.namescope = namescope
import trading_vix_env
import numpy as np
import config as C
import utils
import train
import gym
import policy
if __name__ == '__main__':
np.random.seed(0)
env = trading_vix_env.trading_vix_env()
#env = gym.make('Pendulum-v0')
env.seed(0)
#env = wrappers.Monitor(env, monitor_dir, force=True)
#policy_inputs_size = env.observation_space.shape[0]
policy_outputs_size = env.action_space.shape[0]
policy = policy.Policy(C.extracted_feature_size, policy_outputs_size)
normalizer = utils.Normalizer(C.extracted_feature_size)
train.train(env, policy, normalizer)
target_orig_heights = torch.stack(target_orig_heights)
target_orig_widths = torch.stack(target_orig_widths)
target_orig_sizes = torch.stack((target_orig_heights,
target_orig_widths)).transpose(0, 1)
origsize = (dictionaries[0]['orig_height'].item(),
dictionaries[0]['orig_width'].item())
# Tensor -> float & numpy
if testset.there_is_gt:
target_count = target_count.item()
target_locations = \
target_locations[0].to(device_cpu).numpy().reshape(-1, 2)
target_orig_size = \
target_orig_sizes[0].to(device_cpu).numpy().reshape(2)
normalzr = utils.Normalizer(args.height, args.width)
# Feed forward
with torch.no_grad():
est_maps, est_count = model.forward(imgs)
# Convert to original size
est_map_np = est_maps[0, :, :].to(device_cpu).numpy()
est_map_np_origsize = \
skimage.transform.resize(est_map_np,
output_shape=origsize,
mode='constant')
orig_img_np = imgs[0].to(device_cpu).squeeze().numpy()
orig_img_np_origsize = ((skimage.transform.resize(orig_img_np.transpose((1, 2, 0)),
output_shape=origsize,
mode='constant') + 1) / 2.0 * 255.0).\