import numpy as np
import tensorflow as tf
from environments.coppeliagym import CoppeliaEnv
from utils.utils import setup, create_world
import collections
args = ['--ee_j_pos', '--vrep', '--render']
args = setup(args)
env, agent = create_world(args)
for eps in range(10):
state = env.reset()
env.render()
for t in range(args.time_limit):
if t < 150:
action = env.action_space.high
else:
action = env.action_space.low
np.putmask(action, [1, 1, 1, 1, 0, 1, 1, 1], 0)
next_state, *_ = env.step(action)
print(next_state[7])
images1: np.array, shape: (N, H, W, 3), dtype: np.float32 between 0-1 or np.uint8
images2: np.array, shape: (N, H, W, 3), dtype: np.float32 between 0-1 or np.uint8
use_multiprocessing: If multiprocessing should be used to pre-process the images
batch size: batch size used for inception network
Returns:
FID (scalar)
"""
mu1, sigma1 = calculate_activation_statistics(images1, batch_size, device)
mu2, sigma2 = calculate_activation_statistics(images2, batch_size, device)
fid = calculate_frechet_distance(mu1, sigma1, mu2, sigma2)
return fid
if __name__ == "__main__":
FLAGS, logger = setup(running_script="./utils/models.py",
config="./config.json")
input_path = FLAGS.input
valid_path = FLAGS.valid
device = FLAGS.device if torch.cuda.is_available() else "cpu"
network_dir = f'{FLAGS.path_prefix}/models/{FLAGS.network_name}'
print("\nCalculate FID-Score...")
file = open(f"{network_dir}/fid_{FLAGS.network_name}.txt", 'w')
gen_train_path = f"{network_dir}/generated_train_images/"
gen_valid_path = f"{network_dir}/generated_valid_images/"
num_images = len(os.listdir(gen_train_path))
batch_size = 32
image_size = 128
for path1, path2 in [(gen_train_path, input_path),
np.concatenate([state, agent.transitbuffer.goal]))
return action
def mock_goal_action(agent, state):
agent.select_action(state)
if agent.transitbuffer.meta_time:
agent.transitbuffer.goal = mock_goal
action = agent.sub_agent.select_action(
np.concatenate([state, agent.transitbuffer.goal]))
return action
if __name__ == "__main__":
# Parse Arguments and create directories
args = setup(sys.argv[1:])
env, agent = create_world(args)
#agent.load_model(f"./EXPERIMENTS/directional_10_0_1M/model/{agent.file_name}")
#agent.load_model(f"./EXPERIMENTS/absolute_10_0/model/{agent.file_name}")
#agent.load_model(f"./EXPERIMENTS/absolute_left_right_10/model/{agent.file_name}")
#agent.load_model(f"./EXPERIMENTS/6m_absolute_left_right_up_down/{agent.file_name}")
#agent.load_model(f"./EXPERIMENTS/2m_full_hiro_agent_15_5/model/{agent.file_name}")
agent.load_model(f"./models/{agent.file_name}")
state, done = env.reset(), False
episode_reward = 0
episode_timesteps = 0
episode_num = 0
second_bot = init_kuka()
init_state = state
while True:
second_bot.set_joint_positions(init_state[:7])
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="""Training VAE""")
parser.add_argument('imfolder',
type=str,
default=None,
metavar='image_dir',
help="""The path to the directory which contains
the imgge folder. The images themselves must be
in one or more subdirectories of the imfolder""")
args, rest = parser.parse_known_args()
FLAGS, logger = setup(running_script="./utils/training.py",
args=rest,
config="config.json")
print("FLAGS= ", FLAGS)
imfolder = os.path.abspath(args.imfolder)
device = FLAGS.device if torch.cuda.is_available() else "cpu"
print("\ninput dir: ", imfolder, "\ndevice: ", device)
os.makedirs(FLAGS.path_prefix, exist_ok=True)
if FLAGS.networkname in os.listdir(FLAGS.path_prefix):
input1 = input(
"\nNetwork already exists. Are you sure to proceed? ([y]/n) ")
if not input1 in ['y', 'yes']:
sys.exit()
# Setup logging and creates save dir
if args.logging:
args.experiment_dir, writer = setup_logging(args,
default_args,
args.args_to_ignore.split(','),
exp_name_use_date=True)
args.experiment_dir = Path(args.experiment_dir)
# Dump args
save_yaml(vars(args), f'{args.experiment_dir}/args.yaml')
# Setup everything else
setup(args)
# Load dataloaders
model_native_transform = m['model'].get_native_transform()
dataloader_train = m['dataloader'].get_dataloader(args, model_native_transform, part='train', phase='train')
dataloader_val = m['dataloader'].get_dataloader(args, model_native_transform, part='val', phase='val')
# Load criterion
criterion = get_criterion(args.criterion, args).to(args.device)
# Load model
model = m['model'].get_net(args, dataloader_train, criterion)