def main(env_args, agent_args, buffer_args, render=False):
utils.set_global_seed()
if env_args.get('n_workers', 0) > 1:
ray.init()
agent_name = 'Agent'
agent = Agent(agent_name,
agent_args,
env_args,
save=False,
log=True,
log_tensorboard=True,
log_params=False,
log_stats=True,
device='/gpu:0')
test_agent = None
if render:
env_args['n_envs'] = 1 # run test agent in a single environment
env_args['log_video'] = True
test_agent = Agent(agent_name,
agent_args,
env_args,
save=False,
log_tensorboard=False,
log_params=False,
log_stats=False,
device='/gpu:0',
reuse=True,
graph=agent.graph)
train(agent, agent_args, test_agent)
def test_buffer_op(self):
replay = create_replay(config)
simp_replay = ReplayBuffer(config)
env = gym.make('BipedalWalkerHardcore-v3')
s = env.reset()
for i in range(10000):
a = env.action_space.sample()
ns, r, d, _ = env.step(a)
if d:
ns = env.reset()
replay.add(obs=s.astype(np.float32),
action=a.astype(np.float32),
reward=np.float32(r),
next_obs=ns.astype(np.float32),
done=d)
simp_replay.add(obs=s, action=a, reward=r, next_obs=ns, done=d)
s = ns
if i > 1000:
set_global_seed(i)
sample1 = replay.sample()
set_global_seed(i)
sample2 = simp_replay.sample()
for k in sample1.keys():
np.testing.assert_allclose(sample1[k],
sample2[k],
err_msg=f'{k}')
def main():
set_global_seed()
args_file = 'args.yaml'
args = load_args(args_file)
cmd_args = parse_cmd_args()
if cmd_args.batch_size:
args['eval_batch_size'] = cmd_args
model = SAGAN('model', args, training=False)
model.restore(cmd_args.checkpoint)
model.evaluate(n_iterations=cmd_args.iterations)
def main(env_args, agent_args, buffer_args, render=False):
set_global_seed()
algorithm = agent_args['algorithm']
buffer_args['filename'] = f'experts/data/{env_args["name"]}.pkl'
agent = Agent(agent_args['algorithm'],
agent_args,
env_args,
buffer_args,
log_tensorboard=True,
log_stats=True,
save=True)
train(agent, algorithm, env_args['name'])
def main():
set_global_seed()
args_file = 'args.yaml'
args = load_args(args_file)
cmd_args = parse_cmd_args()
args['eval_image_path'] = cmd_args.image
image = imread(cmd_args.image)
h, w, c = image.shape
h = ceil(h / 4) * 4
w = ceil(w / 4) * 4
args['image_shape'] = (h, w, c)
model = StyleTransferModel('model', args, device='/gpu:0')
model.restore(cmd_args.checkpoint)
model.evaluate(eval_image=True)
def main(args, checkpoint=None):
# you may need this code to train multiple instances on a single GPU
# sess_config = tf.ConfigProto(allow_soft_placement=True)
# sess_config.gpu_options.allow_growth=True
# sess_config.gpu_options.per_process_gpu_memory_fraction = 0.45
# remember to pass sess_config to Model
model = SAGAN('model',
args,
log_tensorboard=True,
save=not checkpoint,
device='/gpu:0')
if checkpoint:
model.restore(checkpoint)
model.train()
if __name__ == '__main__':
cmd_args = parse_cmd_args()
set_global_seed()
args_file = 'args.yaml'
if cmd_args.checkpoint:
args = load_args(args_file)
main(args, cmd_args.checkpoint)
gs = GridSearch(args_file, main, dir_prefix=cmd_args.prefix)
gs()