def test():
log_config()
devs = mx.cpu() if args.gpus is None else [
mx.gpu(int(i)) for i in args.gpus.split(',')
]
# module
dataiter = rl_data.GymDataIter('scenes',
args.batch_size,
args.input_length,
web_viz=True)
print(dataiter.provide_data)
net = sym.get_symbol_thor(dataiter.act_dim)
module = mx.mod.Module(net,
data_names=[d[0] for d in dataiter.provide_data],
label_names=('policy_label', 'value_label'),
context=devs)
module.bind(data_shapes=dataiter.provide_data,
label_shapes=[('policy_label', (args.batch_size, )),
('value_label', (args.batch_size, 1))],
for_training=False)
# load model
assert args.load_epoch is not None
assert args.model_prefix is not None
module.load_params('%s-%04d.params' % (args.model_prefix, args.load_epoch))
N = args.num_epochs * args.num_examples / args.batch_size
R = 0
T = 1e-20
score = np.zeros((args.batch_size, ))
for t in range(N):
dataiter.clear_history()
data = dataiter.next()
module.forward(data, is_train=False)
act = module.get_outputs()[0].asnumpy()
act = [
np.random.choice(dataiter.act_dim, p=act[i])
for i in range(act.shape[0])
]
dataiter.act(act)
time.sleep(0.05)
_, reward, _, done = dataiter.history[0]
T += done.sum()
score += reward
R += (done * score).sum()
score *= (1 - done)
if t % 100 == 0:
logging.info('n %d score: %f T: %f' % (t, R / T, T))
def test():
log_config()
devs = mx.cpu() if args.gpus is None else [
mx.gpu(int(i)) for i in args.gpus.split(',')]
# module
dataiter = robo_data.RobosimsDataIter('scenes', args.batch_size, args.input_length, web_viz=True)
print(dataiter.provide_data)
net = sym.get_symbol_thor(dataiter.act_dim)
module = mx.mod.Module(net, data_names=[d[0] for d in dataiter.provide_data], label_names=('policy_label', 'value_label'), context=devs)
module.bind(data_shapes=dataiter.provide_data,
label_shapes=[('policy_label', (args.batch_size,)), ('value_label', (args.batch_size, 1))],
for_training=False)
# load model
assert args.load_epoch is not None
assert args.model_prefix is not None
module.load_params('%s-%04d.params'%(args.model_prefix, args.load_epoch))
N = args.num_epochs * args.num_examples / args.batch_size
R = 0
T = 1e-20
score = np.zeros((args.batch_size,))
for t in range(N):
dataiter.clear_history()
data = dataiter.next()
module.forward(data, is_train=False)
act = module.get_outputs()[0].asnumpy()
act = [np.random.choice(dataiter.act_dim, p=act[i]) for i in range(act.shape[0])]
dataiter.act(act)
time.sleep(0.05)
_, reward, _, done = dataiter.history[0]
T += done.sum()
score += reward
R += (done*score).sum()
score *= (1-done)
if t % 100 == 0:
logging.info('n %d score: %f T: %f'%(t, R/T, T))