'gpu_rss_mb': detector.gpu_mem / len(img_list),
'gpu_util': detector.gpu_util * 100 / len(img_list)
}
perf_info = detector.det_times.report(average=True)
model_dir = FLAGS.model_dir
mode = FLAGS.run_mode
model_info = {
'model_name': model_dir.strip('/').split('/')[-1],
'precision': mode.split('_')[-1]
}
data_info = {
'batch_size': FLAGS.batch_size,
'shape': "dynamic_shape",
'data_num': perf_info['img_num']
}
det_log = PaddleInferBenchmark(detector.config, model_info,
data_info, perf_info, mems)
det_log('Det')
if __name__ == '__main__':
paddle.enable_static()
parser = argsparser()
FLAGS = parser.parse_args()
print_arguments(FLAGS)
FLAGS.device = FLAGS.device.upper()
assert FLAGS.device in ['CPU', 'GPU',
'XPU'], "device should be CPU, GPU or XPU"
main()
q_vals = session.run(target_dqn.q_values,
feed_dict={target_dqn.input: new_states})
double_q = q_vals[range(batch_size), arg_q_max]
# Bellman equation. Multiplication with (1-terminal_flags) makes sure that
# if the game is over, targetQ=rewards
target_q = rewards + (gamma * double_q * (1 - terminal_flags))
# Gradient descend step to update the parameters of the main network
loss, _ = session.run(
[main_dqn.loss, main_dqn.update],
feed_dict={
main_dqn.input: states,
main_dqn.target_q: target_q,
main_dqn.action: actions
})
return loss
args = utils.argsparser()
tf.random.set_random_seed(args.seed)
np.random.seed(args.seed)
tf.reset_default_graph()
# Control parameters
if args.task == "train":
utils.train(args, DQN, learn, "dist_dqn")
elif args.task == "evaluate":
utils.sample(args, DQN, "dist_dqn", save=False)
elif args.task == "log":
utils.generate_figures("dist_dqn")
else:
utils.sample(args, DQN, "dist_dqn")
