# dqn.load_weights(checkpoint_weights_filename)
# elif os.path.isfile(weights_filename):
# print("Loading previous weights...")
# dqn.load_weights(weights_filename)
dqn.fit(env, callbacks=callbacks, nb_steps=20000000, log_interval=10000)
# After training is done, we save the final weights one more time.
dqn.save_weights(weights_filename, overwrite=True)
# Finally, evaluate our algorithm for 10 episodes.
dqn.test(env, nb_episodes=10, visualize=False)
elif args.mode == 'test':
weights_filename = 'wts/dqn_Breakout-v0_weights_12000000_phyran.h5f'.format(
args.env_name)
if args.weights:
weights_filename = args.weights
print(env.unwrapped.get_action_meanings())
np.random.seed(None)
env.seed(None)
dqn.load_weights(weights_filename)
dqn.training = False
dqn.test_policy = EpsilonPhysicsPolicy(
eps_phy=0.01, eps_ran=0.00
) # set a small epsilon for test policy to avoid getting stuck
env = gym.wrappers.Monitor(env,
"records/",
video_callable=lambda episode_id: True,
force=True)
dqn.test(env, nb_episodes=100, visualize=False)
env.close()
## Init RL agent
agent = DQNAgent(model=model, nb_actions=nb_actions,
memory=memory, nb_steps_warmup=1000,
target_model_update=1e-2, policy=policy,
processor=MultiInputProcessor(2),
# enable_dueling_network=True, dueling_type='avg'
)
agent.compile(Adam(lr=1e-3), metrics=['mae'])
## Comment this row if you want to start learning again
agent.load_weights('{p}/dqn_{fn}_weights.h5f'.format(p=PATH, fn=ENV_NAME))
## Train or evaluate
if TRAIN:
agent.training = True
observation = market.reset()
while True:
try:
# TODO add callbacks?
## Agent vybiraet dejstvie
# (candles=9(mb=>(2,4)?), tickers=4, trades=2)
# TODO actions for multy symbols market
action = agent.forward(observation)
## Execute action
observation, reward, done, info = market.step([action])