def __init__(self, env, nnet, nb_actions):
model = nnet.model
# keras-rl
memory = SequentialMemory(limit=MEMORY_LIMIT,
window_length=WINDOW_LENGTH)
policy = LinearAnnealedPolicy(EpsGreedyQPolicy(),
attr='eps',
value_max=1.,
value_min=.1,
value_test=.05,
nb_steps=EPS_GREEDY_NB_STEPS)
dqn = DQNAgent(model=model,
nb_actions=nb_actions,
policy=policy,
memory=memory,
nb_steps_warmup=NB_STEPS_WARMUP,
target_model_update=TARGET_MODEL_UPDATE
) # gamma=.99, train_interval=4, delta_clip=1.
dqn.compile(Adam(lr=.00025), metrics=['mae'])
self.processor = AgentProcessor()
dqn.processor = self.processor
#for training
self.env = env
self.dqn = dqn
conc = concatenate([model_phase_encoded, model_vehicle_encoded])
hidden = Dense(128)(conc)
hidden = LeakyReLU()(hidden)
hidden = Dense(64)(hidden)
hidden = LeakyReLU()(hidden)
output = Dense(nb_actions, activation='linear')(hidden)
model = Model(inputs=[model_phase_input, model_vehicle_input], outputs=output)
model_path = "dqn_model.h5"
try:
model.load_weights(model_path)
print(f"Success loading previous weights at {model_path}")
except BaseException as e:
print(f"Did not load previous weights due to {e}, {model_path}")
### Policy, Memory & Agent set-up.
policy = LinearAnnealedPolicy(EpsGreedyQPolicy(), attr='eps', value_max=1., value_min=.01, value_test=.01, nb_steps=100000)
memory = SequentialMemory(limit=50000, window_length=1)
dqn = DQNAgent(model=model, nb_actions=nb_actions, memory=memory, policy=policy, batch_size=64, gamma=.95, nb_steps_warmup=2000, target_model_update=.001)
dqn.processor = MultiInputProcessor(2)
dqn.compile(optimizer=Adam(lr=.001))
### Fit.
hist = dqn.fit(env, nb_steps=200, verbose=1, log_interval=10)
dqn.save_weights(model_path, overwrite=True)
print("Saved model to disk")
test_env = CityFlowAgent(mode='predict', config_path=config_path)
start_time = default_timer()
dqn.test(test_env, nb_episodes=1, visualize=False)
print(f"\n Done testing inn {default_timer()-start_time} seconds")
train_processor = ImageProcessor(train_env)
dqn = DQNAgent(enable_double_dqn=False,
model=model,
nb_actions=nb_actions,
gamma=gamma,
batch_size=batch_size,
memory=memory,
nb_steps_warmup=50000,
target_model_update=1e2,
policy=policy,
processor=train_processor)
dqn.compile(Adam(lr=learning_rate), metrics=['accuracy'])
dqn.processor = train_processor
experiment_name = "NoseTip"
history_train = dqn.fit(train_env,
nb_steps=500,
nb_max_episode_steps=100,
log_interval=30000,
visualize=False,
verbose=2)
dqn.save_weights(experiment_name, overwrite=True)
print("******", train_env.wander)
L = Logging()