lstm_layer = agent.model.layers[0]
# Store lstm states
state_record = lstm_layer.states
# Reset states
agent.model.layers[0].reset_states()
agent.target_model.layers[0].reset_states()
agent.train_replay()
# Restore states
agent.model.layers[0].states = state_record
score += reward
state = next_state
if done:
sim.reset()
break
# every episode update the target model to be same with model
agent.update_target_model()
# every episode, plot the play time
scores.append(score)
episodes.append(e)
#plt.plot(episodes, scores, 'b')
'''
Plot normalized data
'''
if False:
try:
t = np.arange(len(data))
agent.load_state()
perturb = torch.from_numpy(np.random.rand(10,10) / 1)
agent.model.state_dict()['linear2.weight'] += perturb.float()
perturb2 = torch.from_numpy(np.random.rand(4,10) / 1)
agent.model.state_dict()['linear3.weight'] += perturb2.float()
print(agent.model.state_dict())
'''
losses, scores, episodes = [], [], []
sim = Simulator(orig_data, data, windowsize=windowsize)
for e in range(EPISODES):
# Write actions to log file
score = 0
state = Tensor(sim.reset())
while not sim.sim_done():
#state = Tensor(sim.state) # Get state
action = agent.get_action(state)
# Simulate trading
#-----------
max_idx = np.argmax(action[:3]) # Choose buy/sell/hold
next_state, reward, done = sim.step(max_idx, action[3])
next_state = Tensor(next_state)
#-----------
# save the sample <s, a, r, s'> to the replay memory
agent.replay_memory(state, action, reward, next_state, done)
state = next_state.clone()