# Loading the best model
model_name = 'model/state_dict2.pt'
policy_net.load_state_dict(torch.load(model_name))
policy_net.to(device)
policy_net.eval(
) # to tell the model how to treat dropout (train: uses dropout, eval: do not use dropout)
state = test_env.reset()
log_probs = []
rewards = []
profits = []
hold_profits = []
for steps in range(test_env.steps_left):
if steps % 500 == 0:
test_env.render()
with torch.no_grad():
action, log_prob = policy_net.get_action(state, device)
new_state, reward, done, _ = test_env.step(action)
log_probs.append(log_prob)
rewards.append(reward)
profits.append(test_env._get_profit())
hold_profits.append(test_env._get_hold_profit())
if done:
print('\nTrading session terminated:')
print(f"total reward: {np.round(np.sum(rewards), decimals=3)}")
print(f"bot profit: {profits[-1]}")
print(f"hold profit: {hold_profits[-1]}")
max_episode_num = 1_000
all_rewards = [0]
avg_rewards = [0]
policy_net.train() # to tell the model how to treat dropout (train: uses dropout, eval: do not use dropout)
for episode in range(max_episode_num):
state = train_env.reset()
log_probs = []
rewards = []
profits = []
hold_profits = []
for steps in range(train_env.steps_left):
if steps % 500 ==0:
train_env.render()
action, log_prob = policy_net.get_action(state, device)
new_state, reward, done, _ = train_env.step(action)
log_probs.append(log_prob)
rewards.append(reward)
profits.append(train_env._get_profit())
hold_profits.append(train_env._get_hold_profit())
if done:
# Evaluate model
all_rewards.append(np.mean(rewards))
avg_rewards.append(np.mean(all_rewards[-10:]))
if all_rewards[-1] > all_rewards[-2]:
c_out=test_env.action_space.n,
seq_len=test_env.observation_space.shape[1])
#acer = ACER(model=model, memory=memory, config=acer_config)
if os.path.exists('./save/model.m5'):
model.load_state_dict(torch.load('./save/model.m5'))
avg_t = 0
avg_r = 0
obs = test_env.reset()
done = False
while True:
prob = model.pi(torch.FloatTensor(obs).unsqueeze(0))
action = Categorical(prob).sample().item()
obs_prime, reward, done, _ = test_env.step(action)
if reward is None:
continue
print("current price : {}".format(test_env._get_current_price()))
print("action : {}".format(action))
print("prob : {}".format(prob))
test_env.render(mode='human')
obs = obs_prime
print('Finished.')
