model = MODEL(c_in=test_env.observation_space.shape[0],
c_out=test_env.action_space.n,
seq_len=test_env.observation_space.shape[1])
model = model.to(device)
agent = PPO(model=model, memory=memory, config=config, device=device)
if os.path.exists('./save/model.m5'):
agent.model.load_state_dict(torch.load('./save/model.m5'))
avg_t = 0
avg_r = 0
for epi in range(1, n_episodes + 1):
print("episode {} start!".format(epi))
obs = test_env.reset()
done = False
while not done:
t = 0
action_list = []
while t < T_horizon:
prob = agent.model.pi(torch.FloatTensor(obs).unsqueeze(0).to(device))
action = Categorical(prob).sample().item()
obs_prime, reward, done, _ = test_env.step(action)
if reward is None:
continue
action_list.append(action)
#device = torch.device('cpu')
print(f"Device used: {device}")
policy_net = PolicyNetwork(input_dim, output_dim1, output_dim2, hidden_dim, n_layers=lstm_layers)
# Loading the best model
model_name = 'model/state_dict3.pt'
policy_net.load_state_dict(torch.load(model_name))
policy_net.to(device)
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())
print(f"Device used: {device}")
policy_net = PolicyNetwork(input_dim,
output_dim1,
output_dim2,
hidden_dim,
n_layers=lstm_layers)
# 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)
