epoch_loss = 0.0
for iteration in range(total_num // batch_size):
seed = np.random.choice(total_num, batch_size)
x_batch = x_train[seed]
y_batch = y_train[seed]
loss = model.loss(x_batch, y_batch)
epoch_loss += loss
model.zero_grad()
model.backward()
model.update()
epoch_loss /= batch_size
pred = model.predict(x_train)
target = np.argmax(y_train, axis=1)
score = len(np.where(pred == target)[0]) / len(target)
val_pred = model.predict(x_val)
val_target = np.argmax(y_val, axis=1)
val_score = len(np.where(val_pred == val_target)[0]) / len(val_target)
acc_stack.append(score)
val_acc_stack.append(val_score)
loss_stack.append(epoch_loss)
if (epoch % verbose) == 0:
spend = round((time.time() - st), 3)
print(
f"epoch : {epoch} | loss : {epoch_loss} | train_score : {score} | validation score : {val_score} | time : {spend} seconds"
def main(args):
# environment initialization
env = BaseEnv(size=args.env_size)
# embedding network initialization
f_s_a = MLP(env.state_size + env.action_size, args.s_a_hidden_size,
args.embedding_dim)
f_s = MLP(env.state_size, args.s_hidden_size, args.embedding_dim)
# buffer initialization
replay_buffer_1 = ReplayBuffer(args, env)
replay_buffer_2 = ReplayBuffer(args, env)
goal_buffer = GoalBuffer()
init_goal = tuple(np.random.randint(1, args.env_size[0] + 1, size=2))
goal_buffer.store(init_goal)
# agent initialization
agent = Agent(env_size=args.env_size)
# optimizer initialization
s_a_optimizer = torch.optim.Adam(f_s_a.parameters(), lr=args.s_a_lr)
s_optimizer = torch.optim.Adam(f_s.parameters(), lr=args.s_a_lr)
log_loss = []
for epoch in tqdm.tqdm(range(args.epoch_num)):
start_position = np.random.randint(1, args.env_size[0] + 1, size=2)
goal = goal_buffer.sample_batch_goal(size=1)[0]
print("goal point is :{}".format(goal))
g_feature = agent.get_state_feature(goal)
ns, r, terminate = env.reset(size=args.env_size,
start_pos=start_position)
for step in range(args.max_step):
s = ns
s_feature = agent.get_state_feature(s)
action, min_dist = agent.get_best_action(s_feature, f_s_a, f_s,
g_feature)
ns, r, terminate = env.step(action)
goal_buffer.store(ns)
ns_feature = agent.get_state_feature(ns)
vec_action = vectorize_action(action)
# print(s_feature.shape, vec_action.shape)
# store one step loss
# s_a_pred = f_s_a.predict(np.concatenate((s_feature, vec_action), axis=0).reshape((1, -1)))
# ns_pred = f_s.predict(np.array(ns_feature).reshape((1, -1)))
e_1 = agent.get_dist(s_feature, vec_action, ns_feature, f_s_a,
f_s)[0]
replay_buffer_1.add([s_feature, vec_action, ns_feature, None, e_1])
# store two step loss
sub_g = goal_buffer.sample_batch_goal(size=1, with_weights=False)
sub_g_feature = agent.get_states_feature(sub_g)[0]
na, min_dist = agent.get_best_action(ns_feature, f_s_a, f_s,
sub_g_feature)
dist_ns = agent.get_dist(ns_feature, vectorize_action(na),
sub_g_feature, f_s_a, f_s)
target = dist_ns + 1
dist_s = agent.get_dist(s_feature, vec_action, sub_g_feature,
f_s_a, f_s)
e_2 = abs(dist_s - target)
replay_buffer_2.add(
[s_feature, vec_action, ns_feature, sub_g_feature, e_2])
if terminate:
break
for step in range(args.random_step):
s = ns
s_feature = agent.get_state_feature(s)
action = agent.get_random_action()
ns, r, terminate = env.step(action)
goal_buffer.store(ns)
ns_feature = agent.get_state_feature(ns)
vec_action = vectorize_action(action)
# print(s_feature.shape, vec_action.shape)
# store one step loss
# s_a_pred = f_s_a.predict(np.concatenate((s_feature, vec_action), axis=0).reshape((1, -1)))
# ns_pred = f_s.predict(np.array(ns_feature).reshape((1, -1)))
e_1 = agent.get_dist(s_feature, vec_action, ns_feature, f_s_a,
f_s)[0]
replay_buffer_1.add([s_feature, vec_action, ns_feature, None, e_1])
# store two step loss
sub_g = goal_buffer.sample_batch_goal(size=1, with_weights=False)
sub_g_feature = agent.get_states_feature(sub_g)[0]
na, min_dist = agent.get_best_action(ns_feature, f_s_a, f_s,
sub_g_feature)
dist_ns = agent.get_dist(ns_feature, vectorize_action(na),
sub_g_feature, f_s_a, f_s)
target = dist_ns + 1
dist_s = agent.get_dist(s_feature, vec_action, sub_g_feature,
f_s_a, f_s)
e_2 = abs(dist_s - target)
replay_buffer_2.add(
[s_feature, vec_action, ns_feature, sub_g_feature, e_2])
batch_1, _, index_1 = replay_buffer_1.get_batch_data()
_, batch_2, index_2 = replay_buffer_2.get_batch_data()
loss_1 = torch.mean(
torch.norm((f_s_a(batch_1['sa']) - f_s(batch_1['ns'])), dim=1))
na = agent.get_best_actions(batch_2['ns'], f_s_a, f_s, batch_2['g'])
target = agent.get_dist(batch_2['ns'], na, batch_2['g'], f_s_a,
f_s) + 1
pred = torch.norm((f_s_a(batch_2['sa']) - f_s(batch_2['g'])), dim=1)
if (epoch + 1) % 100 == 0:
print(pred - target)
print(len(replay_buffer_1), len(replay_buffer_2))
# goal_buffer.goal_visualize()
loss_2 = torch.mean(torch.abs(pred - target))
# if epoch >= 1500:
# args.reg_term = 0.5
loss = (1 - args.reg_term) * loss_1 + args.reg_term * loss_2
log_loss.append(loss)
s_a_optimizer.zero_grad()
s_optimizer.zero_grad()
loss.backward()
# nn.utils.clip_grad_norm_(f_s.parameters(), args.grad_clip)
# nn.utils.clip_grad_norm_(f_s_a.parameters(), args.grad_clip)
s_a_optimizer.step()
s_optimizer.step()
# Update replay buffer
with torch.no_grad():
e_update_1 = torch.norm(torch.FloatTensor(
f_s_a.predict(batch_1['sa']) - f_s.predict(batch_1['ns'])),
dim=1)
na = agent.get_best_actions(batch_2['ns'], f_s_a, f_s,
batch_2['g'])
target = agent.get_dist(batch_2['ns'], na, batch_2['g'], f_s_a,
f_s) + 1
pred = torch.norm(torch.FloatTensor(
f_s_a.predict(batch_2['sa']) - f_s.predict(batch_2['g'])),
dim=1)
e_update_2 = torch.abs(pred - target)
replay_buffer_1.update_error(index_1, e_update_1)
replay_buffer_2.update_error(index_2, e_update_2)
if (epoch + 1) % 100 == 0:
print(
"epoch number: {}/{}, total_loss: {}, loss_normal: {}, loss_update: {}"
.format(epoch + 1, args.epoch_num, loss, loss_1, loss_2))
# if epoch == args.epoch_num - 1:
# print(s_a_embed, s_embed)
if (epoch + 1) % 1000 == 0 and loss < 1:
save_model("./model.pt", f_s_a, f_s, args.epoch_num, loss,
s_a_optimizer, s_optimizer)
print("Saving model at epoch: {}".format(epoch))
if (epoch + 1) % 10000 == 0:
plt.plot(range(epoch + 1), log_loss, color='red')
plt.savefig("./{}.pdf".format(int((epoch + 1) / 10000)),
dpi=1200,
bbox_inches='tight')