1a693073cd01fffa7cb3f018b7221459703119f9,ch09/01_cartpole_dqn.py,,,#,48

Before Change

            optimizer.step()

            // clear batch
            batch_states.clear()
            batch_actions.clear()
            batch_targets.clear()

        new_rewards = exp_source.pop_total_rewards()

After Change

        selector.epsilon = max(EPSILON_STOP, EPSILON_START - step_idx / EPSILON_STEPS)
        replay_buffer.populate(1)

        if len(replay_buffer) < BATCH_SIZE:
            continue

        // sample batch
        batch = replay_buffer.sample(BATCH_SIZE)
        batch_states = [exp.state for exp in batch]
        batch_actions = [exp.action for exp in batch]
        batch_targets = [calc_target(net, exp.reward, exp.last_state)
                         for exp in batch]
        // train
        optimizer.zero_grad()
        states_v = Variable(torch.from_numpy(np.array(batch_states, dtype=np.float32)))
        net_q_v = net(states_v)

Instances