grad_check.py 文件源码

def backprop_check():
    xp = cuda.cupy if config.use_gpu else np
    duel = DDQN()

    state = xp.random.uniform(-1.0, 1.0, (2, config.rl_agent_history_length * config.ale_screen_channels, config.ale_scaled_screen_size[1], config.ale_scaled_screen_size[0])).astype(xp.float32)
    reward = [1, 0]
    action = [3, 4]
    episode_ends = [0, 0]
    next_state = xp.random.uniform(-1.0, 1.0, (2, config.rl_agent_history_length * config.ale_screen_channels, config.ale_scaled_screen_size[1], config.ale_scaled_screen_size[0])).astype(xp.float32)

    optimizer_conv = optimizers.Adam(alpha=config.rl_learning_rate, beta1=config.rl_gradient_momentum)
    optimizer_conv.setup(duel.conv)
    optimizer_fc = optimizers.Adam(alpha=config.rl_learning_rate, beta1=config.rl_gradient_momentum)
    optimizer_fc.setup(duel.fc)

    for i in xrange(10000):
        optimizer_conv.zero_grads()
        optimizer_fc.zero_grads()
        loss, _ = duel.forward_one_step(state, action, reward, next_state, episode_ends)
        loss.backward()
        optimizer_conv.update()
        optimizer_fc.update()
        print loss.data,
        print duel.conv.layer_2.W.data[0, 0, 0, 0],
        print duel.fc.layer_2.W.data[0, 0],