简单的 DQN 可以减慢训练速度答案

【问题标题】：Simple DQN to slow to train简单的 DQN 可以减慢训练速度
【发布时间】：2021-01-03 19:58:02
【问题描述】：

我一直在尝试用取自本文的 DQN 解决 OpenAI 月球着陆器游戏

https://arxiv.org/pdf/2006.04938v2.pdf

问题是训练 50 集需要 12 个小时，所以肯定有问题。

import os
import random
import gym
import numpy as np
from collections import deque
import tensorflow as tf
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense
from tensorflow.keras.optimizers import Adam
from tensorflow.keras import Model

ENV_NAME = "LunarLander-v2"

DISCOUNT_FACTOR = 0.9
LEARNING_RATE = 0.001

MEMORY_SIZE = 2000
TRAIN_START = 1000
BATCH_SIZE = 24

EXPLORATION_MAX = 1.0
EXPLORATION_MIN = 0.01
EXPLORATION_DECAY = 0.99

class MyModel(Model):
    def __init__(self, input_size, output_size):
        super(MyModel, self).__init__()
        self.d1 = Dense(128, input_shape=(input_size,), activation="relu")
        self.d2 = Dense(128, activation="relu")
        self.d3 = Dense(output_size, activation="linear")

    def call(self, x):
        x = self.d1(x)
        x = self.d2(x)
        return self.d3(x)

class DQNSolver():

    def __init__(self, observation_space, action_space):
        self.exploration_rate = EXPLORATION_MAX

        self.action_space = action_space
        self.memory = deque(maxlen=MEMORY_SIZE)

        self.model = MyModel(observation_space,action_space)
        self.model.compile(loss="mse", optimizer=Adam(lr=LEARNING_RATE))

    def remember(self, state, action, reward, next_state, done):
        self.memory.append((state, action, reward, next_state, done))

    def act(self, state):
        if np.random.rand() < self.exploration_rate:
            return random.randrange(self.action_space)
        q_values = self.model.predict(state)
        return np.argmax(q_values[0])

    def experience_replay(self):
        if len(self.memory) < BATCH_SIZE:
            return
        batch = random.sample(self.memory, BATCH_SIZE)
        state_batch, q_values_batch = [], []
        for state, action, reward, state_next, terminal in batch:
            # q-value prediction for a given state
            q_values_cs = self.model.predict(state)
            # target q-value
            max_q_value_ns = np.amax(self.model.predict(state_next)[0])
            # correction on the Q value for the action used
            if terminal:
                q_values_cs[0][action] = reward
            else:
                q_values_cs[0][action] = reward + DISCOUNT_FACTOR * max_q_value_ns
            state_batch.append(state[0])
            q_values_batch.append(q_values_cs[0])
        # train the Q network
        self.model.fit(np.array(state_batch),
                        np.array(q_values_batch),
                        batch_size = BATCH_SIZE,
                        epochs = 1, verbose = 0)
        self.exploration_rate *= EXPLORATION_DECAY
        self.exploration_rate = max(EXPLORATION_MIN, self.exploration_rate)

def lunar_lander():
    env = gym.make(ENV_NAME)
    observation_space = env.observation_space.shape[0]
    action_space = env.action_space.n
    dqn_solver = DQNSolver(observation_space, action_space)
    episode = 0
    print("Running")
    while True:
        episode += 1
        state = env.reset()
        state = np.reshape(state, [1, observation_space])
        scores = []
        score = 0
        while True:
            action = dqn_solver.act(state)
            state_next, reward, terminal, _ = env.step(action)
            state_next = np.reshape(state_next, [1, observation_space])
            dqn_solver.remember(state, action, reward, state_next, terminal)
            dqn_solver.experience_replay()
            state = state_next
            score += reward
            if terminal:
                print("Episode: " + str(episode) + ", exploration: " + str(dqn_solver.exploration_rate) + ", score: " + str(score))
                scores.append(score)
                break
        if np.mean(scores[-min(100, len(scores)):]) >= 195:
            print("Problem is solved in {} episodes.".format(episode))
            break
    env.close
if __name__ == "__main__":
    lunar_lander()

这是日志

root@b11438e3d3e8:~# /usr/bin/python3 /root/test.py
2021-01-03 13:42:38.055593: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1
2021-01-03 13:42:39.338231: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcuda.so.1
2021-01-03 13:42:39.368192: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-01-03 13:42:39.368693: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1716] Found device 0 with properties: 
pciBusID: 0000:01:00.0 name: GeForce GTX 1080 computeCapability: 6.1
coreClock: 1.8095GHz coreCount: 20 deviceMemorySize: 7.92GiB deviceMemoryBandwidth: 298.32GiB/s
2021-01-03 13:42:39.368729: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1
2021-01-03 13:42:39.370269: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcublas.so.10
2021-01-03 13:42:39.371430: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcufft.so.10
2021-01-03 13:42:39.371704: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcurand.so.10
2021-01-03 13:42:39.373318: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcusolver.so.10
2021-01-03 13:42:39.374243: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcusparse.so.10
2021-01-03 13:42:39.377939: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudnn.so.7
2021-01-03 13:42:39.378118: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-01-03 13:42:39.378702: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-01-03 13:42:39.379127: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1858] Adding visible gpu devices: 0
2021-01-03 13:42:39.386525: I tensorflow/core/platform/profile_utils/cpu_utils.cc:104] CPU Frequency: 3411185000 Hz
2021-01-03 13:42:39.386867: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x4fb44c0 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2021-01-03 13:42:39.386891: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): Host, Default Version
2021-01-03 13:42:39.498097: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-01-03 13:42:39.498786: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x4fdf030 initialized for platform CUDA (this does not guarantee that XLA will be used). Devices:
2021-01-03 13:42:39.498814: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): GeForce GTX 1080, Compute Capability 6.1
2021-01-03 13:42:39.498987: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-01-03 13:42:39.499416: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1716] Found device 0 with properties: 
pciBusID: 0000:01:00.0 name: GeForce GTX 1080 computeCapability: 6.1
coreClock: 1.8095GHz coreCount: 20 deviceMemorySize: 7.92GiB deviceMemoryBandwidth: 298.32GiB/s
2021-01-03 13:42:39.499448: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1
2021-01-03 13:42:39.499483: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcublas.so.10
2021-01-03 13:42:39.499504: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcufft.so.10
2021-01-03 13:42:39.499523: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcurand.so.10
2021-01-03 13:42:39.499543: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcusolver.so.10
2021-01-03 13:42:39.499562: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcusparse.so.10
2021-01-03 13:42:39.499581: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudnn.so.7
2021-01-03 13:42:39.499643: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-01-03 13:42:39.500113: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-01-03 13:42:39.500730: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1858] Adding visible gpu devices: 0
2021-01-03 13:42:39.500772: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1
2021-01-03 13:42:39.915228: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1257] Device interconnect StreamExecutor with strength 1 edge matrix:
2021-01-03 13:42:39.915298: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1263]      0 
2021-01-03 13:42:39.915322: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1276] 0:   N 
2021-01-03 13:42:39.915568: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-01-03 13:42:39.916104: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:982] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2021-01-03 13:42:39.916555: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1402] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 6668 MB memory) -> physical GPU (device: 0, name: GeForce GTX 1080, pci bus id: 0000:01:00.0, compute capability: 6.1)
Running
2021-01-03 13:42:40.267699: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcublas.so.10

这是 GPU 统计数据

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 450.66       Driver Version: 450.66       CUDA Version: 11.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce GTX 1080    Off  | 00000000:01:00.0  On |                  N/A |
|  0%   53C    P2    46W / 198W |   7718MiB /  8111MiB |      0%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+

如您所见，TensorFlow 不在 GPU 上计算，而是保留内存，所以我假设这是因为神经网络的输入太小，而是使用 CPU。

为了确保 GPU 安装正确，我从他们的文档中运行了一个示例，它使用了 GPU。也可以在google colabhere上重现问题

是算法问题还是代码问题？

在这种情况下有没有办法利用 GPU？

更新

事实证明，代理学习飞行而不是学习着陆，所以我添加了最大步长 150 来限制剧集时间，但它仍然很慢。

【问题讨论】：

嗨，主要原因似乎是dqn_solver.experience_replay()的使用，尝试在每集结束后使用它，而不是在每集的每个步骤中调用它。

标签： python tensorflow keras deep-learning openai-gym

【解决方案1】：

在笔记本设置中启用 GPU 的情况下在 colab 上运行我能够使用 wandb.ai 监控 GPU 使用情况，它正在尝试使用 GPU，但利用率为 0%。所以我能够复制这个问题，因为它对我来说也运行得很慢。根据this post 和also this one，我最好的猜测是环境中的步进函数正在成为流程的瓶颈。我尝试将批量大小增加 10 倍，以查看这是否会影响速度或 GPU 使用率，但没有成功。对于这个网络的大小和结构，CPU 可能更高效，所以这是唯一被使用的东西。这不能解释为什么你链接的论文在几个小时内就有结果，比我们得到的时间短得多，除非他们使用的笔记本电脑中的 CPU 明显更快。

【讨论】：

感谢您的宝贵时间！会不会是代码的问题？我不太确定重放内存。你见过其他使用 tensorflow2 的 DQN 实现吗？
不客气。抱歉，我无法提供更多帮助。我无法确定代码的问题，但我在深度学习方面的经验太少了，这并不意味着什么。祝你好运。