无法使用自定义预测例程将经过训练的模型部署到 Google Cloud Ai-Platform：模型需要的内存超出了允许的范围答案

【问题标题】：Cannot deploy trained model to Google Cloud Ai-Platform with custom prediction routine: Model requires more memory than allowed无法使用自定义预测例程将经过训练的模型部署到 Google Cloud Ai-Platform：模型需要的内存超出了允许的范围
【发布时间】：2019-12-17 10:52:25
【问题描述】：

我正在尝试使用自定义预测例程将预训练的 pytorch model 部署到 AI Platform。按照here 描述的说明进行操作后，部署失败并出现以下错误：

ERROR: (gcloud.beta.ai-platform.versions.create) Create Version failed. Bad model detected with error: Model requires more memory than allowed. Please try to decrease the model size and re-deploy. If you continue to have error, please contact Cloud ML.

模型文件夹的内容为 83.89 MB 大，低于文档中描述的 250 MB 限制。该文件夹中仅有的文件是模型的检查点文件 (.pth) 和自定义预测例程所需的 tarball。

创建模型的命令：

gcloud beta ai-platform versions create pose_pytorch --model pose --runtime-version 1.15 --python-version 3.5 --origin gs://rcg-models/pytorch_pose_estimation --package-uris gs://rcg-models/pytorch_pose_estimation/my_custom_code-0.1.tar.gz --prediction-class predictor.MyPredictor

将运行时版本更改为1.14 会导致同样的错误。我已经尝试将 --machine-type 参数更改为 mls1-c4-m2，就像 Parth 建议的那样，但我仍然遇到同样的错误。

生成my_custom_code-0.1.tar.gz 的setup.py 文件如下所示：

setup(
    name='my_custom_code',
    version='0.1',
    scripts=['predictor.py'],
    install_requires=["opencv-python", "torch"]
)

来自预测器的相关代码sn-p：

    def __init__(self, model):
        """Stores artifacts for prediction. Only initialized via `from_path`.
        """
        self._model = model
        self._client = storage.Client()

    @classmethod
    def from_path(cls, model_dir):
        """Creates an instance of MyPredictor using the given path.

        This loads artifacts that have been copied from your model directory in
        Cloud Storage. MyPredictor uses them during prediction.

        Args:
            model_dir: The local directory that contains the trained Keras
                model and the pickled preprocessor instance. These are copied
                from the Cloud Storage model directory you provide when you
                deploy a version resource.

        Returns:
            An instance of `MyPredictor`.
        """

        net = PoseEstimationWithMobileNet()
        checkpoint_path = os.path.join(model_dir, "checkpoint_iter_370000.pth")
        checkpoint = torch.load(checkpoint_path, map_location='cpu')
        load_state(net, checkpoint)

        return cls(net)

此外，我在 AI Platform 中为模型启用了日志记录，我得到以下输出：

2019-12-17T09:28:06.208537Z OpenBLAS WARNING - could not determine the L2 cache size on this system, assuming 256k 
2019-12-17T09:28:13.474653Z WARNING:tensorflow:From /usr/local/lib/python3.7/dist-packages/google/cloud/ml/prediction/frameworks/tf_prediction_lib.py:48: The name tf.saved_model.tag_constants.SERVING is deprecated. Please use tf.saved_model.SERVING instead. 
2019-12-17T09:28:13.474680Z {"textPayload":"","insertId":"5df89fad00073e383ced472a","resource":{"type":"cloudml_model_version","labels":{"project_id":"rcg-shopper","region":"","version_id":"lightweight_pose_pytorch","model_id":"pose"}},"timestamp":"2019-12-17T09:28:13.474680Z","logName":"projects/rcg-shopper/logs/ml.googleapis… 
2019-12-17T09:28:13.474807Z WARNING:tensorflow:From /usr/local/lib/python3.7/dist-packages/google/cloud/ml/prediction/frameworks/tf_prediction_lib.py:50: The name tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY is deprecated. Please use tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY instead. 
2019-12-17T09:28:13.474829Z {"textPayload":"","insertId":"5df89fad00073ecd4836d6aa","resource":{"type":"cloudml_model_version","labels":{"project_id":"rcg-shopper","region":"","version_id":"lightweight_pose_pytorch","model_id":"pose"}},"timestamp":"2019-12-17T09:28:13.474829Z","logName":"projects/rcg-shopper/logs/ml.googleapis… 
2019-12-17T09:28:13.474918Z WARNING:tensorflow: 
2019-12-17T09:28:13.474927Z The TensorFlow contrib module will not be included in TensorFlow 2.0. 
2019-12-17T09:28:13.474934Z For more information, please see: 
2019-12-17T09:28:13.474941Z   * https://github.com/tensorflow/community/blob/master/rfcs/20180907-contrib-sunset.md 
2019-12-17T09:28:13.474951Z   * https://github.com/tensorflow/addons 
2019-12-17T09:28:13.474958Z   * https://github.com/tensorflow/io (for I/O related ops) 
2019-12-17T09:28:13.474964Z If you depend on functionality not listed there, please file an issue. 
2019-12-17T09:28:13.474999Z {"textPayload":"","insertId":"5df89fad00073f778735d7c3","resource":{"type":"cloudml_model_version","labels":{"version_id":"lightweight_pose_pytorch","model_id":"pose","project_id":"rcg-shopper","region":""}},"timestamp":"2019-12-17T09:28:13.474999Z","logName":"projects/rcg-shopper/logs/ml.googleapis… 
2019-12-17T09:28:15.283483Z ERROR:root:Failed to import GA GRPC module. This is OK if the runtime version is 1.x 
2019-12-17T09:28:16.890923Z Copying gs://cml-489210249453-1560169483791188/models/pose/lightweight_pose_pytorch/15316451609316207868/user_code/my_custom_code-0.1.tar.gz... 
2019-12-17T09:28:16.891150Z / [0 files][    0.0 B/  8.4 KiB]                                                 
2019-12-17T09:28:17.007684Z / [1 files][  8.4 KiB/  8.4 KiB]                                                 
2019-12-17T09:28:17.009154Z Operation completed over 1 objects/8.4 KiB.                                       
2019-12-17T09:28:18.953923Z Processing /tmp/custom_code/my_custom_code-0.1.tar.gz 
2019-12-17T09:28:19.808897Z Collecting opencv-python 
2019-12-17T09:28:19.868579Z   Downloading https://files.pythonhosted.org/packages/d8/38/60de02a4c9013b14478a3f681a62e003c7489d207160a4d7df8705a682e7/opencv_python-4.1.2.30-cp37-cp37m-manylinux1_x86_64.whl (28.3MB) 
2019-12-17T09:28:21.537989Z Collecting torch 
2019-12-17T09:28:21.552871Z   Downloading https://files.pythonhosted.org/packages/f9/34/2107f342d4493b7107a600ee16005b2870b5a0a5a165bdf5c5e7168a16a6/torch-1.3.1-cp37-cp37m-manylinux1_x86_64.whl (734.6MB) 
2019-12-17T09:28:52.401619Z Collecting numpy>=1.14.5 
2019-12-17T09:28:52.412714Z   Downloading https://files.pythonhosted.org/packages/9b/af/4fc72f9d38e43b092e91e5b8cb9956d25b2e3ff8c75aed95df5569e4734e/numpy-1.17.4-cp37-cp37m-manylinux1_x86_64.whl (20.0MB) 
2019-12-17T09:28:53.550662Z Building wheels for collected packages: my-custom-code 
2019-12-17T09:28:53.550689Z   Building wheel for my-custom-code (setup.py): started 
2019-12-17T09:28:54.212558Z   Building wheel for my-custom-code (setup.py): finished with status 'done' 
2019-12-17T09:28:54.215365Z   Created wheel for my-custom-code: filename=my_custom_code-0.1-cp37-none-any.whl size=7791 sha256=fd9ecd472a6a24335fd24abe930a4e7d909e04bdc4cf770989143d92e7023f77 
2019-12-17T09:28:54.215482Z   Stored in directory: /tmp/pip-ephem-wheel-cache-i7sb0bmb/wheels/0d/6e/ba/bbee16521304fc5b017fa014665b9cae28da7943275a3e4b89 
2019-12-17T09:28:54.222017Z Successfully built my-custom-code 
2019-12-17T09:28:54.650218Z Installing collected packages: numpy, opencv-python, torch, my-custom-code

【问题讨论】：

您可以尝试使用 a) --machine-type=mls-c4-m2 b) --runtime-version 1.14 c) 尝试 a) 和 b) 一起运行吗？
我更新了我的问题以包含您的建议。
抱歉打错了，机器类型应该是 mls1-c4-m2。
在所有情况下都试过了。仍然出现同样的内存错误。
当你测试你的模型时，本地预测工作正常吗？我将查看 Pytorch 模型详细信息，看看是否可以重现。

标签： google-cloud-platform pytorch google-cloud-ml gcp-ai-platform-training

【解决方案1】：

这是一个常见问题，我们理解这是一个痛点。请执行以下操作：

torchvision 具有 torch 作为依赖项，默认情况下，它从 pypi 中提取 torch。

部署模型时，即使你指向使用自定义ai-platform torchvision包它也会这样做，因为torchvision是由PyTorch团队构建的，它被配置为使用torch作为依赖。这个来自 pypi 的 torch 依赖项提供了一个 720mb 的文件，因为它包含 GPU 单元

要解决#1，你需要从源代码buildtorchvision告诉torchvision你想从哪里得到torch，你需要把它设置成包去torch网站更小。使用 Python PEP-0440 direct references 功能重建 torchvision 二进制文件。在 torchvision setup.py 我们有：

pytorch_dep = 'torch'
if os.getenv('PYTORCH_VERSION'):
    pytorch_dep += "==" + os.getenv('PYTORCH_VERSION')

更新torchvision 中的setup.py 以使用直接引用功能：

requirements = [
     #'numpy',
     #'six',
     #pytorch_dep,
     'torch @ https://download.pytorch.org/whl/cpu/torch-1.4.0%2Bcpu-cp37-cp37m-linux_x86_64.whl'
]

*我已经为你做了这个*，所以我构建了 3 个你可以使用的轮子文件：

gs://dpe-sandbox/torchvision-0.4.0-cp37-cp37m-linux_x86_64.whl (torch 1.2.0, vision 0.4.0)
gs://dpe-sandbox/torchvision-0.4.2-cp37-cp37m-linux_x86_64.whl (torch 1.2.0, vision 0.4.2)
gs://dpe-sandbox/torchvision-0.5.0-cp37-cp37m-linux_x86_64.whl (torch 1.4.0  vision 0.5.0)

这些torchvision 包将从torch 站点而不是pypi 获得torch：（例如：https://download.pytorch.org/whl/cpu/torch-1.4.0%2Bcpu-cp37-cp37m-linux_x86_64.whl）

在将模型部署到 AI Platform 时更新您的模型 setup.py，使其不包括 torch 或 torchvision。
重新部署模型如下：

PYTORCH_VISION_PACKAGE=gs://dpe-sandbox/torchvision-0.5.0-cp37-cp37m-linux_x86_64.whl

gcloud beta ai-platform versions create {MODEL_VERSION} --model={MODEL_NAME} \
            --origin=gs://{BUCKET}/{GCS_MODEL_DIR} \
            --python-version=3.7 \
            --runtime-version={RUNTIME_VERSION} \
            --machine-type=mls1-c4-m4 \
            --package-uris=gs://{BUCKET}/{GCS_PACKAGE_URI},{PYTORCH_VISION_PACKAGE}\
            --prediction-class={MODEL_CLASS}

您可以将PYTORCH_VISION_PACKAGE 更改为我在#2 中提到的任何选项

【讨论】：

我也在使用更大的机器：mls1-c4-m4。几周后，自定义预测将支持我们支持的其他 Google 计算引擎实例。

【解决方案2】：

我可以通过调整 setup.py 来取得成功。基本上install_requires 尝试获取 PyPI 托管的torch 包，这是一个巨大的 GPU 构建轮，超过了部署配额。以下setup.py 注入了从官方 pytorch 索引中获取 CPU 构建的 Torch 的安装命令。

from setuptools import setup, find_packages
from setuptools.command.install import install as _install

INSTALL_REQUIRES = ['pillow']

CUSTOM_INSTALL_COMMANDS = [
    # Install torch here.
    [
        'python-default', '-m', 'pip', 'install', '--target=/tmp/custom_lib',
        '-b', '/tmp/pip_builds', 'torch==1.4.0+cpu', 'torchvision==0.5.0+cpu',
        '-f', 'https://download.pytorch.org/whl/torch_stable.html'
    ],
]

class Install(_install):
    def run(self):
        import sys
        if sys.platform == 'linux':
            import subprocess
            import logging
            for command in CUSTOM_INSTALL_COMMANDS:
                logging.info('Custom command: ' + ' '.join(command))
                result = subprocess.run(
                    command, check=True, stdout=subprocess.PIPE
                )
                logging.info(result.stdout.decode('utf-8', 'ignore'))
        _install.run(self)

setup(
    name='predictor',
    version='0.1',
    packages=find_packages(),
    install_requires=INSTALL_REQUIRES,
    cmdclass={'install': Install},
)

【讨论】：

我有同样的问题，但是当我根据你的更改我的 setup.py 时，我得到这个：错误：（gcloud.beta.ai-platform.versions.create）创建版本失败。检测到错误模型并出现错误：“处理用户代码时出现问题：无法找到 predictor.MyPredictor。请确保 (1) prediction_class 是完全限定的函数名称

【解决方案3】：

经过好几个小时的旧尝试错误，我得出了与@kyamagu 相同的结论，“install_requires 尝试获取 PyPI 托管的 Torch 包，它是一个巨大的 GPU 构建轮，超过了部署配额。”

但是，他的解决方案对我不起作用。因此，经过更多小时的试验错误（由于缺乏文档和错误的错误），我想出了这个解决方案：

我们需要获取大约 100 MB 的 Pytorch 的 cpu 构建轮，而不是默认托管 PyPI 的 700 MB GPU 构建。你可以在这里找到它们：https://download.pytorch.org/whl/cpu/torch_stable.html

接下来，我们需要将它们放在我们的 gs 存储中，然后将路径作为 --package-uris 的一部分提供，如下所示：

gcloud beta ai-platform versions create v17 \
    --model=newest \
    --origin=gs://bucket \
    --runtime-version=1.15 \
    --python-version=3.7 \
    --package-uris=gs://bucket/predictor-0.1.tar.gz,gs://bucket/torch-1.3.0+cpu-cp37-cp37m-linux_x86_64.whl \
    --prediction-class=predictor.MyPredictor \
    --machine-type=mls1-c4-m4

另外，注意package-uris 的顺序，predictor 包应该在前，逗号后面不能有空格。

希望这会有所帮助。干杯！

【讨论】：