在 docker 容器中运行 tensorflow 时出错

Question

我正在尝试在 Docker 容器中运行的 Python 应用程序中使用 Tensorflow 模块（实际上我使用的是 Keras，但错误来自 Tensorflow）

我想加载模型（.json 和 .h5 文件）以便使用它：

import logging
import os
from keras.models import model_from_json # library for machine learning
from numpy import array
import json

def load_models():
    global loaded_h_model
    global loaded_u_model
    global loaded_r_model
    global loaded_c_model

    modelPath = os.getenv("MODELPATH", "./models/")

    # load models
    json_h_file = open(modelPath+'model_HD.json', 'r')
    loaded_model_h_json = json_h_file.read()
    json_h_file.close()
    loaded_h_model = model_from_json(loaded_model_h_json)
    loaded_h_model.load_weights(modelPath+"model_HD.h5")

    json_u_file = open(modelPath+'model_UD.json', 'r')
    loaded_model_u_json = json_u_file.read()
    json_u_file.close()
    loaded_u_model = model_from_json(loaded_model_u_json)
    loaded_u_model.load_weights(modelPath+"model_UD.h5")

    json_r_file = open(modelPath+'model_RD.json', 'r')
    loaded_model_r_json = json_r_file.read()
    json_r_file.close()
    loaded_r_model = model_from_json(loaded_model_r_json)
    loaded_r_model.load_weights(modelPath+"model_RD.h5")

    json_c_file = open(modelPath+'model_CD.json', 'r')
    loaded_model_c_json = json_c_file.read()
    json_c_file.close()
    loaded_c_model = model_from_json(loaded_model_c_json)
    loaded_c_model.load_weights(modelPath+"model_CD.h5")

这是我使用的 Dockerfile：

FROM python:3.7

# copy source code files
COPY machinelearning.py ./

# copy models files
COPY models/* ./models/

# install dependencies
RUN pip3 install --upgrade pip \
    && pip3 install h5py \
    && pip3 install tensorflow \
    && pip3 install keras

# run script
CMD [ "python", "./machinelearning.py" ]

但是当我运行 Docker 容器时，我有以下警告/错误：

2020-01-29 09:40:24.542588: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'libnvinfer.so.6'; dlerror: libnvinfer.so.6: cannot open shared object file: No such file or directory
2020-01-29 09:40:24.542727: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'libnvinfer_plugin.so.6'; dlerror: libnvinfer_plugin.so.6: cannot open shared object file: No such file or directory
2020-01-29 09:40:24.542743: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:30] Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.
Using TensorFlow backend.
2020-01-29 09:40:25.394254: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'libcuda.so.1'; dlerror: libcuda.so.1: cannot open shared object file: No such file or directory
2020-01-29 09:40:25.394289: E tensorflow/stream_executor/cuda/cuda_driver.cc:351] failed call to cuInit: UNKNOWN ERROR (303)
2020-01-29 09:40:25.394321: I tensorflow/stream_executor/cuda/cuda_diagnostics.cc:156] kernel driver does not appear to be running on this host (dd231f397f1f): /proc/driver/nvidia/version does not exist
2020-01-29 09:40:25.394539: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2020-01-29 09:40:25.419513: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 1992000000 Hz
2020-01-29 09:40:25.420250: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x55cab5bf9760 initialized for platform Host (this does not guarantee that XLA will be used). Devices:
2020-01-29 09:40:25.420299: I tensorflow/compiler/xla/service/service.cc:176]   StreamExecutor device (0): Host, Default Version

我认为我需要在我的 Dockerfile 中安装库或不同版本的 Tensorflow/Keras。

我该如何解决这个问题？ 谢谢

Answer 1

首先，您需要COPY requirements.txt /to/destination。你的 requirements.txt 应该包含 版本号的依赖关系。

FROM python:latest
COPY requirements.txt /usr/src/code/

之后运行

RUN pip3 install -r requirements.txt

而不是 Dockerfile 中的以下代码

RUN pip3 install --upgrade pip \
    && pip3 install h5py \
    && pip3 install tensorflow \
    && pip3 install keras 

我希望通过在 requirements.txt 中提及版本号来解决问题，而不仅仅是 --upgrade 标签。

如果不需要，也不要运行升级。