无法从keras.engine导入get_source_inputs的错误解决求助

问题场景 #

运行以下Python代码时触发ImportError：

import os
import cv2
import numpy as np
from matplotlib import pyplot as plt
from patchify import patchify
from PIL import Image
import segmentation_models as sm
from tensorflow.keras.metrics import MeanIoU

错误信息 #

ImportError: cannot import name 'get_source_inputs' from 'keras.engine' (C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\keras\engine\__init__.py)

完整报错回溯 #

2022-09-20 07:56:06.902214: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'cudart64_110.dll'; dlerror: cudart64_110.dll not found
2022-09-20 07:56:06.902313: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.
Traceback (most recent call last):
  File "C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\spyder_kernels\py3compat.py", line 356, in compat_exec
    exec(code, globals, locals)

  File "c:\users\xpapa\untitled2.py", line 38, in <module>
    import segmentation_models as sm

  File "C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\segmentation_models\__init__.py", line 5, in <module>
    from .unet import Unet

  File "C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\segmentation_models\unet\__init__.py", line 1, in <module>
    from .model import Unet

  File "C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\segmentation_models\unet\model.py", line 4, in <module>
    from ..backbones import get_backbone, get_feature_layers

  File "C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\segmentation_models\backbones\__init__.py", line 1, in <module>
    from classification_models import Classifiers

  File "C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\classification_models\__init__.py", line 3, in <module>
    from . import resnet as rn

  File "C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\classification_models\resnet\__init__.py", line 1, in <module>
    from .models import *

  File "C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\classification_models\resnet\models.py", line 1, in <module>
    from .builder import build_resnet

  File "C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\classification_models\resnet\builder.py", line 11, in <module>
    from keras.engine import get_source_inputs

ImportError: cannot import name 'get_source_inputs' from 'keras.engine' (C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\keras\engine\__init__.py)

解决方法 #

错误根源是classification_models库调用了旧版Keras的API，而新版TensorFlow/Keras中get_source_inputs的模块路径已变更，以下是三种可行解决方案：

• 修改库源码：找到文件C:\Users\xpapa\.conda\envs\TensorFlow\lib\site-packages\classification_models\resnet\builder.py，将第11行的from keras.engine import get_source_inputs替换为from keras.utils.layer_utils import get_source_inputs，保存后重新运行代码。
• 安装适配版库：执行命令安装兼容TensorFlow 2.x的分支版本：pip install git+https://github.com/qubvel/classification_models.git
• 降级Keras版本：如果项目对Keras版本无特殊要求，执行命令降级到兼容版本：pip install keras==2.3.1

内容的提问来源于stack exchange，提问作者XPAPA