Keras数据增强中出现SciPy未定义的NameError问题求助

问题 #

在Conda环境miniconda3/envs/tf中已成功安装SciPy，Python脚本开头也执行了import scipy，但使用Keras的ImageDataGenerator做数据增强时，调用gen_img[0]触发NameError，提示name 'scipy' is not defined。已确认SciPy安装状态正常，但错误持续。

相关代码：

import scipy
gen = ImageDataGenerator(rotation_range=10)
img = np.array(cv2.imread(img_path))
gen_img = gen.flow(np.expand_dims(img, axis=0))
plt.imshow(gen_img[0])

报错栈追踪：

---------------------------------------------------------------------------
NameError                                 Traceback (most recent call last)
Cell In[52], line 1
----> 1 gen_img[0]

File c:\Users\ADMIN\miniconda3\envs\tf\lib\site-packages\keras\preprocessing\image.py:116, in Iterator.__getitem__(self, idx)
    112     self._set_index_array()
    113 index_array = self.index_array[
    114     self.batch_size * idx : self.batch_size * (idx + 1)
    115 ]
--> 116 return self._get_batches_of_transformed_samples(index_array)

File c:\Users\ADMIN\miniconda3\envs\tf\lib\site-packages\keras\preprocessing\image.py:801, in NumpyArrayIterator._get_batches_of_transformed_samples(self, index_array)
    799 x = self.x[j]
    800 params = self.image_data_generator.get_random_transform(x.shape)
--> 801 x = self.image_data_generator.apply_transform(
    802     x.astype(self.dtype), params
    803 )
    804 x = self.image_data_generator.standardize(x)
    805 batch_x[i] = x

File c:\Users\ADMIN\miniconda3\envs\tf\lib\site-packages\keras\preprocessing\image.py:2013, in ImageDataGenerator.apply_transform(self, x, transform_parameters)
   2010 img_col_axis = self.col_axis - 1
   2011 img_channel_axis = self.channel_axis - 1
--> 2013 x = apply_affine_transform(
   2014     x,
   2015     transform_parameters.get("theta", 0),
   2016     transform_parameters.get("tx", 0),
   2017     transform_parameters.get("ty", 0),
   2018     transform_parameters.get("shear", 0),
   2019     transform_parameters.get("zx", 1),
   2020     transform_parameters.get("zy", 1),
   2021     row_axis=img_row_axis,
   2022     col_axis=img_col_axis,
   2023     channel_axis=img_channel_axis,
   2024     fill_mode=self.fill_mode,
   2025     cval=self.cval,
   2026     order=self.interpolation_order,
   2027 )
   2029 if transform_parameters.get("channel_shift_intensity") is not None:
   2030     x = apply_channel_shift(
   2031         x,
   2032         transform_parameters["channel_shift_intensity"],
   2033         img_channel_axis,
   2034     )

File c:\Users\ADMIN\miniconda3\envs\tf\lib\site-packages\keras\preprocessing\image.py:2529, in apply_affine_transform(x, theta, tx, ty, shear, zx, zy, row_axis, col_axis, channel_axis, fill_mode, cval, order)
   2485 @keras_export("keras.preprocessing.image.apply_affine_transform")
   2486 def apply_affine_transform(
   2487     x,
   (...)
   2499     order=1,
   2500 ):
   2501     """Applies an affine transformation specified by the parameters given.
   2502 
   2503     Args:
   (...)
   2527         ImportError: if SciPy is not available.
   2528     """
--> 2529     if scipy is None:
   2530         raise ImportError(
   2531             "Image transformations require SciPy. " "Install SciPy."
   2532         )
   2534     # Input sanity checks:
   2535     # 1. x must 2D image with one or more channels (i.e., a 3D tensor)
   2536     # 2. channels must be either first or last dimension

NameError: name 'scipy' is not defined

原因分析 #

这个错误并非你的代码问题，而是Keras内部的apply_affine_transform函数存在导入bug：

• 函数内部直接使用了scipy变量，但既没有在函数内部导入SciPy，也没有在模块顶部完成正确导入；
• 你自己脚本里的import scipy仅在当前脚本的命名空间生效，无法被Keras内部函数访问。

解决方法 #

方法1：降级Keras到无bug的稳定版本 #

该bug常见于部分较新的Keras 2.x版本，可降级到经过验证的稳定版本：

conda activate tf
pip install keras==2.10.0

方法2：手动修改Keras源码（临时应急） #

找到报错文件c:\Users\ADMIN\miniconda3\envs\tf\lib\site-packages\keras\preprocessing\image.py，在apply_affine_transform函数开头添加SciPy导入：

def apply_affine_transform(...):
    import scipy  # 添加此行
    if scipy is None:
        raise ImportError(...)
    # 函数其余代码保持不变

方法3：改用TensorFlow官方图像增强API #

放弃Keras旧版ImageDataGenerator，使用TensorFlow原生的增强模块，无需依赖SciPy：

import tensorflow as tf
import cv2
import matplotlib.pyplot as plt
import numpy as np

# 定义增强层
data_augmentation = tf.keras.Sequential([
    tf.keras.layers.RandomRotation(factor=0.1)  # 对应原rotation_range=10
])

img = np.array(cv2.imread(img_path))
img = tf.expand_dims(img, 0)  # 增加batch维度
gen_img = data_augmentation(img)
plt.imshow(gen_img[0].numpy().astype("uint8"))
plt.show()

内容的提问来源于stack exchange，提问作者pragateesh