已安装SciPy 1.5.3仍触发ImportError：TensorFlow图像生成代码报错排查与解决

Problem Description #

I'm running an image generation script using TensorFlow and NumPy, but keep hitting an ImportError even though I've installed SciPy 1.5.3. Here's my code:

def img_gen(pair_list, file):
    image = []
    for loc in pair_list:
        shutil.copy(loc, target_folder)
    img_gen = ImageDataGenerator(rescale=1 / 255)
    img_tranform = img_gen.flow_from_directory(temp_folder, target_size=(112, 112), batch_size=1, class_mode='sparse', shuffle=False, color_mode='rgb', seed=42)
    x_data, y_data = next(img_tranform)
    image.append(x_data)
    remove_file = os.path.join(target_folder,file)
    remove_file = remove_file.replace("\\", "/")
    os.remove(remove_file)
    return image

img_class = []
img_pairs = []
for folders1, subfolders1, filenames1 in os.walk(orig_img_dir):
    for folders2, subfolders2, filenames2 in os.walk(segm_img_dir):
        for file1 in filenames1:
            for file2 in filenames2:
                if file1.endswith('.png') and file1==file2:
                    audit_class = folders1.partition('\\')
                    audit_class = audit_class[-1]
                    img_class.append(audit_class)
                    file1_loc = os.path.join(folders1, file1)
                    file2_loc = os.path.join(folders2, file2)
                    img_pair = [file1_loc, file2_loc]
                    con_image = img_gen(img_pair, file1)
                    img = tf.concat([con_image[0][0], con_image[1][0]], axis=-1)
                    img_pairs.append(img)

When I run this, I get the following error:

C:\Program Files\WinPython64-3.8.6\python-3.8.6.amd64\lib\site-packages\keras_preprocessing\image\affine_transformations.py in apply_affine_transform(x, theta, tx, ty, shear, zx, zy, row_axis, col_axis, channel_axis, fill_mode, cval, order)
279     """
280     if scipy is None:
---> 281         raise ImportError('Image transformations require SciPy. Install SciPy.')
282     transform_matrix = None

ImportError: Image transformations require SciPy. Install SciPy.

I already have SciPy 1.5.3 installed, so what's causing this and how can I fix it?

Root Causes & Solutions #

This error typically pops up when the keras_preprocessing library can't access your SciPy installation in the active Python environment. Here are the most likely fixes based on common scenarios:

1. SciPy isn't installed in your WinPython environment #

You might have installed SciPy in a different Python environment, but your script is running in the WinPython 3.8.6 instance.

Fix:

• Open the WinPython Command Prompt from your installation folder
• Run pip list to check if SciPy 1.5.3 appears in the list. If not, install it with:

  pip install scipy==1.5.3
  

• If it is listed, reinstall to fix potential corrupted files:

  pip uninstall scipy -y
  pip install scipy==1.5.3
  

2. Version mismatch between keras_preprocessing and SciPy #

Older versions of keras_preprocessing can have trouble detecting specific SciPy versions like 1.5.3.

Fix:
Upgrade the keras_preprocessing library to a version compatible with Python 3.8:

pip install --upgrade keras-preprocessing

3. Your script is using the wrong Python interpreter #

Your system might be pointing to a different Python installation instead of the WinPython one you intended.

Fix:
Add these debug lines at the top of your script to verify the environment:

import scipy
import sys
print(f"Python executable path: {sys.executable}")
print(f"SciPy version: {scipy.__version__}")
print(f"SciPy installation path: {scipy.__file__}")

• If the path doesn't match your WinPython directory (C:\Program Files\WinPython64-3.8.6\python-3.8.6.amd64\...), switch to the WinPython environment to run your script.

4. Missing underlying SciPy dependencies (rare) #

SciPy relies on low-level libraries like OpenBLAS or MKL. If these are missing, keras_preprocessing might fail to detect SciPy.

Fix:
Use WinPython's built-in package manager to reinstall SciPy (it handles dependencies automatically) or download a pre-built wheel for Python 3.8/Windows 64-bit from PyPI and install it manually.

内容的提问来源于stack exchange，提问作者Raj Rajeshwari Prasad