调用keras.Sequential.predict时Python无报错崩溃，退出码异常

问题现象 #

运行基于TensorFlow的GAN代码时，Python无报错直接崩溃，仅输出以下日志：

2022-07-16 09:11:13.307094: I tensorflow/core/platform/cpu_feature_guard.cc:151] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations: AVX AVX2 To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2022-07-16 09:11:13.832861: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1525] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 1342 MB memory: -> device: 0, name: NVIDIA GeForce MX330, pci bus id: 0000:02:00.0, compute capability: 6.1
2022-07-16 09:11:13.911731: W tensorflow/core/kernels/data/cache_dataset_ops.cc:768] The calling iterator did not fully read the dataset being cached. In order to avoid unexpected truncation of the dataset, the partially cached contents of the dataset will be discarded. This can happen if you have an input pipeline similar to dataset.cache().take(k).repeat(). You should use dataset.take(k).cache().repeat() instead.
Process finished with exit code -1073740791 (0xC0000409)

代码如下：

# Bringing in tensorflow
import tensorflow as tf
# Brining in tensorflow datasets for fashion mnist
import tensorflow_datasets as tfds
# Bringing in matplotlib for viz stuff
from matplotlib import pyplot as plt
# Do some data transformation
import numpy as np
# Bring in the sequential api for the generator and discriminator
from keras import Sequential
# Bring in the layers for the neural network
from keras.layers import Conv2D, Dense, Flatten, Reshape, LeakyReLU, Dropout, UpSampling2D


# Scale and return images only
def scale_images(data):
    image = data['image']
    return image / 255


def build_generator():
    model = Sequential()

    # Takes in random values and reshapes it to 7x7x128
    # Beginnings of a generated image
    model.add(Dense(7 * 7 * 128, input_dim=128))
    model.add(LeakyReLU(0.2))
    model.add(Reshape((7, 7, 128)))

    # Upsampling block 1
    model.add(UpSampling2D())
    model.add(Conv2D(128, 5, padding='same'))
    model.add(LeakyReLU(0.2))

    # Upsampling block 2
    model.add(UpSampling2D())
    model.add(Conv2D(128, 5, padding='same'))
    model.add(LeakyReLU(0.2))

    # Convolutional block 1
    model.add(Conv2D(128, 4, padding='same'))
    model.add(LeakyReLU(0.2))

    # Convolutional block 2
    model.add(Conv2D(128, 4, padding='same'))
    model.add(LeakyReLU(0.2))

    # Conv layer to get to one channel
    model.add(Conv2D(1, 4, padding='same', activation='sigmoid'))

    return model


gpus = tf.config.experimental.list_physical_devices('GPU')
for gpu in gpus:
    tf.config.experimental.set_memory_growth(gpu, True)

# Use the tensorflow datasets api to bring in the data source
ds = tfds.load('fashion_mnist', split='train')

var = ds.as_numpy_iterator().next()['label']

# Setup connection aka iterator
data_iterator = ds.as_numpy_iterator()


# Getting data out of the pipeline
var2 = data_iterator.next()['image']


# Setup the subplot formatting
fig, ax = plt.subplots(ncols=4, figsize=(20, 20))
# Loop four times and get images
for idx in range(4):
    # Grab an image and label
    sample = data_iterator.next()
    # Plot the image using a specific subplot
    ax[idx].imshow(np.squeeze(sample['image']))
    # Appending the image label as the plot title
    ax[idx].title.set_text(sample['label'])
# plt.show()


# Reload the dataset
ds = tfds.load('fashion_mnist', split='train')
# Running the dataset through the scale_images preprocessing step
ds = ds.map(scale_images)
# Cache the dataset for that batch
ds = ds.cache()
# Shuffle it up
ds = ds.shuffle(60000)
# Batch into 128 images per sample
ds = ds.batch(128)
# Reduces the likelihood of bottlenecking
ds = ds.prefetch(64)

generator = build_generator()

# Generate new fashion
img = generator.predict(np.random.randn(4, 128, 1)) # crash happens here

教程评论区有用户在同一行遇到类似错误：

UnimplementedError: Graph execution error: Node: 'sequential/conv2d/Conv2D' DNN library is not found. [[{{node sequential/conv2d/Conv2D}}]] [Op:__inference_predict_function_140492]

使用NVIDIA GeForce MX330显卡（仅2GB显存），调整生成器参数后问题仍存在；已手动安装CUDA和cudNN，无CUDA相关警告；尝试不同脚本均在调用predict或类似步骤时崩溃。后续从命令行启动程序时发现关键错误：

Could not locate zlibwapi.dll. Please make sure it is in your library path!

解决方法 #

1. 修复zlibwapi.dll缺失问题 #

• 从NVIDIA官方CUDA工具包安装路径中提取zlibwapi.dll（通常位于C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\vX.X\bin，X.X为CUDA版本号）
• 将该文件复制到以下任意位置：
  • 系统目录（如C:\Windows\System32）
  • Python环境的Lib/site-packages/tensorflow目录
  • 添加该文件所在路径到系统环境变量PATH中

2. 优化显存占用（适配低显存显卡） #

• 减少模型通道数：将生成器中所有128通道数改为64或32，例如：

  def build_generator():
      model = Sequential()
      model.add(Dense(7 * 7 * 64, input_dim=128))
      model.add(LeakyReLU(0.2))
      model.add(Reshape((7, 7, 64)))
      model.add(UpSampling2D())
      model.add(Conv2D(64, 5, padding='same'))
      model.add(LeakyReLU(0.2))
      # 后续Conv2D层同理修改通道数
  

• 减小Batch Size：将ds.batch(128)改为ds.batch(32)或ds.batch(16)
• 强制使用CPU运行：若显存仍不足，在代码开头添加：

  tf.config.set_visible_devices([], 'GPU')
  

3. 验证CUDA环境完整性 #

• 确认TensorFlow版本与CUDA版本匹配（如TensorFlow 2.10+对应CUDA 11.2+）
• 确认cudNN版本与CUDA版本兼容，并已正确放置到CUDA安装目录的include和lib/x64文件夹中

内容的提问来源于stack exchange，提问作者stefanp99