系统信息:

• 笔记本电脑
• 操作系统平台和发行版:Ubuntu Linux，18.04，x64
• TensorFlow 安装自:pip
• TensorFlow 版本:2.1.0
• Python版本:3.6.9
• GPU 型号和内存:nVidia RTX2060 6GB
• CPU型号:i7-9850H
• 内存:16GB

我在另一台 PC 的 CPU 上使用 TensorFlow 2.0。

我安装了(使用 https://www.tensorflow.org/install/gpu 处的指南)CUDA 10.1。

我开始使用 ResNet50V2 在包含 26998 个训练图像和 1000 个作为验证图像的数据集上运行用于神经网络的旧脚本，其中包含 2 个类别。

网络

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
keras_layer (KerasLayer)     (None, 1792)              4363712   
_________________________________________________________________
dense (Dense)                (None, 64)                114752    
_________________________________________________________________
dropout (Dropout)            (None, 64)                0         
_________________________________________________________________
dense_1 (Dense)              (None, 2)                 130       
=================================================================
Total params: 4,478,594
Trainable params: 114,882
Non-trainable params: 4,363,712
_________________________________________________________________

其中 keras_layer 是从 tensorflow_hub 得到的 resnet。

作为第一期，我得到了一个 CUDA_ERROR_OUT_OF_MEMORY 我解决了添加

physical_devices = tf.config.experimental.list_physical_devices('GPU')
for dev in physical_devices:
  try:
    tf.config.experimental.set_memory_growth(dev, True)
    print(dev, "SET MEMORY GROWTH")
  except:
    print("Device config error")
    sys.exit(1)

但是现在我收到了类似的警告:

2020-04-07 01:39:57.857284: I tensorflow/stream_executor/cuda/cuda_driver.cc:801] failed to allocate 2.70G (2897281024 bytes) from device: CUDA_ERROR_OUT_OF_MEMORY: out of memory

2020-04-07 01:39:58.035192: W tensorflow/core/common_runtime/bfc_allocator.cc:309] Garbage collection: deallocate free memory regions (i.e., allocations) so that we can re-allocate a larger region to avoid OOM due to memory fragmentation. If you see this message frequently, you are running near the threshold of the available device memory and re-allocation may incur great performance overhead. You may try smaller batch sizes to observe the performance impact. Set TF_ENABLE_GPU_GARBAGE_COLLECTION=false if you'd like to disable this feature.

都打印了几次。

在此之后我得到:

2020-04-07 01:41:59.069302: W tensorflow/core/kernels/data/generator_dataset_op.cc:103] Error occurred when finalizing GeneratorDataset iterator: Cancelled: Operation was cancelled

我读到它们没有关系，但我不清楚是什么导致了第二次警告。

最后是这样的:

WARNING:tensorflow:sample_weight modes were coerced from
  ...
    to  
  ['...']

(我认为它们是由三个不同的问题引起的，我决定将所有问题都发布在一个问题中以避免垃圾邮件，但如果这是一个问题，我可以分成不同的线程。)

我使用 ImageDataGenerator 生成数据集:

train_image_generator = ImageDataGenerator(rescale=1./255., rotation_range=10., horizontal_flip=True) # Generator for our training data
validation_image_generator = ImageDataGenerator(rescale=1./255.) # Generator for our validation data

train_data_gen = train_image_generator.flow_from_directory(batch_size=batch_size,
                                                        directory=train_dir,
                                                        shuffle=True,
                                                        target_size=(IMG_H, IMG_W),
                                                        class_mode='sparse')

validation_data_gen = validation_image_generator.flow_from_directory(batch_size=batch_size,
                                                          directory=validation_dir,
                                                          shuffle=True,
                                                          target_size=(IMG_H, IMG_W),
                                                          class_mode='sparse')

如果需要其他代码，我会添加。

谢谢。

编辑 1:

对于警告:

2020-04-07 01:41:59.069302: W tensorflow/core/kernels/data/generator_dataset_op.cc:103] Error occurred when finalizing GeneratorDataset iterator: Cancelled: Operation was cancelled

我试图在 fit() 中设置 workers=1 并且它消失了，但我仍然不知道这个警告的原因和后果。

最佳答案

此错误是由于您之前运行程序而导致 GPU 已被占用。现在，当您尝试重新运行时，没有内存可用于再次占用模型。

执行以下操作 -

1. 打开终端并输入 nivida-smi
2. 找到占用您 GPU 的进程 ID (PID)
3. 使用 kill -9 PID 杀死占用 gpu 的进程 (PID)

Note - You can also kill process using top

https://stackoverflow.com/questions/61070699/