ResourceExhaustedError: OOM when allocating tensor with shape[32,32,239,239] and type float

Question

我正在尝试使用不同的图像从 this paper(model 1) 重新创建 CNN 图像识别模型。然而，在第一个时期拟合模型 returns 我一个 ResourceExhaustedError 。批量大小已经相当小，所以我猜问题出在我从论文中复制的模型定义上。任何关于模型更改内容的建议都将不胜感激。谢谢！

#Load dataset
BATCH_SIZE = 32
IMG_SIZE = (244,244)
train_set = tf.keras.preprocessing.image_dataset_from_directory(
    main_dir, 
    shuffle = True,
    image_size = IMG_SIZE,
    batch_size = BATCH_SIZE)
val_set = tf.keras.preprocessing.image_dataset_from_directory(
    main_dir, 
    shuffle = True, 
    image_size = IMG_SIZE,
    batch_size = BATCH_SIZE)
class_names = train_set.class_names
print(class_names)

#Augment data by flipping image and random rotation
data_augmentation = tf.keras.Sequential([
    tf.keras.layers.experimental.preprocessing.RandomFlip('horizontal'),
    tf.keras.layers.experimental.preprocessing.RandomRotation(0.2),
])

#Model definition 
model = Sequential([
    data_augmentation,
    tf.keras.layers.experimental.preprocessing.Rescaling(1./255),
    Conv2D(filters=64,kernel_size=(4,4), activation='relu'),
    Conv2D(filters=32,kernel_size=(3,3), activation='relu'),
    AveragePooling2D(pool_size=(4,4)),

    Conv2D(filters=32,kernel_size=(3,3), activation='relu'),
    Conv2D(filters=32,kernel_size=(3,3), activation='relu'),
    Conv2D(filters=32,kernel_size=(3,3), activation='relu'),
    AveragePooling2D(pool_size=(2,2)),
    Flatten(),
    
    Dense(256, activation='relu'),
    Dense(256, activation='relu'),
    Dense(128, activation='relu'),
    Dense(128, activation='relu'),
    Dense(128, activation='tanh'),
    Dense(1, activation='softmax')

])

model.compile(optimizer='RMSprop',
              loss=keras.losses.CategoricalCrossentropy(from_logits=True),
              metrics=[keras.metrics.CategoricalAccuracy()])

history = model.fit(train_set,validation_data=val_set, epochs=150)

拟合模型后出错：

ResourceExhaustedError:  OOM when allocating tensor with shape[32,32,239,239] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc
     [[node gradient_tape/sequential_1/average_pooling2d/AvgPoolGrad (defined at <ipython-input-10-ef749d320491>:1) ]]

nvidia-smi

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 460.91.03    Driver Version: 460.91.03    CUDA Version: 11.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|                               |                      |               MIG M. |
|===============================+======================+======================|
|   0  GeForce 940MX       Off  | 00000000:01:00.0 Off |                  N/A |
| N/A   46C    P0    N/A /  N/A |   1938MiB /  2004MiB |      2%      Default |
|                               |                      |                  N/A |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                                  |
|  GPU   GI   CI        PID   Type   Process name                  GPU Memory |
|        ID   ID                                                   Usage      |
|=============================================================================|
|    0   N/A  N/A       959      G   /usr/lib/xorg/Xorg                 97MiB |
|    0   N/A  N/A      1270      G   /usr/bin/gnome-shell               25MiB |
|    0   N/A  N/A      4635      G   /usr/lib/firefox/firefox          212MiB |
|    0   N/A  N/A      5843      C   /usr/bin/python3                 1595MiB |
+-----------------------------------------------------------------------------+

Answer 1

错误是由您的 GPU 运行内存不足引起的。这可能是因为 1) 即使对于你的 GPU，你每个时期加载的数据太多，或者 2) 如果你碰巧有足够的 VRAM，另一个进程会在时期之间保留一些内存。这可能是因为 tensorflow 在运行时保留了 100% 的 VRAM。 You can limit amount of VRAM reserved to just what's needed.

import tensorflow as tf
for gpu in tf.config.list_physical_devices("GPU")
    tf.config.experimental.set_memory_growth(gpu, True)

编辑：看看你的显卡，我很确定你没有足够的 VRAM 来满足你的批量大小和输入大小（所以问题 1）。您应该降低批量大小。

ResourceExhaustedError: OOM when allocating tensor with shape[32,32,239,239] and type float

ResourceExhaustedError: OOM when allocating tensor with shape[32,32,239,239] and type float

python

tensorflow

keras

conv-neural-network