使用TensorFlow训练某些较大模型时会发生内存溢出,如果 已经安装了TensorFlow-GPU版本,训练时会优先调用GPU版本的TensorFlow,而一般电脑上显存比较小,很容易发生溢出,就会出现如下报错:

ResourceExhaustedError: OOM when allocating tensor with shape[1024,728,1,1] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc [[node model/block13_sepconv2/separable_conv2d (defined at <ipython-input-41-425b3e9b7078>:11) ]] Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info. [Op:__inference_train_function_41706] Function call stack: train_function

 

ResourceExhaustedError:  OOM when allocating tensor with shape[1024,728,1,1] and type float on /job:localhost/replica:0/task:0/device:GPU:0 by allocator GPU_0_bfc
	 [[node model/block13_sepconv2/separable_conv2d (defined at <ipython-input-41-425b3e9b7078>:11) ]]
Hint: If you want to see a list of allocated tensors when OOM happens, add report_tensor_allocations_upon_oom to RunOptions for current allocation info.
 [Op:__inference_train_function_41706]

Function call stack:
train_function

解决方案:

1. 尝试使用CPU进行训练,将model.fit()代码做如下修改:

with tf.device("/cpu:0"):
    history = model.fit(替换成自己的代码)

输出:

Epoch 1/50
21/86 [======>.......................] - ETA: 16:08 - loss: 0.4574 - accuracy: 0.8438

这样就可以使用CPU进行训练了。

2. 如果是在jupyter notebook中运行代码,则尝试先运行下面的代码:

tf.keras.backend.clear_session()

如果在notebook中运行了很多代码,则会占用一定的内存,上面的代码顾名思义就是清楚掉之前运行的一些session,以释放空间。

3. 如果还是不行,则只能修改代码,将批次数batch_size改小一些,每次给模型喂入小批量的数据。

 

 

 

 

 

 

Logo

开放原子开发者工作坊旨在鼓励更多人参与开源活动,与志同道合的开发者们相互交流开发经验、分享开发心得、获取前沿技术趋势。工作坊有多种形式的开发者活动,如meetup、训练营等,主打技术交流,干货满满,真诚地邀请各位开发者共同参与!

更多推荐