np.concatenate的内存错误

时间:2018-01-09 19:14:41

标签: python numpy tensorflow keras

当我在iPython笔记本中运行流动代码时:

_x = np.concatenate([_batches.next() for i in range(_batches.samples)])

我收到此错误消息

---------------------------------------------------------------
MemoryError                   Traceback (most recent call last)
<ipython-input-14-313ecf2ea184> in <module>()
----> 1 _x = np.concatenate([_batches.next() for i in 
range(_batches.samples)])

MemoryError:

迭代器有9200个元素。

next(_batch)返回一个np.array形状:(1,400,400,3)

我有30GB内存和16GB GPU。

当我在Keras中使用predict_generator()时,我遇到了类似的问题。我运行以下代码:

bottleneck_features_train = bottleneck_model.predict_generator(batches, len(batches), verbose=1) 

使用verbose = 1时,我可以看到进度指示器一直显示,但后来我收到以下错误:

2300/2300 [==============================] - 177s 77ms/step
---------------------------------------------------------------
MemoryError                   Traceback (most recent call last)
<ipython-input-19-d0e463f64f5a> in <module>()
----> 1 bottleneck_features_train = 
bottleneck_model.predict_generator(batches, len(batches), verbose=1)

~/anaconda3/lib/python3.6/site-packages/keras/legacy/interfaces.py in 
wrapper(*args, **kwargs)
     85                 warnings.warn('Update your `' + object_name +
     86                               '` call to the Keras 2 API: ' + 
signature, stacklevel=2)
---> 87             return func(*args, **kwargs)
     88         wrapper._original_function = func
     89         return wrapper

~/anaconda3/lib/python3.6/site-packages/keras/engine/training.py in 
predict_generator(self, generator, steps, max_queue_size, workers, 
use_multiprocessing, verbose)
   2345                 return all_outs[0][0]
   2346             else:
-> 2347                 return np.concatenate(all_outs[0])
   2348         if steps_done == 1:
   2349             return [out for out in all_outs]

MemoryError: 

请问这个内存问题的解决方案吗?谢谢!

1 个答案:

答案 0 :(得分:2)

对于第一个错误,数据太大了。假设数据类型为int64或float64(每个元素8个字节),则总数据为9200 * 400 * 400 * 3 * 8字节,即35GB。所有这些数据都以块的形式收集,然后通过串联复制成一个大数组。

你可以预先分配数组,也许它可以工作:

x_ = np.empty((9200,400,400,3))
for i in range(9200): 
    x_[i] = batches.next()