Keras model.predict 函数写入文件（而不是内存）

Question

I have a fairly large image dataset and I want to know if there's a way I can make keras model.predict 'dump' results on a file on my disk intead of loading a large numpy array on memory, to avoid out of memory issues.

I know I can split the file in small batches and process each batch separately, but I'm wondering if I can alter keras (or python) behavior to direct the result to file (maybe a pickle object?).

Answer 1

You can probably use np.savetxt to save the ndarray returned by model.predict directly,

np.savetxt( 'results.gz' , model.predict( test_images ) )

Also, the documentation for np.savetxt suggests,

If the filename ends in .gz, the file is automatically saved in compressed gzip format. loadtxt understands gzipped files transparently.