编译器在编译时会自动使用GPU还是必须手动使用它？

Question

I have been compiling this program for bird image preprocessing on my pc with 8GB RAM, intel core i7-4th gen. processor but everytime is gets stuck after running for a while showing this error:

File "bird.py", line 13, in <module>
    X,Y,X_test,Y_test = pickle.load(open("full_dataset.pkl", "rb"))
MemoryError

Being a newbie to tensorflow, I am not sure whether CUDA (which I've installed successfully) is being used or is it just the CPU still performing the computations.

Answer 1

The error you are seeing is raised before any GPU computation (or indeed any TensorFlow computation) has been issued. The MemoryError implies that unpickling the dataset is either leading to running out of memory, or that there is some corruption in the data. Try loading a smaller dataset to see if the problem persists.

As to your original question, if you have installed the GPU version of TensorFlow, and a compatible version of CUDA (ie CUDA 8.0) and cuDNN, TensorFlow will offload much of the computation to your GPU, including the matrix multiplications and convolutions that will comprise the bulk of the time spent during training.