Problem in acc metric

[solin319]

In #7561, label and pred_label were convert to int32 in NDArray. This may slowdown the speed in train.
I use "train_imagenet.py" in example to train some models in 4 gpus in single machine., the batch-size is 128. The speed was shown as below.

                   convert in NDArray (samples/sec)             change to numpy first (samples/sec)
vgg16                      124                                             144
resnet-50                  183                                             185
resnet152                  77.8                                            78.3
inception-v3               116                                             117

[mli]

can we just use ndarray to compute instead of converting to numpy? if there is a number of classes, say 10K, then it could be problematic

a sample code:

def accuracy(output, label):
    # both output and label are ndarray
    return (output.argmax(axis=1)==label).sum().asscalar()