ZhennanQin opened a new issue #15078: Naive engine produce incorrect result on MKLDNN backend URL: https://github.com/apache/incubator-mxnet/issues/15078 Reproducible test: ``` import numpy as np import mxnet as mx from mxnet import gluon, nd, image from mxnet.gluon.data.vision import transforms from gluoncv.model_zoo import get_model ctx = [mx.cpu()] # Load Model model_name = "cifar_resnet20_v1" kwargs = {'classes': 10, 'pretrained': True} net = get_model(model_name, **kwargs) net.hybridize(static_alloc=True, static_shape=True) def test(ctx, val_data): metric = mx.metric.Accuracy() for i, batch in enumerate(val_data): data = gluon.utils.split_and_load(batch[0], ctx_list=ctx, batch_axis=0) label = gluon.utils.split_and_load(batch[1], ctx_list=ctx, batch_axis=0) outputs = [net(X) for X in data] metric.update(label, outputs) break return metric.get() transform_test = transforms.Compose([ transforms.ToTensor(), transforms.Normalize([0.4914, 0.4822, 0.4465], [0.2023, 0.1994, 0.2010]) ]) val_data = gluon.data.DataLoader( gluon.data.vision.CIFAR10(train=False).transform_first(transform_test), batch_size=128, shuffle=False, num_workers=2) name, val_acc = test(ctx, val_data) print('val=%f' % val_acc) ``` With threaded_engine, result is correct: ``` $ unset MXNET_ENGINE_TYPE $ python small.py val=0.929688 ``` with naive_engine, result is incorrect: ``` $ export MXNET_ENGINE_TYPE=NaiveEngine $ python small.py val=0.085938 ``` If remove below line, the issue can get fix: https://github.com/apache/incubator-mxnet/blob/master/src/operator/nn/mkldnn/mkldnn_convolution.cc#L418 `weight.MKLDNNDataReorderAsync(fwd->fwd_pd.weights_primitive_desc());` The reason behind is, naive_engine won't do this as lazy as threaded_engine, causing different execution order, and finally make result incorrect.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
