Cannot train custom object detector when using mxnet-mkl

Hi there,
I installed mxnet with MKL-DNN by running:

pip install --upgrade mxnet-mkl gluoncv

I tried reinstalling it just now, so I’m getting the latest version of the packages.

Everything works fine, and I’m getting a much better performance when running inference. However, whenever I try to finetune a pre-trained model (basically, I’m using a modified version of the scripts provided in the Tutorials), I get the error:

!IsMKLDNNData(): We can't generate TBlob for MKLDNN data. Please use Reorder2Default() to generate a new NDArray first

If I switch to a different environment, where I installed native mxnet (pip install --upgrade mxnet gluoncv) the training process goes on without any errors.
Here another user had a similar problem and in the answers it was said the issue had been fixed. Am I doing something wrong? Is the MKL-DNN supposed to be used only for inference?

Hi @LewsTherin511, thank you for reporting the issue. It’s really great to hear that your get performance boost with mxnet-mkl.
No, MKL-DNN is not only for inference. It should also work for training. To address the issue, could you please share the version of your mxnet? Is it possible for you to try the latest nightly release? You can install it by:
pip install --pre mxnet -f
The MKL-DNN backend has been enabled by default in the mxnet package.

If the problem still exists, could you please share a simple reproducer?

@TaoLv, thanks for your reply!

In the environment where I was getting the error reported in the first post, I have:

mxnet-mkl -> 1.6.0
gluoncv -> 0.6.0

Also, I created a new environment where I installed the nightly version as you suggested, and added gluoncv by the usual pip install --upgrade gluoncv…so, in this environment now I have:

mxnet -> 2.0.0b20200513
gluoncv -> 0.7.0

but I’m still getting the same error:

File "../src/ndarray/", line 840
MXNetError: Check failed: !IsMKLDNNData(): We can't generate TBlob for MKLDNN data. Please use Reorder2Default() to generate a new NDArray first

The error is returned from the line:

sum_loss, cls_loss, box_loss = mbox_loss(cls_preds, box_preds, cls_targets, box_targets)

in the code:

import mxnet as mx
from mxnet import autograd, gluon
import gluoncv as gcv
from gluoncv import data as gdata
from gluoncv.utils import download, viz
from import Tuple, Stack, Pad
from import SSDDefaultValTransform
from import SSDDefaultTrainTransform
from gluoncv.utils.metrics.coco_detection import COCODetectionMetric
from gluoncv.utils.metrics.voc_detection import VOC07MApMetric
from mxboard import SummaryWriter

def main():

    ctx = [mx.cpu()]

    model_name = "ssd_512_resnet50_v1_coco"
    classes = ['Guitar', 'human_face']
    batch_size = 8
    net = gcv.model_zoo.get_model(model_name, pretrained=True)
    saved_weights_path = "saved_weights/random_test/"

    # prepare data
    data_shape = 512
    train_dataset ='custom_dataset/train_guitar_face_00.rec', coord_normalized=True)
    val_dataset  ='custom_dataset/test_guitar_face_00.rec', coord_normalized=True)

    # summary file for tensorboard
    sw = SummaryWriter(logdir=saved_weights_path+'logs/', flush_secs=30)

    eval_metric = VOC07MApMetric(iou_thresh=0.5, class_names=classes)

    # create data batches from dataset (net, train_dataset, data_shape, batch_size, num_workers):
    train_data, val_data = get_dataloader(net, train_dataset, val_dataset, data_shape, batch_size, num_workers=0)

    # ---------------------
    #   Training
    # ---------------------
    # configuration
    trainer = gluon.Trainer(net.collect_params(), 'sgd', {'learning_rate': 0.001, 'wd': 0.0005, 'momentum': 0.9})
    mbox_loss = gcv.loss.SSDMultiBoxLoss()
    ce_metric = mx.metric.Loss('CrossEntropy')
    smoothl1_metric = mx.metric.Loss('SmoothL1')

    for epoch in range(0, 100):
        tic = time.time()
        btic = time.time()
        net.hybridize(static_alloc=True, static_shape=True)
        loss1 = 0
            loss2 = 0
            name1 = ''
            name2 = ''
            for i, batch in enumerate(train_data):
                batch_size = batch[0].shape[0]
                data = gluon.utils.split_and_load(batch[0], ctx_list=ctx, batch_axis=0)
                cls_targets = gluon.utils.split_and_load(batch[1], ctx_list=ctx, batch_axis=0)
                box_targets = gluon.utils.split_and_load(batch[2], ctx_list=ctx, batch_axis=0)
                with autograd.record():
                    cls_preds = []
                    box_preds = []
                    for x in data:
				cls_pred, box_pred, _ = net(x)
                    sum_loss, cls_loss, box_loss = mbox_loss(cls_preds, box_preds, cls_targets, box_targets)
                # since we have already normalized the loss, we don't want to normalize by batch-size anymore
                ce_metric.update(0, [l * batch_size for l in cls_loss])
                smoothl1_metric.update(0, [l * batch_size for l in box_loss])
                name1, loss1 = ce_metric.get()
                name2, loss2 = smoothl1_metric.get()

                if i % 20 == 0:
                    print('[Epoch {}][Batch {}], Speed: {:.3f} samples/sec, {}={:.3f}, {}={:.3f}'.format(
				epoch, i, batch_size / (time.time() - btic), name1, loss1, name2, loss2))
                btic = time.time()

            print(f"\nValidation accuracies for epoch: {epoch}")
            for k,v in zip(map_name, mean_ap):
                sw.add_scalar(tag=f'{k}_val_acc', value=v, global_step=epoch)
            # save model parameters for inference every 5 epochs
            if epoch%5==0:
                net.export(f"{saved_weights_path}", epoch=epoch)

def get_dataloader(net, train_dataset, val_dataset, data_shape, batch_size, num_workers):

def validate(net, val_data, ctx, eval_metric):

if __name__ == "__main__":

Thank you for trying it out. I will try to run the code snippet and get back to you later.