mxnet-ssd: Metric error on the last version of ssd

On the last version of ssd (I don’t know since when) I got this error when I try to train with all default parameters on pascal voc :

[18:45:00] src/operator/convolution.cu:87: This convolution is not supported by cudnn, MXNET convolution is applied.
Traceback (most recent call last):
  File "train.py", line 134, in <module>
    voc07_metric=args.use_voc07_metric)
  File "/home/rd/Documents/RD3/users/Hinde/mxnet/example/ssd/train/train_net.py", line 266, in train_net
    monitor=monitor)
  File "/home/rd/Documents/RD3/lib/mxnet/python/mxnet/module/base_module.py", line 481, in fit
    self.update_metric(eval_metric, data_batch.label)
  File "/home/rd/Documents/RD3/lib/mxnet/python/mxnet/module/module.py", line 664, in update_metric
    self._exec_group.update_metric(eval_metric, labels)
  File "/home/rd/Documents/RD3/lib/mxnet/python/mxnet/module/executor_group.py", line 558, in update_metric
    eval_metric.update_dict(labels_, preds)
  File "/home/rd/Documents/RD3/lib/mxnet/python/mxnet/metric.py", line 82, in update_dict
    pred = [pred[name] for name in self.output_names]
TypeError: 'int' object is not iterable

The problem seem to come from prediction that return an index instead a list of index. I use mxnet 0.9.5 and cudnn 6. edit : I use ssd code from official mxnet repository (mxnet/example/ssd)