Mingxiao Huang

worse chainer convnet-benchmarks performance on cupy-2.0.0 as compared to cupy-1.0.0.1

2

Hello, would you please help explain this issue? Thanks in advance. We found that convnet-benchmarks performance on cupy-2.0.0 is worse than that on cupy-1.0.0.1. We don't know whether it is...

ResNet50 evaluation on Imagenet only got 74% top1 accuracy.

1

Hello, we tried to reproduce the ResNet50 inference, but only got 73+% top1, it is still far from SOTA(75%) and your data(76%). Any comments here? Details: torch (1.1.0a0+a3933b8) torchvision (0.2.2.post3)...

How did you train resnet50 model

1

hello, I am trying to train resnet50 based on pytorch, however, I only got 57% top1 validation acuracy on imagenet, which is far from SOTA. I tried to do inference...

hello, would you please also consider CPU(especially iDeep enbaled) case? In some files, we can see that you did "model.to_gpu()" or "chainer.cuda.to_cpu(**)" without checking directly. Thanks.

fix the bug in xpu

4

only when import intel_extension_for_pytorch will torch has xpu attribute. fix the device configure bug for DDP. enable accelerate launch for xpu. fix the bug in wait_for_everyone for xpu exposed in...

risk brought by no pre-allocate output for allgather&reducescatter&alltoall

8

Hello, recently I was trying to enable horovod torch UT on xpu backend, and then I noticed one issue, torch allgather&reducescatter UT will fail randomly when hvd.size() > 4, if...

hello, may I know where to get the models mentioned by ./models/resnet_v1_50/run_inference.sh \ --use_xla --no_tf32 \ --data_dir=/data/imagenet/train-val-tfrecord --input_saved_model_dir=/models in https://github.com/tensorflow/tensorrt/tree/master/tftrt/benchmarking-python/image_classification