I was trying to reproduce the result of resnet50_v1 in mxnet.model_zoo.vision
and gluoncv.model_zoo
. I use the scripts in
https://github.com/apache/incubator-mxnet/blob/master/example/gluon/image_classification.py
with
python image_classification.py \
--dataset imagenet \
--data-dir /mnt/ssd/imagenet \
--num-worker 16 \
--batch-size 64 \
--lr 0.2 \
--model resnet50_v1
and
https://github.com/dmlc/gluon-cv/blob/master/scripts/classification/imagenet/train_imagenet.py
with
python train_imagenet.py \
--data-dir /mnt/ssd/imagenet \
--model resnet50_v1 --mode hybrid \
--lr 0.2 --num-epochs 120 --batch-size 64 --num-gpus 4 -j 32 \
--save-dir params_resnet50_v1_best \
--logging-file resnet50_v1_best.log
The accuracy is around 74%, which is below the 76.15% reported by torchvision. I don’t use label smoothing and fp16. So what’s wrong with my settings? Is there anyone who reproduces the accuracy higher than 76%?
My mxnet version is 1.5.1
, mxnet-cu101mkl
from pip install
.