Joshua Z. Zhang

because you are assigning your customized weights to voc models, the shape of class prediction is different. You can load the same net in your training script. I have added...

`Isn't larger image supposed to give better result?` No, receptive field is fixed, therefore for image classification, the performance may drop if the input is too big. Note that the...

in c++, it does not matter what's the input resolution is, just make sure it's in good range (300-600) which the model has been trained with

Did you check out these? https://github.com/dmlc/gluon-cv/tree/master/gluoncv/data/transforms/presets

@zhangpzh change the number of gpus will affect the effective batch size on each GPU if syncBN is used, which is not present for SSD training example right now. I...

@zhangpzh Can you follow exactly the same bash command provided to see whether you can reproduce?

@zhangpzh Thanks for the report. I will investigate once I got time on it.

For now, the memory is always released to a memory pool for reuse. There's no API to release the memory pool to system, unless you exit the process.

I can help you raise it if you can open an feature request first https://github.com/apache/incubator-mxnet/issues

TLDR. It's moved to MultiBoxLoss. Reason: OHEM is dependent on prediction value, but we want to pre-compute all targets using cpu cores during data loading. Instead we can use cpu...