Joshua Z. Zhang

you have to use ImageNet pre-trained weights, otherwise you need a DSSD variant.

@titikid For maximum flexibility I suggest you to use broadcast multiply instead of batchnorm itself. You have full control of how the behavior is without hacking batchnorm itself.

You need to cast not only data but also label to float16 in symbol. You can follow the instruction on mxnet/example/image-classifcation/symbols/resnet.py Basically it's a mx.sym.cast('float16') for all inputs.

balancing class examples is important, I guess your number characters are very few compared to others.

caffe version is forcing every thing resized to 300x300 in the preprocessing step, so that's basically not different.

For (1), update submodule, it's indicating you don't have the correct python module (2) use num_parts and part_index instead of rand and nworker in cfg.train/valid to split dataset. The second...

starting from 0

very likely the mxnet is not correctly installed

python 3 don't support print syntax, you need to modify the code a bit or use python2.7 instead

I have a plan to add this. If you are in hurry, you can implement one in the meantime. Just a couple of lines of code IMO.