Insightface: Running performance tests to find the best convolution algorithm, this can take a while

[meixitu]

Hi ,
Thanks for your wonderful work.
I tried to use your pre-trained Retinaface-R50 model to run inference.

When I run inference for megaface dataset, it always has below infomation:
Running performance tests to find the best convolution algorithm, this can take a while... (setting env variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)

This test slow down the inference so much, I set the MXNET_CUDNN_AUTOTUNE_DEFAULT=0 in environment, but it is not disabled.

I use mxnet-cu10.0 1.3.0 version, Is anyone can help me?

[bityigoss]

in the R50-symbol.json file, reset param cudnn_tune's value with "None" would help. the default value of cudnn_tune is limited_workspace.
mxnet reference

[zzh928253398]

i reset param cudnn_tune's value with "off", it works.

cudnn_tune: enable this option leads to higher startup time but may give faster speed. Options are
off: no tuning
limited_workspace:run test and pick the fastest algorithm that doesn’t exceed workspace limit.
fastest: pick the fastest algorithm and ignore workspace limit.
None (default): the behavior is determined by environment variable MXNET_CUDNN_AUTOTUNE_DEFAULT.
0 for off,
1 for limited workspace (default),
2 for fastest.

[pasa13142]

Does accuracy change with these values and there almost 50 cudnn tune value, we should change all of them ?

[Programmerwyl]

I set

os.environ["MXNET_CUDNN_AUTOTUNE_DEFAULT"] = "2"

but it does not work,how to set it
thanks

[KEVIN666666666]

I set

os.environ["MXNET_CUDNN_AUTOTUNE_DEFAULT"] = "2"

but it does not work,how to set it
thanks

os.environ['MXNET_CUDNN_AUTOTUNE_DEFAULT'] = '0'

This works for me.

[j-t-fraser]

The following steps combined removed all autotunes for me. Neither of the two steps worked alone:

• Replacing the cudnn_tune value "limited_workspace" with "off" in the R50-symbol.json file
• os.environ['MXNET_CUDNN_AUTOTUNE_DEFAULT'] = '0'