[SOLVED] Some warning message for running on CUDA-9.0

Questions
1

Hi,

I configured the TVM framework with CUDA-9.0, and run the “from_darknet.py” with the correct result, but has some warning messages during the running process, I don’t know why this happened?

Compiling the model…
WARNING:autotvm:Cannot find config for target=cuda, workload=(‘conv2d’, (1, 3, 416, 416, ‘float32’), (32, 3, 3, 3, ‘float32’), (1, 1), (1, 1), ‘NCHW’, ‘float32’). A fallback configuration is used, which may bring great performance regression.
WARNING:autotvm:Cannot find config for target=cuda, workload=(‘conv2d’, (1, 32, 416, 416, ‘float32’), (64, 32, 3, 3, ‘float32’), (2, 2), (1, 1), ‘NCHW’, ‘float32’). A fallback configuration is used, which may bring great performance regression.

Loading the test image…
Running the test image…

Please give me some advices.
Thanks so much.

2

To get the best performance, we need to do auto-tuning for every conv2d operators in the network.
We released pre-tuned parameters for some common convolution networks (resnet-50, vgg-16, mobilenet, etc) and tvm will download them during compilation.

For untuned shapes like the shapes reported by you, we will use some fallback configurations.
The performance is acceptable but is not the best. If you only want to try some examples, you can ignore these warning.
If you care about the performance, you can follow the tutorial to tune by yourself.

1 Like
3

Thanks for your detailed explanation. I understand the waring message now. Thanks so much.