DLR inference issue on raspberry pi about neo-ai-dlr

[haojin2]

Hello,
I just used your code to export the resnet50_v1 model from GluonCV and managed to get a successful compilation job on my side. Note that you're using layout='HWC' when you're exporting the model, so you should specify the input configuration to be {'data': (1, 224, 224, 3)}, and you should do the same for the TF model if it's in NHWC format.
Please let me know if you ever encounter any more problems with compiling models. Thanks!
Hao

[wweic]

@fast4break We looked at the model you use and think it's because its format might not be supported by Neo(the models were last updated on 2016-10-21, which means it was generated by a MXNet version at that time). Is it possible for you to try latest Gluon model?

[fast4break]

@wweic sure, I will try it asap

[fast4break]

@wweic Could not find pretrained Gluon models, so I use the following codes to generate .params and .json files:
net = gcv.model_zoo.get_model('resnet50_v1', pretrained=True) export_block('resnet50_v1', net, preprocess=True, layout='HWC')
Compress two files into a .tar.gz file and upload to s3, use Sagemaker console to compile the model, failed with following error:

Client Error: InputConfiguration: Check 1. Input names and shapes matches with trained model. 2. If you are compiling MXNet model, make sure .params file only contains model weights.

I also tried a tensorflow model http://download.tensorflow.org/models/official/20181001_resnet/savedmodels/resnet_v1_fp16_savedmodel_NHWC.tar.gz
Compress the .pb file and the variables directory into a .tar.gz file and upload to s3, compile from sagemaker console, failed with the following error:

Unexpected error during compilation

[fast4break]

@haojin2 Thanks, problem solved and models run on pi.

[wweic]

thanks @haojin2 @fast4break .