[Model Zoo] Potential additions to available catalogue

[afzalxo]

Hi,

We have been working on FPGA-aware neural architecture search with the objective of maximizing accuracy-throughput tradeoffs of the searched models on Image classification tasks with ImageNet2012. We have come up with a few efficientnet-like models that, in some cases, outperform most high-quality models available in the model zoo in both accuracy and throughput. Here is a list of models that we have found:	Target Device	DPU Arch	Model Name	Top-1 Accuracy Float	Top-1 Accuracy Quantized	E2E throughput (fps), Multi Thread, measured on target
VCK190	DPUCVDX8G_ISA3_C32B6	efficientnet-vck190-a	77.568	76.966	2805.14
VCK190	DPUCVDX8G_ISA3_C32B6	efficientnet-vck190-b	76.656	76.300	4015.04
ZCU102	DPUCZDX8G_ISA1_B4096	efficientnet-zcu102-a	77.698	77.432	271.94
ZCU102	DPUCZDX8G_ISA1_B4096	efficientnet-zcu102-b	76.602	76.314	398.54

The following plots show the comparison in accuracy/throughput against some of the high-quality models currently available in the model zoo. We are comparing against models that take 224 x 224 inputs, and are not pruned, on the ImageNet2012. The numbers for existing models in the plots are obtained from the spreadsheet provided here, while the numbers for our models are measured on the device.

VCK190 Models:

ZCU102 Models:

Would you be interested in adding these models to the model zoo if we can provide the relevant pre-trained models, code, quantized weights, etc?

[wangxd-xlnx]

Hi @afzalxo

Thanks for your comment. Overall I think it's a good idea. We will discuss with marketing in what form it would be better to add your model into model zoo.

Considering the user experience, we hope to be unified in model packaging. It should include float/quantized model, code, and readme.

Could you send us a sample at your convenience? Thanks again for your support.