Anton Lokhmotov

Results 273 comments of Anton Lokhmotov

Thank you @MatthewARM. For the official [MLPerf Inference v2.0](https://mlcommons.org/en/inference-edge-20/) submission, we measured 339 ms for Odroid N2+ and 349 ms for Raspberry Pi 4. Seems like a RPi4 regression to...

@ckstanton Does this look like a manifestation of the same bug to you? ``` ================================================ MLPerf Results Summary ================================================ SUT name : ArmNN_TFLite_SUT Scenario : SingleStream Mode : PerformanceOnly 90th...

A question to the WG. We have already collected **lots** of results like above, which would be rather onerous to rerun before the submission deadline. I'm wondering if we can...

@nv-jinhosuh That's my feeling too. But IIRC we decided that the early estimate would be used as the metric regardless. And it is what the submission checker appears to do....

Taken from [the latest meeting slides](https://docs.google.com/presentation/d/1x7goS8XnuwlS2PVlsoE93KvbQQA-GhDjEEYGORQ5pfk/edit#slide=id.g12733238521_0_1) ``` Accuracy target FP32: 0.375 mAP Latency target WG Approves:100 msec ``` [Copy of 2022.08.02 MLCommons Inference WG Meeting - RetinaNet.pdf](https://github.com/mlcommons/inference/files/9259454/Copy.of.2022.08.02.MLCommons.Inference.WG.Meeting.-.RetinaNet.pdf)

Sorry, I missed [PR1168](https://github.com/mlcommons/inference/pull/1168) (submission checker change) 5 weeks ago. But [PR](https://github.com/mlcommons/inference_policies/pull/251) was only two weeks ago. According to [the attached slide](https://github.com/mlcommons/inference/files/9261585/Copy.of.2022.07.05.MLCommons.Inference.WG.Meeting.-.v2.1.timeline.pdf) from [the WG meeting slide deck](https://docs.google.com/presentation/d/1NYiCaWGFUc45gKZvAbH-Jooqw37lR_d71W9RDbKqc-E/edit#slide=id.gbb938d29b3_3_0) 4 weeks...

Where do we draw a line between cases like this and (non-WG approved) late changes like https://github.com/mlcommons/inference/pull/1206?

Thank you Arjun. Your accuracy is still higher than 37.55%.

> [Inference Policies](https://github.com/mlcommons/inference_policies/blob/master/inference_rules.adoc#minimum-requirements-of-a-datacenter-system) say the accuracy requirement as 99% of FP32 model, but does not clearly say which exact model (lowest accuracy one?) nor say up to how many digits...