create monthly mobile optimized CV models

[alexshepard]

The mobile team would like mobile CV models to be available in lockstep with the web CV model as soon as convenient. For the mobile MVP, this does not need to be an automatic pipeline for us in the way we integrate the model into the apps. It also does not need to be hosted somewhere so that we can download it at app runtime. It would be included manually at compile time.

Johannes says:

Seek should be ready to consume the latest model already, given that the model API does not change. But there is no priority organization-wide to always have the latest model in Seek. iNatRN has a current estimate of an MVP at the end of June with a current track record of moving back two weeks every two weeks. For the iNatRN MVP, we want to be able to manually include the latest model on every update to the web model. So, the earliest potential time for high priority is the end of June. Currently still a low priority.

[alexshepard]

In order to make the CV model small enough to deploy on mobile, we can apply a few strategies right now.

• fine tune a CV model with a factorized head to reduce the size of the final linear layer, and
• after training, quantize weights during CoreML and TFLite conversion.

Future strategies that might be worth exploring could include:

• exploring mobile optimized model architectures like one of the smaller efficientnet models,
• dual headed training to produce both web and mobile CV models during the same fine tuning run,
• quantization during training,
• etc

But for now let's just get something simple going so we can produce mobile models for the mobile beta team every month. I'll start the first pass of this for 2.13 on bluey since nobody seems to be using it right now.