Working OCR

[dtronmans]

Purpose

Depends on luxonis/luxonis-train#340

• 800 train images with 62 characters is a frequency of 100 only per character, which was not enough to learn. Moved it to 2000 images which is an average frequency of 250 per character.
• Gradient clip, CosineAnnealingLR instead of ConstantLR and AdamW optimizer all make the training more stable, which it seems OCR is very sensitive to.
• accumulate_grad_batches lowered, so a lower effective batch size per optimizer step also happens to make the training more stable and the train loss follows the val loss more closely

Specification

Dependencies & Potential Impact

Deployment Plan

Testing & Validation