Lower mAP on custom dataset, with comparison

[scaramouche88]

Hi @eriklindernoren , I'm training with a custom dataset (1 class, 2900 images for train, 900 for evaluation). With another repo of Yolov3 (this one) I get an mAP of 76%. With the same config (as it should be), with this repo I arrive to 55%. Adapting the config (longer training), I reach 63%.

However if I test, as suggested in the Readme, with test.py I have a difference of 3% in mAP on COCO (52.6% vs 55.5% reported).

From where such discrepancy might come from?