mAP NAN while training with custom dataset

[YCAyca]

Hello, I try to train YOLOX-X using my custom dataset in COCO format. While it was fine with a small version of my dataset (~3K images) using the default settings in yolox_base.py (only batch size set to 4, due to the lack of GPU memory), when I train it using the big version of my dataet (~10K) I cant find any solution to prevent getting NAN for each class, since the first epoch. I have tried many things:

• Decreasing self.basic_lr_per_img by 10
• Decreasing self.basic_lr_per_img by 100
• Decreasing epoch number to 100, to 75 (thinking that the yolox cosine warmup learning rate scheduler gets arranged according to the total iteration number, and since my iteration per epoch is 3x more now, decreasing max epoch could...)
• Using multi GPU so that I can put batch size = 16, which gives me very similar iteration number per epoch with my previous training

NONE of them worked. I don't know what to do else, is anyone have any idea?
Apart from that, I have checked my COCO format labels in the platform that I used to convert my labels, and they all seem fine. But maybe in YOLOX dataloader something is wrong, how can I visualize my ground truths easily after loading a batch in YOLOX training???

[YCAyca]

Another weird thing is that all the losses are 0 since the first iteration, except conf_loss which becomes directly the total loss: