Number of images increasing in each epoch of YoloV4 training

Question

Couldn't put images due to low reputation

epoch 45: 950.664856 avg loss, 2880 images, 19.969 hours left

epoch 46: 953.999146 avg loss, 2944 images, 19.879 hours left

The number of images is increasing from one epoch to the next. I don't have an in-depth understanding of how it works. I followed the "how to train to detect custom objects" section of the alexeyAB repo. The number of images I have given as the dataset is 50, but after each layer, the number of images goes up by 64. Why could this be happening? Somehow the estimated training time is 16 hours, even though the loss has already gone down to 100 from the initial 4500 in half an hour.

First I stopped the training and checked my cfg file, but I had the correct changes according to the repo. I know that the model already has a data augmentation step, so I did not try to apply any data augmentation beforehand. The images come from a CFD simulation, so getting more is a difficult situation. I went to the issues section on the repo, but didn't find one that matched what I was seeing. On google searching "Number of images increasing in each epoch of YoloV4 training" gave me results only about data augmentation.

By the time I wrote this question (after 50 mins of training), the avgloss has gone down to 8.4. What would be an appropriate loss to stop at considering my dataset size? And why is the image count increasing each epoch?