What if accuracy is high but loss is not that good for text detection training?

[light42]

I'm training text detection with ResNet_v50 model, and dataset synthetically generated from this repo.
I am using default configs, with slight adjusment at learning_rate(0.0006) and weight_decay(0.0001).

It achieved very high accuracy (~0.99) yet the loss has been so-so (dropped from ~6.00 to ~1.00).
My question is, is this already good enough? And if it is not good, what can I do to improved it further?

[LDOUBLEV]

Loss is not an indicator for evaluating the quality of the model, acc is.

The calculation formula of loss is different, and the output value will be stable within a certain range, not necessarily a small value.

BTW, the loss of GAN will even have negative values.