When I train a neural network, the validation accuracy was 0 at first 3000 steps, after that it went up, why did this happen?

Question

During the training of CNN, I encountered some problems. As shown in the following figure, In the early stage of neural network training, the validation loss drops very slowly, and the validation accuracy is almost zero, but after some steps it goes up. Why is validation curve not in step with the training curve in the early stage? What could be the reason? Hope someone can answer, thanks!