I have lost in image classification on Imagenet dataset

I have downdload Imagenet validation dataset.It contain 50000 data.I have a plan that train the validation dataset, not training dataset

I have make the one-hot label,and I used binary_cross_entropy, KLD and MSE as my loss function. Backbone is resent34. I found it can not be convergence.

Then I maked the index label,for example, label :{10,15,480,1…} .This time I used cross_entropy as my loss function.I found it still can not be convergence.Even though I only train 700 data.

So ,I do not know what is the happend