the training part of wider face dataset is about 15K
and the batch size is 1, using 4 GPUs,
the iteration is setting to 21K to get the best result.
how did the 21K come from?