How to speed up training for large batchsize

[aihao2000]

Currently my batchsize is 100 and each sample contains three images. Do I need to change write_batch_size in datasets.map? How should I estimate this value? I specified DEFAULT_WRITE_BATCH_SIZE=100 in the constructor of the bulider class in the my_dataset.py script. Will it be the default write_batch_size=1000 with datasets.map?

[aihao2000]

After my experiment, I found that with_transform is much faster than map when I have a powerful cpu