The training time of per iteration changes when the batch_size is changed

[chenchen278]

The training time of per iteration changes when the batch_size is changed.For example, a cost time about 75s and 270s as the batch_size of 10 and 32. And in my opinion, this time should keep the same as the images are passed in parallel to the model. Also, I check my gpu's status and find its utilization is quit low. Was it cause by the dataset_generated_program in this project?