Created by: luotao1

when I test the inference benchmark of https://github.com/chengduoZH/benchmark/blob/add_resnet_50_v2/fluid/ResNet_50/train_resnet.py#L182 The correct script is:

test_program = fluid.default_main_program().clone(for_test=True)
optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
optimizer.minimize(avg_cost)

But if I write as

optimizer = fluid.optimizer.Momentum(learning_rate=0.01, momentum=0.9)
optimizer.minimize(avg_cost)
test_program = fluid.default_main_program().clone(for_test=True)

The inference elasped time is the same as the training time.

The reason is that clone(for_test=True) must be used before backward and optimization please. After optimizer.minimize(avg_cost), the backward information will be added in program_desc.