Thank you for contributing to PaddlePaddle.
Before submitting the issue, you could search issue in the github in case that there was a similar issue submitted or resolved before.
If there is no solution,please make sure that this is a training issue including the following details:
System information
-PaddlePaddle version: v2.2
-CPU: None
-GPU: V100 CUDA11.4
-OS Platform: Ubuntu 20.04
-Other imformation: None
Note: You can get most of the information by running summary_env.py .

To Reproduce
Run RN50-static-train with AMP-O1 configuration.
Print "eval_prog" and check the log.

Describe your current behavior
When testing RN50 AMP training, I found that eval program use FP32 instead of FP16, and the throughput of evaluation is much lower than training.
This may be due to some operations being pruned during eval_prog.clone(for_test=True) .
As a workaround, I can manually call rewrite_program to cast ops to FP16 in eval prog.
logically, train prog and eval prog are both supposed to use mixed precision in AMP training. And this issue would not only appear in RN50, but also in all AMP training with eval(test) program.
So I think this may be a design bug.

Code to reproduce the issue
Clone PaddleClas and prepare relevant training environment.

git clone https://github.com/PaddlePaddle/PaddleClas.git

Insert code to train.py#L167 to print program:

print(train_prog)
print(eval_prog)

Then run training with AMP-O1 config:

bash ./ppcls/static/run_dali.sh

Then check the log. In train prog, some variables are casted to "float16". But in eval prog, all variables are "float32". And the throughput of evaluation is much lower than training, which is unreasonable.

Manually call rewrite_program here train.py#L167 to cast ops to FP16 in eval prog.

from paddle.fluid.contrib.mixed_precision.fp16_utils import rewrite_program
from paddle.fluid.contrib.mixed_precision.fp16_lists import AutoMixedPrecisionLists
rewrite_program(eval_prog, amp_lists=AutoMixedPrecisionLists())

Then print the eval prog and check the log. Some variables are casted to FP16 and the throughput of evaluation becomes higher than training. This is more like the result I expected.

Other info / logs
In eval prog, all variables are "float32".:

After rewrite eval prog, some variables are casted to "float16" according to white/black list.