the converted model get more larger about condensenetv2 HOT 6 CLOSED

Did you use this function to convert your model?

from condensenetv2.

yes,i did,
and i tried to converted the model by convert_and_eval.py on cifar10, the result is same.
and then i tried to replace the function of CondensingSFR in layers.py to CondensingConv in projects Condensenet，the models got smaller，but it got wrongs when test.
what i think the question of the converted models get larger is in the function of CondensingSFR, would you please have a see.
thanks

from condensenetv2.

I test the model configuration which you provide.

The original model's FLOPs/Parameters are 203.17M/1.49M. After converting, the model's FLOPs/Param will be 40.60M/1.22M which are less than original model.

Could you give model details about your experiments? For example, what is the meaning of the number 10723->24418?

from condensenetv2.

oh，the 10723KB means the param of models that convert to onnx，and the 24418KB means the param of converted models that converted to onnx.
could you please give me your email, then i send my codes to you

from condensenetv2.

To be honest, we do not test what will happen if we convert the pytorch model to onnx model. Our experiments are all conducted on GPU device. You can send your code to my email([email protected]), I will check your code later(project ddl is coming, sorry) see what I can do to help you.

from condensenetv2.

OK, thanks
i will send my projects to you, and the converted model on GPU device is also get larger, so i will hope you help me solve the problem,

from condensenetv2.