I tested the parameters and FLOPs using this code:
https://github.com/fdbtrs/AdaDistill/blob/main/backbones/mobilefacenet.py
The computation code is as follows:
from ptflops import get_model_complexity_info backbone = MobileFaceNet([112,112]) data = torch.zeros([1,3,112,112]) backbone.eval() T1 = time.time() feature = backbone(data) T2 = time.time() macs, params = get_model_complexity_info(backbone, (3, 112, 112), as_strings=True, print_per_layer_stat=False, verbose=False) log = '' log+= '{} {}'.format('Params: ', params) log+= '\t{} {}'.format('FLOPs: ', macs) log+= '\toutput: {}'.format((np.squeeze(feature.detach().numpy())).shape) log+= '\tcost : %0.2f ms'%((T2 - T1) * 1000) print(log)
The result is :
Params: 1.2 M FLOPs: 233.09 MMac output: (512,) cost : 43.12 ms
However, the paper states in Section 4.3:
"We use MobileFaceNet [9] (1.19M parameters and 0.45 GFLOPs), noted as MFN, as the student."
I would appreciate it if the author could explain this discrepancy.
I tested the parameters and FLOPs using this code:
https://github.com/fdbtrs/AdaDistill/blob/main/backbones/mobilefacenet.py
The computation code is as follows:
from ptflops import get_model_complexity_info backbone = MobileFaceNet([112,112]) data = torch.zeros([1,3,112,112]) backbone.eval() T1 = time.time() feature = backbone(data) T2 = time.time() macs, params = get_model_complexity_info(backbone, (3, 112, 112), as_strings=True, print_per_layer_stat=False, verbose=False) log = '' log+= '{} {}'.format('Params: ', params) log+= '\t{} {}'.format('FLOPs: ', macs) log+= '\toutput: {}'.format((np.squeeze(feature.detach().numpy())).shape) log+= '\tcost : %0.2f ms'%((T2 - T1) * 1000) print(log)The result is :
Params: 1.2 M FLOPs: 233.09 MMac output: (512,) cost : 43.12 msHowever, the paper states in Section 4.3:
"We use MobileFaceNet [9] (1.19M parameters and 0.45 GFLOPs), noted as MFN, as the student."
I would appreciate it if the author could explain this discrepancy.