您好，想问一下FLOPs指标的问题

[zushiye1]

请问，这些模型中的FLOPs指标是如何得到的呢

[Dachuan Shi]

By fvcore package. See here

UPop/deit/main.py

Lines 343 to 350 in 6aae798

 #### Count parameters and FLOPs #### 
 search_model.eval() 
 with torch.no_grad(): 
     flop = FlopCountAnalysis(search_model, torch.randn(1, 3, 224, 224).to(device)) 
     print(flop_count_table(flop, max_depth=7, show_param_shapes=True)) 
     print("Total", flop.total() / 1e9) 
 search_model.train()  
 #######################################