85426e4964fb6ee30ca5b14c5c79f4843e948962,memcnn/models/tests/test_revop.py,,test_normal_vs_revblock,#,215

Before Change

            e.train()

        // define an arbitrary reversible function and define graph for model 1
        Xin = X.clone()
        fn = ReversibleBlock(c1_2, c2_2, keep_input=False, coupling=coupling, adapter=AffineAdapterNaive,
                                   implementation_fwd=implementation_fwd, implementation_bwd=implementation_bwd)
        Y = fn.forward(Xin)
        loss2 = torch.mean(Y)

        // define the reversible function without custom backprop and define graph for model 2
        XX = X.clone().data

After Change

            e.train()

        // define an arbitrary reversible function and define graph for model 1
        Xin = X.clone().requires_grad_()
        fn = ReversibleBlock(c1_2, c2_2, keep_input=False, coupling=coupling, adapter=AffineAdapterNaive,
                                   implementation_fwd=implementation_fwd, implementation_bwd=implementation_bwd)
        Y = fn.forward(Xin)
        loss2 = torch.mean(Y)

        // define the reversible function without custom backprop and define graph for model 2
        XX = X.clone().detach().requires_grad_()

Instances