masahi opened a new issue #4739: FoldScaleAxis + FoldConstant suboptimal on Conv + BN + Relu? URL: https://github.com/apache/incubator-tvm/issues/4739 Hi, when I run the following script I get an IR at the bottom which looks suboptimal to me. I expected to be left with only conv2d + bias add + relu. Am I missing something? @tqchen @vinx13 ``` def test_fold_bn(): def get_layers(prefix, data): weight = relay.var(prefix+"weight") bn_gamma = relay.var(prefix+"bn_gamma") bn_beta = relay.var(prefix+"bn_beta") bn_mmean = relay.var(prefix+"bn_mean") bn_mvar = relay.var(prefix+"bn_var") layer = relay.nn.conv2d(data=data, weight=weight, kernel_size=(3,3), channels=16, padding=(1, 1)) layer = relay.nn.batch_norm(layer, bn_gamma, bn_beta, bn_mmean, bn_mvar)[0] layer = relay.nn.relu(layer) return layer data = relay.var("data", relay.TensorType((1, 3, 224, 224), "float32")) layer1 = get_layers("layer1_", data) last = layer1 net = relay.Function(relay.analysis.free_vars(last), last) ishape = (1, 3, 224, 224) mod, params = tvm.relay.testing.create_workload(net) with relay.build_config(opt_level=3, disabled_pass=["AlterOpLayout"]): opt_mod, params = relay.build_module.optimize(mod, "llvm") print(opt_mod["main"].astext()) ``` ``` fn (%data: Tensor[(1, 3, 224, 224), float32], %layer1_weight: Tensor[(16, 3, 3, 3), float32], %layer1_bn_gamma: Tensor[(16), float32], %layer1_bn_beta: Tensor[(16), float32], %layer1_bn_mean: Tensor[(16), float32], %layer1_bn_var: Tensor[(16), float32]) -> Tensor[(1, 16, 224, 224), float32] { %3 = fn (%p0: Tensor[(16), float32], %p1: Tensor[(16), float32], Primitive=1) -> Tensor[(16), float32] { %0 = add(%p0, 1e-05f /* ty=float32 */) /* ty=Tensor[(16), float32] */; %1 = sqrt(%0) /* ty=Tensor[(16), float32] */; %2 = divide(1f /* ty=float32 */, %1) /* ty=Tensor[(16), float32] */; multiply(%2, %p1) /* ty=Tensor[(16), float32] */ }; %4 = %3(%layer1_bn_var, %layer1_bn_gamma) /* ty=Tensor[(16), float32] */; %8 = fn (%p01: Tensor[(16, 3, 3, 3), float32], %p11: Tensor[(16), float32], Primitive=1) -> Tensor[(16, 3, 3, 3), float32] { %5 = expand_dims(%p11, axis=1, num_newaxis=2) /* ty=Tensor[(16, 1, 1), float32] */; %6 = squeeze(%5, axis=[1, 2]) /* ty=Tensor[(16), float32] */; %7 = expand_dims(%6, axis=1, num_newaxis=3) /* ty=Tensor[(16, 1, 1, 1), float32] */; multiply(%p01, %7) /* ty=Tensor[(16, 3, 3, 3), float32] */ }; %9 = %8(%layer1_weight, %4) /* ty=Tensor[(16, 3, 3, 3), float32] */; %16 = fn (%p02: Tensor[(1, 3, 224, 224), float32], %p12: Tensor[(16, 3, 3, 3), float32], %p2: Tensor[(16), float32], %p3: Tensor[(16), float32], %p4: Tensor[(16), float32], Primitive=1) -> Tensor[(1, 16, 224, 224), float32] { %10 = nn.conv2d(%p02, %p12, padding=[1, 1], channels=16, kernel_size=[3, 3]) /* ty=Tensor[(1, 16, 224, 224), float32] */; %11 = negative(%p2) /* ty=Tensor[(16), float32] */; %12 = multiply(%11, %p3) /* ty=Tensor[(16), float32] */; %13 = add(%12, %p4) /* ty=Tensor[(16), float32] */; %14 = expand_dims(%13, axis=1, num_newaxis=2) /* ty=Tensor[(16, 1, 1), float32] */; %15 = add(%10, %14) /* ty=Tensor[(1, 16, 224, 224), float32] */; nn.relu(%15) /* ty=Tensor[(1, 16, 224, 224), float32] */ }; %16(%data, %9, %layer1_bn_mean, %4, %layer1_bn_beta) /* ty=Tensor[(1, 16, 224, 224), float32] */ } ```
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
