program(1.3)
[buildInfo = dict<string, string>({{"coremlc-component-MIL", "3500.14.1"}, {"coremlc-version", "3500.32.1"},{"model-version", "tinny_mbv2_p9fku39q44_440000_2.0.0"}, {"model-bundle-name", "LearnedFeatures"}, {"model-name", "DenseFeat2"}}), mldb_token = string("mldb-rxg5j37ao8")]
{
    func main<ios18>(tensor<uint8, [1, 1, 320, 320]> image_input) {
            string chema_fp16 = const()[name = string("chema_fp16"), val = string("fp16")];
            tensor<fp16, [1, 1, 320, 320]> image_input_fp16 = cast(dtype = chema_fp16, x = image_input)[name = string("chema_cast_to_fp16")];
            fp16 image__scaled___y_0 = const()[name = string("image__scaled___y_0"), val = fp16(0.00392150879)];
            tensor<fp16, [1, 1, 320, 320]> image = mul(x = image_input_fp16, y = image__scaled___y_0)[name = string("image__scaled__")];
            string input_1_pad_type_0 = const()[name = string("input_1_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_1_pad_0 = const()[name = string("input_1_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> input_1_strides_0 = const()[name = string("input_1_strides_0"), val = tensor<int32, [2]>([2, 2])];
            tensor<int32, [2]> input_1_dilations_0 = const()[name = string("input_1_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_1_groups_0 = const()[name = string("input_1_groups_0"), val = int32(1)];
            string image_to_fp16_dtype_0 = const()[name = string("image_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [32, 1, 3, 3]> const_28_to_fp16 = const()[name = string("const_28_to_fp16"), val = tensor<fp16, [32, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(64)))];
            tensor<fp16, [32]> const_29_to_fp16 = const()[name = string("const_29_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(704)))];
            tensor<fp16, [1, 1, 320, 320]> image_to_fp16 = cast(dtype = image_to_fp16_dtype_0, x = image)[name = string("cast_49")];
            tensor<fp16, [1, 32, 160, 160]> input_3_cast_fp16 = conv(bias = const_29_to_fp16, dilations = input_1_dilations_0, groups = input_1_groups_0, pad = input_1_pad_0, pad_type = input_1_pad_type_0, strides = input_1_strides_0, weight = const_28_to_fp16, x = image_to_fp16)[name = string("input_3_cast_fp16")];
            fp16 input_5_alpha_0_to_fp16 = const()[name = string("input_5_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_5_beta_0_to_fp16 = const()[name = string("input_5_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 32, 160, 160]> input_5_cast_fp16 = clip(alpha = input_5_alpha_0_to_fp16, beta = input_5_beta_0_to_fp16, x = input_3_cast_fp16)[name = string("input_5_cast_fp16")];
            string input_7_pad_type_0 = const()[name = string("input_7_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_7_pad_0 = const()[name = string("input_7_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            int32 input_7_groups_0 = const()[name = string("input_7_groups_0"), val = int32(32)];
            tensor<int32, [2]> input_7_strides_0 = const()[name = string("input_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_7_dilations_0 = const()[name = string("input_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [32, 1, 3, 3]> const_30_to_fp16 = const()[name = string("const_30_to_fp16"), val = tensor<fp16, [32, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(832)))];
            tensor<fp16, [32]> const_31_to_fp16 = const()[name = string("const_31_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1472)))];
            tensor<fp16, [1, 32, 160, 160]> input_9_cast_fp16 = conv(bias = const_31_to_fp16, dilations = input_7_dilations_0, groups = input_7_groups_0, pad = input_7_pad_0, pad_type = input_7_pad_type_0, strides = input_7_strides_0, weight = const_30_to_fp16, x = input_5_cast_fp16)[name = string("input_9_cast_fp16")];
            fp16 input_11_alpha_0_to_fp16 = const()[name = string("input_11_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_11_beta_0_to_fp16 = const()[name = string("input_11_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 32, 160, 160]> input_11_cast_fp16 = clip(alpha = input_11_alpha_0_to_fp16, beta = input_11_beta_0_to_fp16, x = input_9_cast_fp16)[name = string("input_11_cast_fp16")];
            string input_13_pad_type_0 = const()[name = string("input_13_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_13_strides_0 = const()[name = string("input_13_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_13_pad_0 = const()[name = string("input_13_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_13_dilations_0 = const()[name = string("input_13_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_13_groups_0 = const()[name = string("input_13_groups_0"), val = int32(1)];
            tensor<fp16, [16, 32, 1, 1]> const_32_to_fp16 = const()[name = string("const_32_to_fp16"), val = tensor<fp16, [16, 32, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(1600)))];
            tensor<fp16, [16]> const_33_to_fp16 = const()[name = string("const_33_to_fp16"), val = tensor<fp16, [16]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2688)))];
            tensor<fp16, [1, 16, 160, 160]> input_15_cast_fp16 = conv(bias = const_33_to_fp16, dilations = input_13_dilations_0, groups = input_13_groups_0, pad = input_13_pad_0, pad_type = input_13_pad_type_0, strides = input_13_strides_0, weight = const_32_to_fp16, x = input_11_cast_fp16)[name = string("input_15_cast_fp16")];
            string input_17_pad_type_0 = const()[name = string("input_17_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_17_strides_0 = const()[name = string("input_17_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_17_pad_0 = const()[name = string("input_17_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_17_dilations_0 = const()[name = string("input_17_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_17_groups_0 = const()[name = string("input_17_groups_0"), val = int32(1)];
            tensor<fp16, [96, 16, 1, 1]> const_34_to_fp16 = const()[name = string("const_34_to_fp16"), val = tensor<fp16, [96, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(2816)))];
            tensor<fp16, [96]> const_35_to_fp16 = const()[name = string("const_35_to_fp16"), val = tensor<fp16, [96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(5952)))];
            tensor<fp16, [1, 96, 160, 160]> input_19_cast_fp16 = conv(bias = const_35_to_fp16, dilations = input_17_dilations_0, groups = input_17_groups_0, pad = input_17_pad_0, pad_type = input_17_pad_type_0, strides = input_17_strides_0, weight = const_34_to_fp16, x = input_15_cast_fp16)[name = string("input_19_cast_fp16")];
            fp16 input_21_alpha_0_to_fp16 = const()[name = string("input_21_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_21_beta_0_to_fp16 = const()[name = string("input_21_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 96, 160, 160]> input_21_cast_fp16 = clip(alpha = input_21_alpha_0_to_fp16, beta = input_21_beta_0_to_fp16, x = input_19_cast_fp16)[name = string("input_21_cast_fp16")];
            string input_23_pad_type_0 = const()[name = string("input_23_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_23_pad_0 = const()[name = string("input_23_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> input_23_strides_0 = const()[name = string("input_23_strides_0"), val = tensor<int32, [2]>([2, 2])];
            int32 input_23_groups_0 = const()[name = string("input_23_groups_0"), val = int32(96)];
            tensor<int32, [2]> input_23_dilations_0 = const()[name = string("input_23_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [96, 1, 3, 3]> const_36_to_fp16 = const()[name = string("const_36_to_fp16"), val = tensor<fp16, [96, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(6208)))];
            tensor<fp16, [96]> const_37_to_fp16 = const()[name = string("const_37_to_fp16"), val = tensor<fp16, [96]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8000)))];
            tensor<fp16, [1, 96, 80, 80]> input_25_cast_fp16 = conv(bias = const_37_to_fp16, dilations = input_23_dilations_0, groups = input_23_groups_0, pad = input_23_pad_0, pad_type = input_23_pad_type_0, strides = input_23_strides_0, weight = const_36_to_fp16, x = input_21_cast_fp16)[name = string("input_25_cast_fp16")];
            fp16 input_27_alpha_0_to_fp16 = const()[name = string("input_27_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_27_beta_0_to_fp16 = const()[name = string("input_27_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 96, 80, 80]> input_27_cast_fp16 = clip(alpha = input_27_alpha_0_to_fp16, beta = input_27_beta_0_to_fp16, x = input_25_cast_fp16)[name = string("input_27_cast_fp16")];
            string input_29_pad_type_0 = const()[name = string("input_29_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_29_strides_0 = const()[name = string("input_29_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_29_pad_0 = const()[name = string("input_29_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_29_dilations_0 = const()[name = string("input_29_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_29_groups_0 = const()[name = string("input_29_groups_0"), val = int32(1)];
            tensor<fp16, [24, 96, 1, 1]> const_38_to_fp16 = const()[name = string("const_38_to_fp16"), val = tensor<fp16, [24, 96, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(8256)))];
            tensor<fp16, [24]> const_39_to_fp16 = const()[name = string("const_39_to_fp16"), val = tensor<fp16, [24]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(12928)))];
            tensor<fp16, [1, 24, 80, 80]> input_31_cast_fp16 = conv(bias = const_39_to_fp16, dilations = input_29_dilations_0, groups = input_29_groups_0, pad = input_29_pad_0, pad_type = input_29_pad_type_0, strides = input_29_strides_0, weight = const_38_to_fp16, x = input_27_cast_fp16)[name = string("input_31_cast_fp16")];
            string input_33_pad_type_0 = const()[name = string("input_33_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_33_strides_0 = const()[name = string("input_33_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_33_pad_0 = const()[name = string("input_33_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_33_dilations_0 = const()[name = string("input_33_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_33_groups_0 = const()[name = string("input_33_groups_0"), val = int32(1)];
            tensor<fp16, [144, 24, 1, 1]> const_40_to_fp16 = const()[name = string("const_40_to_fp16"), val = tensor<fp16, [144, 24, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(13056)))];
            tensor<fp16, [144]> const_41_to_fp16 = const()[name = string("const_41_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20032)))];
            tensor<fp16, [1, 144, 80, 80]> input_35_cast_fp16 = conv(bias = const_41_to_fp16, dilations = input_33_dilations_0, groups = input_33_groups_0, pad = input_33_pad_0, pad_type = input_33_pad_type_0, strides = input_33_strides_0, weight = const_40_to_fp16, x = input_31_cast_fp16)[name = string("input_35_cast_fp16")];
            fp16 input_37_alpha_0_to_fp16 = const()[name = string("input_37_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_37_beta_0_to_fp16 = const()[name = string("input_37_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 144, 80, 80]> input_37_cast_fp16 = clip(alpha = input_37_alpha_0_to_fp16, beta = input_37_beta_0_to_fp16, x = input_35_cast_fp16)[name = string("input_37_cast_fp16")];
            string input_39_pad_type_0 = const()[name = string("input_39_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_39_pad_0 = const()[name = string("input_39_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            int32 input_39_groups_0 = const()[name = string("input_39_groups_0"), val = int32(144)];
            tensor<int32, [2]> input_39_strides_0 = const()[name = string("input_39_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_39_dilations_0 = const()[name = string("input_39_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [144, 1, 3, 3]> const_42_to_fp16 = const()[name = string("const_42_to_fp16"), val = tensor<fp16, [144, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(20416)))];
            tensor<fp16, [144]> const_43_to_fp16 = const()[name = string("const_43_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23104)))];
            tensor<fp16, [1, 144, 80, 80]> input_41_cast_fp16 = conv(bias = const_43_to_fp16, dilations = input_39_dilations_0, groups = input_39_groups_0, pad = input_39_pad_0, pad_type = input_39_pad_type_0, strides = input_39_strides_0, weight = const_42_to_fp16, x = input_37_cast_fp16)[name = string("input_41_cast_fp16")];
            fp16 input_43_alpha_0_to_fp16 = const()[name = string("input_43_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_43_beta_0_to_fp16 = const()[name = string("input_43_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 144, 80, 80]> input_43_cast_fp16 = clip(alpha = input_43_alpha_0_to_fp16, beta = input_43_beta_0_to_fp16, x = input_41_cast_fp16)[name = string("input_43_cast_fp16")];
            string input_45_pad_type_0 = const()[name = string("input_45_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_45_strides_0 = const()[name = string("input_45_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_45_pad_0 = const()[name = string("input_45_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_45_dilations_0 = const()[name = string("input_45_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_45_groups_0 = const()[name = string("input_45_groups_0"), val = int32(1)];
            tensor<fp16, [24, 144, 1, 1]> const_44_to_fp16 = const()[name = string("const_44_to_fp16"), val = tensor<fp16, [24, 144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(23488)))];
            tensor<fp16, [24]> const_45_to_fp16 = const()[name = string("const_45_to_fp16"), val = tensor<fp16, [24]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30464)))];
            tensor<fp16, [1, 24, 80, 80]> var_170_cast_fp16 = conv(bias = const_45_to_fp16, dilations = input_45_dilations_0, groups = input_45_groups_0, pad = input_45_pad_0, pad_type = input_45_pad_type_0, strides = input_45_strides_0, weight = const_44_to_fp16, x = input_43_cast_fp16)[name = string("op_170_cast_fp16")];
            tensor<fp16, [1, 24, 80, 80]> input_47_cast_fp16 = add(x = input_31_cast_fp16, y = var_170_cast_fp16)[name = string("input_47_cast_fp16")];
            string input_49_pad_type_0 = const()[name = string("input_49_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_49_strides_0 = const()[name = string("input_49_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_49_pad_0 = const()[name = string("input_49_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_49_dilations_0 = const()[name = string("input_49_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_49_groups_0 = const()[name = string("input_49_groups_0"), val = int32(1)];
            tensor<fp16, [144, 24, 1, 1]> const_46_to_fp16 = const()[name = string("const_46_to_fp16"), val = tensor<fp16, [144, 24, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(30592)))];
            tensor<fp16, [144]> const_47_to_fp16 = const()[name = string("const_47_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37568)))];
            tensor<fp16, [1, 144, 80, 80]> input_51_cast_fp16 = conv(bias = const_47_to_fp16, dilations = input_49_dilations_0, groups = input_49_groups_0, pad = input_49_pad_0, pad_type = input_49_pad_type_0, strides = input_49_strides_0, weight = const_46_to_fp16, x = input_47_cast_fp16)[name = string("input_51_cast_fp16")];
            fp16 input_53_alpha_0_to_fp16 = const()[name = string("input_53_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_53_beta_0_to_fp16 = const()[name = string("input_53_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 144, 80, 80]> input_53_cast_fp16 = clip(alpha = input_53_alpha_0_to_fp16, beta = input_53_beta_0_to_fp16, x = input_51_cast_fp16)[name = string("input_53_cast_fp16")];
            string input_55_pad_type_0 = const()[name = string("input_55_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_55_pad_0 = const()[name = string("input_55_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            tensor<int32, [2]> input_55_strides_0 = const()[name = string("input_55_strides_0"), val = tensor<int32, [2]>([2, 2])];
            int32 input_55_groups_0 = const()[name = string("input_55_groups_0"), val = int32(144)];
            tensor<int32, [2]> input_55_dilations_0 = const()[name = string("input_55_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [144, 1, 3, 3]> const_48_to_fp16 = const()[name = string("const_48_to_fp16"), val = tensor<fp16, [144, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(37952)))];
            tensor<fp16, [144]> const_49_to_fp16 = const()[name = string("const_49_to_fp16"), val = tensor<fp16, [144]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(40640)))];
            tensor<fp16, [1, 144, 40, 40]> input_57_cast_fp16 = conv(bias = const_49_to_fp16, dilations = input_55_dilations_0, groups = input_55_groups_0, pad = input_55_pad_0, pad_type = input_55_pad_type_0, strides = input_55_strides_0, weight = const_48_to_fp16, x = input_53_cast_fp16)[name = string("input_57_cast_fp16")];
            fp16 input_59_alpha_0_to_fp16 = const()[name = string("input_59_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_59_beta_0_to_fp16 = const()[name = string("input_59_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 144, 40, 40]> input_59_cast_fp16 = clip(alpha = input_59_alpha_0_to_fp16, beta = input_59_beta_0_to_fp16, x = input_57_cast_fp16)[name = string("input_59_cast_fp16")];
            string input_61_pad_type_0 = const()[name = string("input_61_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_61_strides_0 = const()[name = string("input_61_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_61_pad_0 = const()[name = string("input_61_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_61_dilations_0 = const()[name = string("input_61_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_61_groups_0 = const()[name = string("input_61_groups_0"), val = int32(1)];
            tensor<fp16, [32, 144, 1, 1]> const_50_to_fp16 = const()[name = string("const_50_to_fp16"), val = tensor<fp16, [32, 144, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(41024)))];
            tensor<fp16, [32]> const_51_to_fp16 = const()[name = string("const_51_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50304)))];
            tensor<fp16, [1, 32, 40, 40]> input_63_cast_fp16 = conv(bias = const_51_to_fp16, dilations = input_61_dilations_0, groups = input_61_groups_0, pad = input_61_pad_0, pad_type = input_61_pad_type_0, strides = input_61_strides_0, weight = const_50_to_fp16, x = input_59_cast_fp16)[name = string("input_63_cast_fp16")];
            string input_65_pad_type_0 = const()[name = string("input_65_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_65_strides_0 = const()[name = string("input_65_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_65_pad_0 = const()[name = string("input_65_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_65_dilations_0 = const()[name = string("input_65_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_65_groups_0 = const()[name = string("input_65_groups_0"), val = int32(1)];
            tensor<fp16, [192, 32, 1, 1]> const_52_to_fp16 = const()[name = string("const_52_to_fp16"), val = tensor<fp16, [192, 32, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(50432)))];
            tensor<fp16, [192]> const_53_to_fp16 = const()[name = string("const_53_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(62784)))];
            tensor<fp16, [1, 192, 40, 40]> input_67_cast_fp16 = conv(bias = const_53_to_fp16, dilations = input_65_dilations_0, groups = input_65_groups_0, pad = input_65_pad_0, pad_type = input_65_pad_type_0, strides = input_65_strides_0, weight = const_52_to_fp16, x = input_63_cast_fp16)[name = string("input_67_cast_fp16")];
            fp16 input_69_alpha_0_to_fp16 = const()[name = string("input_69_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_69_beta_0_to_fp16 = const()[name = string("input_69_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 192, 40, 40]> input_69_cast_fp16 = clip(alpha = input_69_alpha_0_to_fp16, beta = input_69_beta_0_to_fp16, x = input_67_cast_fp16)[name = string("input_69_cast_fp16")];
            string input_71_pad_type_0 = const()[name = string("input_71_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_71_pad_0 = const()[name = string("input_71_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            int32 input_71_groups_0 = const()[name = string("input_71_groups_0"), val = int32(192)];
            tensor<int32, [2]> input_71_strides_0 = const()[name = string("input_71_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_71_dilations_0 = const()[name = string("input_71_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [192, 1, 3, 3]> const_54_to_fp16 = const()[name = string("const_54_to_fp16"), val = tensor<fp16, [192, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(63232)))];
            tensor<fp16, [192]> const_55_to_fp16 = const()[name = string("const_55_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(66752)))];
            tensor<fp16, [1, 192, 40, 40]> input_73_cast_fp16 = conv(bias = const_55_to_fp16, dilations = input_71_dilations_0, groups = input_71_groups_0, pad = input_71_pad_0, pad_type = input_71_pad_type_0, strides = input_71_strides_0, weight = const_54_to_fp16, x = input_69_cast_fp16)[name = string("input_73_cast_fp16")];
            fp16 input_75_alpha_0_to_fp16 = const()[name = string("input_75_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_75_beta_0_to_fp16 = const()[name = string("input_75_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 192, 40, 40]> input_75_cast_fp16 = clip(alpha = input_75_alpha_0_to_fp16, beta = input_75_beta_0_to_fp16, x = input_73_cast_fp16)[name = string("input_75_cast_fp16")];
            string input_77_pad_type_0 = const()[name = string("input_77_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_77_strides_0 = const()[name = string("input_77_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_77_pad_0 = const()[name = string("input_77_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_77_dilations_0 = const()[name = string("input_77_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_77_groups_0 = const()[name = string("input_77_groups_0"), val = int32(1)];
            tensor<fp16, [32, 192, 1, 1]> const_56_to_fp16 = const()[name = string("const_56_to_fp16"), val = tensor<fp16, [32, 192, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(67200)))];
            tensor<fp16, [32]> const_57_to_fp16 = const()[name = string("const_57_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79552)))];
            tensor<fp16, [1, 32, 40, 40]> var_259_cast_fp16 = conv(bias = const_57_to_fp16, dilations = input_77_dilations_0, groups = input_77_groups_0, pad = input_77_pad_0, pad_type = input_77_pad_type_0, strides = input_77_strides_0, weight = const_56_to_fp16, x = input_75_cast_fp16)[name = string("op_259_cast_fp16")];
            tensor<fp16, [1, 32, 40, 40]> input_79_cast_fp16 = add(x = input_63_cast_fp16, y = var_259_cast_fp16)[name = string("input_79_cast_fp16")];
            string input_81_pad_type_0 = const()[name = string("input_81_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_81_strides_0 = const()[name = string("input_81_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_81_pad_0 = const()[name = string("input_81_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_81_dilations_0 = const()[name = string("input_81_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_81_groups_0 = const()[name = string("input_81_groups_0"), val = int32(1)];
            tensor<fp16, [192, 32, 1, 1]> const_58_to_fp16 = const()[name = string("const_58_to_fp16"), val = tensor<fp16, [192, 32, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(79680)))];
            tensor<fp16, [192]> const_59_to_fp16 = const()[name = string("const_59_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92032)))];
            tensor<fp16, [1, 192, 40, 40]> input_83_cast_fp16 = conv(bias = const_59_to_fp16, dilations = input_81_dilations_0, groups = input_81_groups_0, pad = input_81_pad_0, pad_type = input_81_pad_type_0, strides = input_81_strides_0, weight = const_58_to_fp16, x = input_79_cast_fp16)[name = string("input_83_cast_fp16")];
            fp16 input_85_alpha_0_to_fp16 = const()[name = string("input_85_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_85_beta_0_to_fp16 = const()[name = string("input_85_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 192, 40, 40]> input_85_cast_fp16 = clip(alpha = input_85_alpha_0_to_fp16, beta = input_85_beta_0_to_fp16, x = input_83_cast_fp16)[name = string("input_85_cast_fp16")];
            string input_87_pad_type_0 = const()[name = string("input_87_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> input_87_pad_0 = const()[name = string("input_87_pad_0"), val = tensor<int32, [4]>([1, 1, 1, 1])];
            int32 input_87_groups_0 = const()[name = string("input_87_groups_0"), val = int32(192)];
            tensor<int32, [2]> input_87_strides_0 = const()[name = string("input_87_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [2]> input_87_dilations_0 = const()[name = string("input_87_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [192, 1, 3, 3]> const_60_to_fp16 = const()[name = string("const_60_to_fp16"), val = tensor<fp16, [192, 1, 3, 3]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(92480)))];
            tensor<fp16, [192]> const_61_to_fp16 = const()[name = string("const_61_to_fp16"), val = tensor<fp16, [192]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96000)))];
            tensor<fp16, [1, 192, 40, 40]> input_89_cast_fp16 = conv(bias = const_61_to_fp16, dilations = input_87_dilations_0, groups = input_87_groups_0, pad = input_87_pad_0, pad_type = input_87_pad_type_0, strides = input_87_strides_0, weight = const_60_to_fp16, x = input_85_cast_fp16)[name = string("input_89_cast_fp16")];
            fp16 input_91_alpha_0_to_fp16 = const()[name = string("input_91_alpha_0_to_fp16"), val = fp16(0x0p+0)];
            fp16 input_91_beta_0_to_fp16 = const()[name = string("input_91_beta_0_to_fp16"), val = fp16(0x1.8p+2)];
            tensor<fp16, [1, 192, 40, 40]> input_91_cast_fp16 = clip(alpha = input_91_alpha_0_to_fp16, beta = input_91_beta_0_to_fp16, x = input_89_cast_fp16)[name = string("input_91_cast_fp16")];
            string input_93_pad_type_0 = const()[name = string("input_93_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_93_strides_0 = const()[name = string("input_93_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_93_pad_0 = const()[name = string("input_93_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_93_dilations_0 = const()[name = string("input_93_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_93_groups_0 = const()[name = string("input_93_groups_0"), val = int32(1)];
            tensor<fp16, [32, 192, 1, 1]> const_62_to_fp16 = const()[name = string("const_62_to_fp16"), val = tensor<fp16, [32, 192, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(96448)))];
            tensor<fp16, [32]> const_63_to_fp16 = const()[name = string("const_63_to_fp16"), val = tensor<fp16, [32]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108800)))];
            tensor<fp16, [1, 32, 40, 40]> var_304_cast_fp16 = conv(bias = const_63_to_fp16, dilations = input_93_dilations_0, groups = input_93_groups_0, pad = input_93_pad_0, pad_type = input_93_pad_type_0, strides = input_93_strides_0, weight = const_62_to_fp16, x = input_91_cast_fp16)[name = string("op_304_cast_fp16")];
            tensor<fp16, [1, 32, 40, 40]> input_99_cast_fp16 = add(x = input_79_cast_fp16, y = var_304_cast_fp16)[name = string("input_99_cast_fp16")];
            string input_95_pad_type_0 = const()[name = string("input_95_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_95_strides_0 = const()[name = string("input_95_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_95_pad_0 = const()[name = string("input_95_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_95_dilations_0 = const()[name = string("input_95_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_95_groups_0 = const()[name = string("input_95_groups_0"), val = int32(1)];
            tensor<fp16, [44, 16, 1, 1]> densefeat_df_decoder_conv1_weight_to_fp16 = const()[name = string("densefeat_df_decoder_conv1_weight_to_fp16"), val = tensor<fp16, [44, 16, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(108928)))];
            tensor<fp16, [44]> densefeat_df_decoder_conv1_bias_to_fp16 = const()[name = string("densefeat_df_decoder_conv1_bias_to_fp16"), val = tensor<fp16, [44]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110400)))];
            tensor<fp16, [1, 44, 160, 160]> input_95_cast_fp16 = conv(bias = densefeat_df_decoder_conv1_bias_to_fp16, dilations = input_95_dilations_0, groups = input_95_groups_0, pad = input_95_pad_0, pad_type = input_95_pad_type_0, strides = input_95_strides_0, weight = densefeat_df_decoder_conv1_weight_to_fp16, x = input_15_cast_fp16)[name = string("input_95_cast_fp16")];
            tensor<fp16, [1, 44, 160, 160]> input_103_cast_fp16 = relu(x = input_95_cast_fp16)[name = string("input_103_cast_fp16")];
            string input_97_pad_type_0 = const()[name = string("input_97_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_97_strides_0 = const()[name = string("input_97_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_97_pad_0 = const()[name = string("input_97_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_97_dilations_0 = const()[name = string("input_97_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_97_groups_0 = const()[name = string("input_97_groups_0"), val = int32(1)];
            tensor<fp16, [42, 24, 1, 1]> densefeat_df_decoder_conv2_weight_to_fp16 = const()[name = string("densefeat_df_decoder_conv2_weight_to_fp16"), val = tensor<fp16, [42, 24, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(110592)))];
            tensor<fp16, [42]> densefeat_df_decoder_conv2_bias_to_fp16 = const()[name = string("densefeat_df_decoder_conv2_bias_to_fp16"), val = tensor<fp16, [42]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112704)))];
            tensor<fp16, [1, 42, 80, 80]> input_97_cast_fp16 = conv(bias = densefeat_df_decoder_conv2_bias_to_fp16, dilations = input_97_dilations_0, groups = input_97_groups_0, pad = input_97_pad_0, pad_type = input_97_pad_type_0, strides = input_97_strides_0, weight = densefeat_df_decoder_conv2_weight_to_fp16, x = input_47_cast_fp16)[name = string("input_97_cast_fp16")];
            tensor<fp16, [1, 42, 80, 80]> input_105_cast_fp16 = relu(x = input_97_cast_fp16)[name = string("input_105_cast_fp16")];
            string input_101_pad_type_0 = const()[name = string("input_101_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_101_strides_0 = const()[name = string("input_101_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_101_pad_0 = const()[name = string("input_101_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_101_dilations_0 = const()[name = string("input_101_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_101_groups_0 = const()[name = string("input_101_groups_0"), val = int32(1)];
            tensor<fp16, [42, 32, 1, 1]> densefeat_df_decoder_conv3_weight_to_fp16 = const()[name = string("densefeat_df_decoder_conv3_weight_to_fp16"), val = tensor<fp16, [42, 32, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(112896)))];
            tensor<fp16, [42]> densefeat_df_decoder_conv3_bias_to_fp16 = const()[name = string("densefeat_df_decoder_conv3_bias_to_fp16"), val = tensor<fp16, [42]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115648)))];
            tensor<fp16, [1, 42, 40, 40]> input_101_cast_fp16 = conv(bias = densefeat_df_decoder_conv3_bias_to_fp16, dilations = input_101_dilations_0, groups = input_101_groups_0, pad = input_101_pad_0, pad_type = input_101_pad_type_0, strides = input_101_strides_0, weight = densefeat_df_decoder_conv3_weight_to_fp16, x = input_99_cast_fp16)[name = string("input_101_cast_fp16")];
            tensor<fp16, [1, 42, 40, 40]> input_107_cast_fp16 = relu(x = input_101_cast_fp16)[name = string("input_107_cast_fp16")];
            string input_109_pad_type_0 = const()[name = string("input_109_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_109_strides_0 = const()[name = string("input_109_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_109_pad_0 = const()[name = string("input_109_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_109_dilations_0 = const()[name = string("input_109_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_109_groups_0 = const()[name = string("input_109_groups_0"), val = int32(1)];
            tensor<fp16, [1, 44, 1, 1]> densefeat_df_decoder_convhead2_1_sm_weight_to_fp16 = const()[name = string("densefeat_df_decoder_convhead2_1_sm_weight_to_fp16"), val = tensor<fp16, [1, 44, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(115840)))];
            tensor<fp16, [1]> densefeat_df_decoder_convhead2_1_sm_bias_to_fp16 = const()[name = string("densefeat_df_decoder_convhead2_1_sm_bias_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 1, 160, 160]> input_109_cast_fp16 = conv(bias = densefeat_df_decoder_convhead2_1_sm_bias_to_fp16, dilations = input_109_dilations_0, groups = input_109_groups_0, pad = input_109_pad_0, pad_type = input_109_pad_type_0, strides = input_109_strides_0, weight = densefeat_df_decoder_convhead2_1_sm_weight_to_fp16, x = input_103_cast_fp16)[name = string("input_109_cast_fp16")];
            string input_111_pad_type_0 = const()[name = string("input_111_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_111_strides_0 = const()[name = string("input_111_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_111_pad_0 = const()[name = string("input_111_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_111_dilations_0 = const()[name = string("input_111_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_111_groups_0 = const()[name = string("input_111_groups_0"), val = int32(1)];
            tensor<fp16, [1, 42, 1, 1]> densefeat_df_decoder_convhead2_2_sm_weight_to_fp16 = const()[name = string("densefeat_df_decoder_convhead2_2_sm_weight_to_fp16"), val = tensor<fp16, [1, 42, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116032)))];
            tensor<fp16, [1]> densefeat_df_decoder_convhead2_2_sm_bias_to_fp16 = const()[name = string("densefeat_df_decoder_convhead2_2_sm_bias_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 1, 80, 80]> input_111_cast_fp16 = conv(bias = densefeat_df_decoder_convhead2_2_sm_bias_to_fp16, dilations = input_111_dilations_0, groups = input_111_groups_0, pad = input_111_pad_0, pad_type = input_111_pad_type_0, strides = input_111_strides_0, weight = densefeat_df_decoder_convhead2_2_sm_weight_to_fp16, x = input_105_cast_fp16)[name = string("input_111_cast_fp16")];
            string input_113_pad_type_0 = const()[name = string("input_113_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> input_113_strides_0 = const()[name = string("input_113_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> input_113_pad_0 = const()[name = string("input_113_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> input_113_dilations_0 = const()[name = string("input_113_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 input_113_groups_0 = const()[name = string("input_113_groups_0"), val = int32(1)];
            tensor<fp16, [1, 42, 1, 1]> densefeat_df_decoder_convhead2_3_sm_weight_to_fp16 = const()[name = string("densefeat_df_decoder_convhead2_3_sm_weight_to_fp16"), val = tensor<fp16, [1, 42, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116224)))];
            tensor<fp16, [1]> densefeat_df_decoder_convhead2_3_sm_bias_to_fp16 = const()[name = string("densefeat_df_decoder_convhead2_3_sm_bias_to_fp16"), val = tensor<fp16, [1]>([0x0p+0])];
            tensor<fp16, [1, 1, 40, 40]> input_113_cast_fp16 = conv(bias = densefeat_df_decoder_convhead2_3_sm_bias_to_fp16, dilations = input_113_dilations_0, groups = input_113_groups_0, pad = input_113_pad_0, pad_type = input_113_pad_type_0, strides = input_113_strides_0, weight = densefeat_df_decoder_convhead2_3_sm_weight_to_fp16, x = input_107_cast_fp16)[name = string("input_113_cast_fp16")];
            fp32 var_363_scale_factor_height_0 = const()[name = string("op_363_scale_factor_height_0"), val = fp32(0x1p+1)];
            fp32 var_363_scale_factor_width_0 = const()[name = string("op_363_scale_factor_width_0"), val = fp32(0x1p+1)];
            bool var_363_align_corners_0 = const()[name = string("op_363_align_corners_0"), val = bool(true)];
            tensor<fp16, [1, 1, 320, 320]> var_363_cast_fp16 = upsample_bilinear(align_corners = var_363_align_corners_0, scale_factor_height = var_363_scale_factor_height_0, scale_factor_width = var_363_scale_factor_width_0, x = input_109_cast_fp16)[name = string("op_363_cast_fp16")];
            fp32 var_365_scale_factor_height_0 = const()[name = string("op_365_scale_factor_height_0"), val = fp32(0x1p+2)];
            fp32 var_365_scale_factor_width_0 = const()[name = string("op_365_scale_factor_width_0"), val = fp32(0x1p+2)];
            bool var_365_align_corners_0 = const()[name = string("op_365_align_corners_0"), val = bool(true)];
            tensor<fp16, [1, 1, 320, 320]> var_365_cast_fp16 = upsample_bilinear(align_corners = var_365_align_corners_0, scale_factor_height = var_365_scale_factor_height_0, scale_factor_width = var_365_scale_factor_width_0, x = input_111_cast_fp16)[name = string("op_365_cast_fp16")];
            tensor<fp16, [1, 1, 320, 320]> var_366_cast_fp16 = add(x = var_363_cast_fp16, y = var_365_cast_fp16)[name = string("op_366_cast_fp16")];
            fp32 var_368_scale_factor_height_0 = const()[name = string("op_368_scale_factor_height_0"), val = fp32(0x1p+3)];
            fp32 var_368_scale_factor_width_0 = const()[name = string("op_368_scale_factor_width_0"), val = fp32(0x1p+3)];
            bool var_368_align_corners_0 = const()[name = string("op_368_align_corners_0"), val = bool(true)];
            tensor<fp16, [1, 1, 320, 320]> var_368_cast_fp16 = upsample_bilinear(align_corners = var_368_align_corners_0, scale_factor_height = var_368_scale_factor_height_0, scale_factor_width = var_368_scale_factor_width_0, x = input_113_cast_fp16)[name = string("op_368_cast_fp16")];
            tensor<fp16, [1, 1, 320, 320]> input_115_cast_fp16 = add(x = var_366_cast_fp16, y = var_368_cast_fp16)[name = string("input_115_cast_fp16")];
            string x_score_pad_type_0 = const()[name = string("x_score_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> x_score_strides_0 = const()[name = string("x_score_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> x_score_pad_0 = const()[name = string("x_score_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> x_score_dilations_0 = const()[name = string("x_score_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 x_score_groups_0 = const()[name = string("x_score_groups_0"), val = int32(1)];
            tensor<fp16, [1, 1, 1, 1]> densefeat_df_decoder_convhead2_sm_weight_to_fp16 = const()[name = string("densefeat_df_decoder_convhead2_sm_weight_to_fp16"), val = tensor<fp16, [1, 1, 1, 1]>([[[[0x1p+0]]]])];
            tensor<fp16, [1]> densefeat_df_decoder_convhead2_sm_bias_to_fp16 = const()[name = string("densefeat_df_decoder_convhead2_sm_bias_to_fp16"), val = tensor<fp16, [1]>([0x1.fb4p-4])];
            tensor<fp16, [1, 1, 320, 320]> x_score_cast_fp16 = conv(bias = densefeat_df_decoder_convhead2_sm_bias_to_fp16, dilations = x_score_dilations_0, groups = x_score_groups_0, pad = x_score_pad_0, pad_type = x_score_pad_type_0, strides = x_score_strides_0, weight = densefeat_df_decoder_convhead2_sm_weight_to_fp16, x = input_115_cast_fp16)[name = string("x_score_cast_fp16")];
            tensor<fp16, [1, 1, 320, 320]> scores_map_1_cast_fp16 = sigmoid(x = x_score_cast_fp16)[name = string("scores_map_1_cast_fp16")];
            int32 var_403 = const()[name = string("op_403"), val = int32(32)];
            int32 var_410 = const()[name = string("op_410"), val = int32(0)];
            tensor<int32, [4]> var_426_begin_0 = const()[name = string("op_426_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_426_end_0 = const()[name = string("op_426_end_0"), val = tensor<int32, [4]>([1, 1, 160, 320])];
            tensor<bool, [4]> var_426_end_mask_0 = const()[name = string("op_426_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [1, 1, 160, 320]> var_426_cast_fp16 = slice_by_index(begin = var_426_begin_0, end = var_426_end_0, end_mask = var_426_end_mask_0, x = scores_map_1_cast_fp16)[name = string("op_426_cast_fp16")];
            tensor<int32, [4]> var_427_begin_0 = const()[name = string("op_427_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_427_end_0 = const()[name = string("op_427_end_0"), val = tensor<int32, [4]>([1, 1, 160, 160])];
            tensor<bool, [4]> var_427_end_mask_0 = const()[name = string("op_427_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 160, 160]> var_427_cast_fp16 = slice_by_index(begin = var_427_begin_0, end = var_427_end_0, end_mask = var_427_end_mask_0, x = var_426_cast_fp16)[name = string("op_427_cast_fp16")];
            tensor<int32, [4]> var_433_begin_0 = const()[name = string("op_433_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 160])];
            tensor<int32, [4]> var_433_end_0 = const()[name = string("op_433_end_0"), val = tensor<int32, [4]>([1, 1, 160, 1])];
            tensor<bool, [4]> var_433_end_mask_0 = const()[name = string("op_433_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 160, 160]> var_433_cast_fp16 = slice_by_index(begin = var_433_begin_0, end = var_433_end_0, end_mask = var_433_end_mask_0, x = var_426_cast_fp16)[name = string("op_433_cast_fp16")];
            tensor<int32, [4]> var_438_begin_0 = const()[name = string("op_438_begin_0"), val = tensor<int32, [4]>([0, 0, 160, 0])];
            tensor<int32, [4]> var_438_end_0 = const()[name = string("op_438_end_0"), val = tensor<int32, [4]>([1, 1, 1, 320])];
            tensor<bool, [4]> var_438_end_mask_0 = const()[name = string("op_438_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 160, 320]> var_438_cast_fp16 = slice_by_index(begin = var_438_begin_0, end = var_438_end_0, end_mask = var_438_end_mask_0, x = scores_map_1_cast_fp16)[name = string("op_438_cast_fp16")];
            tensor<int32, [4]> var_439_begin_0 = const()[name = string("op_439_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_439_end_0 = const()[name = string("op_439_end_0"), val = tensor<int32, [4]>([1, 1, 160, 160])];
            tensor<bool, [4]> var_439_end_mask_0 = const()[name = string("op_439_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [1, 1, 160, 160]> var_439_cast_fp16 = slice_by_index(begin = var_439_begin_0, end = var_439_end_0, end_mask = var_439_end_mask_0, x = var_438_cast_fp16)[name = string("op_439_cast_fp16")];
            tensor<int32, [4]> var_447_begin_0 = const()[name = string("op_447_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 160])];
            tensor<int32, [4]> var_447_end_0 = const()[name = string("op_447_end_0"), val = tensor<int32, [4]>([1, 1, 160, 1])];
            tensor<bool, [4]> var_447_end_mask_0 = const()[name = string("op_447_end_mask_0"), val = tensor<bool, [4]>([true, true, true, true])];
            tensor<fp16, [1, 1, 160, 160]> var_447_cast_fp16 = slice_by_index(begin = var_447_begin_0, end = var_447_end_0, end_mask = var_447_end_mask_0, x = var_438_cast_fp16)[name = string("op_447_cast_fp16")];
            bool scores_map_interleave_0 = const()[name = string("scores_map_interleave_0"), val = bool(false)];
            tensor<fp16, [4, 1, 160, 160]> scores_map_cast_fp16 = concat(axis = var_410, interleave = scores_map_interleave_0, values = (var_427_cast_fp16, var_433_cast_fp16, var_439_cast_fp16, var_447_cast_fp16))[name = string("scores_map_cast_fp16")];
            tensor<int32, [2]> var_466 = const()[name = string("op_466"), val = tensor<int32, [2]>([5, 5])];
            tensor<int32, [2]> var_467 = const()[name = string("op_467"), val = tensor<int32, [2]>([1, 1])];
            string var_470_pad_type_0 = const()[name = string("op_470_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_470_pad_0 = const()[name = string("op_470_pad_0"), val = tensor<int32, [4]>([2, 2, 2, 2])];
            bool var_470_ceil_mode_0 = const()[name = string("op_470_ceil_mode_0"), val = bool(false)];
            tensor<fp16, [4, 1, 160, 160]> var_470_cast_fp16 = max_pool(ceil_mode = var_470_ceil_mode_0, kernel_sizes = var_466, pad = var_470_pad_0, pad_type = var_470_pad_type_0, strides = var_467, x = scores_map_cast_fp16)[name = string("op_470_cast_fp16")];
            tensor<bool, [4, 1, 160, 160]> max_mask_1_cast_fp16 = equal(x = scores_map_cast_fp16, y = var_470_cast_fp16)[name = string("max_mask_1_cast_fp16")];
            string max_mask_3_dtype_0 = const()[name = string("max_mask_3_dtype_0"), val = string("fp16")];
            tensor<int32, [2]> var_474 = const()[name = string("op_474"), val = tensor<int32, [2]>([5, 5])];
            tensor<int32, [2]> var_475 = const()[name = string("op_475"), val = tensor<int32, [2]>([1, 1])];
            string var_478_pad_type_0 = const()[name = string("op_478_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_478_pad_0 = const()[name = string("op_478_pad_0"), val = tensor<int32, [4]>([2, 2, 2, 2])];
            bool var_478_ceil_mode_0 = const()[name = string("op_478_ceil_mode_0"), val = bool(false)];
            tensor<fp16, [4, 1, 160, 160]> max_mask_3 = cast(dtype = max_mask_3_dtype_0, x = max_mask_1_cast_fp16)[name = string("cast_48")];
            tensor<fp16, [4, 1, 160, 160]> var_478_cast_fp16 = max_pool(ceil_mode = var_478_ceil_mode_0, kernel_sizes = var_474, pad = var_478_pad_0, pad_type = var_478_pad_type_0, strides = var_475, x = max_mask_3)[name = string("op_478_cast_fp16")];
            fp16 var_410_promoted_to_fp16 = const()[name = string("op_410_promoted_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [4, 1, 160, 160]> supp_mask_1_cast_fp16 = greater(x = var_478_cast_fp16, y = var_410_promoted_to_fp16)[name = string("supp_mask_1_cast_fp16")];
            tensor<fp16, [4, 1, 160, 160]> zeros_to_fp16 = const()[name = string("zeros_to_fp16"), val = tensor<fp16, [4, 1, 160, 160]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(116416)))];
            tensor<fp16, [4, 1, 160, 160]> input_119_cast_fp16 = select(a = zeros_to_fp16, b = scores_map_cast_fp16, cond = supp_mask_1_cast_fp16)[name = string("input_119_cast_fp16")];
            tensor<int32, [2]> var_481 = const()[name = string("op_481"), val = tensor<int32, [2]>([5, 5])];
            tensor<int32, [2]> var_482 = const()[name = string("op_482"), val = tensor<int32, [2]>([1, 1])];
            string var_485_pad_type_0 = const()[name = string("op_485_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_485_pad_0 = const()[name = string("op_485_pad_0"), val = tensor<int32, [4]>([2, 2, 2, 2])];
            bool var_485_ceil_mode_0 = const()[name = string("op_485_ceil_mode_0"), val = bool(false)];
            tensor<fp16, [4, 1, 160, 160]> var_485_cast_fp16 = max_pool(ceil_mode = var_485_ceil_mode_0, kernel_sizes = var_481, pad = var_485_pad_0, pad_type = var_485_pad_type_0, strides = var_482, x = input_119_cast_fp16)[name = string("op_485_cast_fp16")];
            tensor<bool, [4, 1, 160, 160]> var_486_cast_fp16 = equal(x = input_119_cast_fp16, y = var_485_cast_fp16)[name = string("op_486_cast_fp16")];
            string new_max_mask_1_dtype_0 = const()[name = string("new_max_mask_1_dtype_0"), val = string("fp16")];
            tensor<bool, [4, 1, 160, 160]> var_488 = logical_not(x = supp_mask_1_cast_fp16)[name = string("op_488")];
            string var_489_dtype_0 = const()[name = string("op_489_dtype_0"), val = string("fp16")];
            tensor<fp16, [4, 1, 160, 160]> var_489 = cast(dtype = var_489_dtype_0, x = var_488)[name = string("cast_46")];
            tensor<fp16, [4, 1, 160, 160]> new_max_mask_1 = cast(dtype = new_max_mask_1_dtype_0, x = var_486_cast_fp16)[name = string("cast_47")];
            tensor<fp16, [4, 1, 160, 160]> var_490 = minimum(x = new_max_mask_1, y = var_489)[name = string("op_490")];
            tensor<fp16, [4, 1, 160, 160]> max_mask_5 = maximum(x = max_mask_3, y = var_490)[name = string("max_mask_5")];
            tensor<int32, [2]> var_493 = const()[name = string("op_493"), val = tensor<int32, [2]>([5, 5])];
            tensor<int32, [2]> var_494 = const()[name = string("op_494"), val = tensor<int32, [2]>([1, 1])];
            string var_497_pad_type_0 = const()[name = string("op_497_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_497_pad_0 = const()[name = string("op_497_pad_0"), val = tensor<int32, [4]>([2, 2, 2, 2])];
            bool var_497_ceil_mode_0 = const()[name = string("op_497_ceil_mode_0"), val = bool(false)];
            tensor<fp16, [4, 1, 160, 160]> var_497_cast_fp16 = max_pool(ceil_mode = var_497_ceil_mode_0, kernel_sizes = var_493, pad = var_497_pad_0, pad_type = var_497_pad_type_0, strides = var_494, x = max_mask_5)[name = string("op_497_cast_fp16")];
            fp16 var_410_promoted_1_to_fp16 = const()[name = string("op_410_promoted_1_to_fp16"), val = fp16(0x0p+0)];
            tensor<bool, [4, 1, 160, 160]> supp_mask_cast_fp16 = greater(x = var_497_cast_fp16, y = var_410_promoted_1_to_fp16)[name = string("supp_mask_cast_fp16")];
            tensor<fp16, [4, 1, 160, 160]> input_123_cast_fp16 = select(a = zeros_to_fp16, b = scores_map_cast_fp16, cond = supp_mask_cast_fp16)[name = string("input_123_cast_fp16")];
            tensor<int32, [2]> var_500 = const()[name = string("op_500"), val = tensor<int32, [2]>([5, 5])];
            tensor<int32, [2]> var_501 = const()[name = string("op_501"), val = tensor<int32, [2]>([1, 1])];
            string var_504_pad_type_0 = const()[name = string("op_504_pad_type_0"), val = string("custom")];
            tensor<int32, [4]> var_504_pad_0 = const()[name = string("op_504_pad_0"), val = tensor<int32, [4]>([2, 2, 2, 2])];
            bool var_504_ceil_mode_0 = const()[name = string("op_504_ceil_mode_0"), val = bool(false)];
            tensor<fp16, [4, 1, 160, 160]> var_504_cast_fp16 = max_pool(ceil_mode = var_504_ceil_mode_0, kernel_sizes = var_500, pad = var_504_pad_0, pad_type = var_504_pad_type_0, strides = var_501, x = input_123_cast_fp16)[name = string("op_504_cast_fp16")];
            tensor<bool, [4, 1, 160, 160]> var_505_cast_fp16 = equal(x = input_123_cast_fp16, y = var_504_cast_fp16)[name = string("op_505_cast_fp16")];
            string new_max_mask_dtype_0 = const()[name = string("new_max_mask_dtype_0"), val = string("fp16")];
            tensor<bool, [4, 1, 160, 160]> var_507 = logical_not(x = supp_mask_cast_fp16)[name = string("op_507")];
            string var_508_dtype_0 = const()[name = string("op_508_dtype_0"), val = string("fp16")];
            tensor<fp16, [4, 1, 160, 160]> var_508 = cast(dtype = var_508_dtype_0, x = var_507)[name = string("cast_44")];
            tensor<fp16, [4, 1, 160, 160]> new_max_mask = cast(dtype = new_max_mask_dtype_0, x = var_505_cast_fp16)[name = string("cast_45")];
            tensor<fp16, [4, 1, 160, 160]> var_509 = minimum(x = new_max_mask, y = var_508)[name = string("op_509")];
            tensor<fp16, [4, 1, 160, 160]> max_mask = maximum(x = max_mask_5, y = var_509)[name = string("max_mask")];
            string var_511_dtype_0 = const()[name = string("op_511_dtype_0"), val = string("bool")];
            tensor<bool, [4, 1, 160, 160]> var_511 = cast(dtype = var_511_dtype_0, x = max_mask)[name = string("cast_43")];
            tensor<fp16, [4, 1, 160, 160]> nms_scores_1_cast_fp16 = select(a = scores_map_cast_fp16, b = zeros_to_fp16, cond = var_511)[name = string("nms_scores_1_cast_fp16")];
            tensor<int32, [4]> var_515_begin_0 = const()[name = string("op_515_begin_0"), val = tensor<int32, [4]>([0, 0, 2, 0])];
            tensor<int32, [4]> var_515_end_0 = const()[name = string("op_515_end_0"), val = tensor<int32, [4]>([4, 1, 158, 160])];
            tensor<bool, [4]> var_515_end_mask_0 = const()[name = string("op_515_end_mask_0"), val = tensor<bool, [4]>([true, true, false, true])];
            tensor<fp16, [4, 1, 156, 160]> var_515_cast_fp16 = slice_by_index(begin = var_515_begin_0, end = var_515_end_0, end_mask = var_515_end_mask_0, x = nms_scores_1_cast_fp16)[name = string("op_515_cast_fp16")];
            tensor<int32, [4]> input_125_begin_0 = const()[name = string("input_125_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 2])];
            tensor<int32, [4]> input_125_end_0 = const()[name = string("input_125_end_0"), val = tensor<int32, [4]>([4, 1, 156, 158])];
            tensor<bool, [4]> input_125_end_mask_0 = const()[name = string("input_125_end_mask_0"), val = tensor<bool, [4]>([true, true, true, false])];
            tensor<fp16, [4, 1, 156, 156]> input_125_cast_fp16 = slice_by_index(begin = input_125_begin_0, end = input_125_end_0, end_mask = input_125_end_mask_0, x = var_515_cast_fp16)[name = string("input_125_cast_fp16")];
            tensor<int32, [8]> nms_scores_pad_0 = const()[name = string("nms_scores_pad_0"), val = tensor<int32, [8]>([0, 0, 0, 0, 2, 2, 2, 2])];
            string nms_scores_mode_0 = const()[name = string("nms_scores_mode_0"), val = string("constant")];
            fp16 const_13_to_fp16 = const()[name = string("const_13_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [4, 1, 160, 160]> nms_scores_cast_fp16 = pad(constant_val = const_13_to_fp16, mode = nms_scores_mode_0, pad = nms_scores_pad_0, x = input_125_cast_fp16)[name = string("nms_scores_cast_fp16")];
            tensor<int32, [2]> var_519 = const()[name = string("op_519"), val = tensor<int32, [2]>([4, -1])];
            tensor<fp16, [4, 25600]> var_520_cast_fp16 = reshape(shape = var_519, x = nms_scores_cast_fp16)[name = string("op_520_cast_fp16")];
            int32 var_521_axis_0 = const()[name = string("op_521_axis_0"), val = int32(-1)];
            bool var_521_ascending_0 = const()[name = string("op_521_ascending_0"), val = bool(false)];
            bool var_521_sort_0 = const()[name = string("op_521_sort_0"), val = bool(true)];
            bool var_521_return_indices_0 = const()[name = string("op_521_return_indices_0"), val = bool(true)];
            string var_521_cast_fp16_cast_int16_output_indices_dtype_0 = const()[name = string("op_521_cast_fp16_cast_int16_output_indices_dtype_0"), val = string("uint16")];
            tensor<fp16, [4, 32]> var_521_cast_fp16_cast_int16_0, tensor<uint16, [4, 32]> var_521_cast_fp16_cast_int16_1 = topk(ascending = var_521_ascending_0, axis = var_521_axis_0, k = var_403, output_indices_dtype = var_521_cast_fp16_cast_int16_output_indices_dtype_0, return_indices = var_521_return_indices_0, sort = var_521_sort_0, x = var_520_cast_fp16)[name = string("op_521_cast_fp16_cast_int16")];
            int32 var_525_axis_0 = const()[name = string("op_525_axis_0"), val = int32(0)];
            int32 var_525_batch_dims_0 = const()[name = string("op_525_batch_dims_0"), val = int32(0)];
            bool var_525_validate_indices_0 = const()[name = string("op_525_validate_indices_0"), val = bool(false)];
            tensor<uint16, [25600]> var_523_to_uint16 = const()[name = string("op_523_to_uint16"), val = tensor<uint16, [25600]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(321280)))];
            int32 var_529_axis_0 = const()[name = string("op_529_axis_0"), val = int32(0)];
            int32 var_529_batch_dims_0 = const()[name = string("op_529_batch_dims_0"), val = int32(0)];
            bool var_529_validate_indices_0 = const()[name = string("op_529_validate_indices_0"), val = bool(false)];
            tensor<uint16, [25600]> var_527_to_uint16 = const()[name = string("op_527_to_uint16"), val = tensor<uint16, [25600]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(372544)))];
            // BEFORE
            tensor<uint16, [4, 32]> var_525_cast_uint16 = gather(axis = var_525_axis_0, batch_dims = var_525_batch_dims_0, indices = var_521_cast_fp16_cast_int16_1, validate_indices = var_525_validate_indices_0, x = var_523_to_uint16)[name = string("op_525_cast_uint16")];
            tensor<uint16, [4, 32]> var_529_cast_uint16 = gather(axis = var_529_axis_0, batch_dims = var_529_batch_dims_0, indices = var_521_cast_fp16_cast_int16_1, validate_indices = var_529_validate_indices_0, x = var_527_to_uint16)[name = string("op_529_cast_uint16")];
            
            int32 keypoints_1_axis_0 = const()[name = string("keypoints_1_axis_0"), val = int32(-1)];
            string var_526_to_fp16_dtype_0 = const()[name = string("op_526_to_fp16_dtype_0"), val = string("fp16")];
            string var_530_to_fp16_dtype_0 = const()[name = string("op_530_to_fp16_dtype_0"), val = string("fp16")];
            tensor<fp16, [4, 32]> var_529_cast_uint16_to_fp16 = cast(dtype = var_530_to_fp16_dtype_0, x = var_529_cast_uint16)[name = string("cast_41")];
            tensor<fp16, [4, 32]> var_525_cast_uint16_to_fp16 = cast(dtype = var_526_to_fp16_dtype_0, x = var_525_cast_uint16)[name = string("cast_42")];
            tensor<fp16, [4, 32, 2]> keypoints_1_cast_fp16 = stack(axis = keypoints_1_axis_0, values = (var_525_cast_uint16_to_fp16, var_529_cast_uint16_to_fp16))[name = string("keypoints_1_cast_fp16")];
            tensor<int32, [1]> var_564_axes_0 = const()[name = string("op_564_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 32, 1, 2]> var_564_cast_fp16 = expand_dims(axes = var_564_axes_0, x = keypoints_1_cast_fp16)[name = string("op_564_cast_fp16")];
            tensor<fp16, [1, 25, 2]> var_559_to_fp16 = const()[name = string("op_559_to_fp16"), val = tensor<fp16, [1, 25, 2]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(423808)))];
            tensor<fp16, [4, 32, 25, 2]> patch_indexes_cast_fp16 = add(x = var_559_to_fp16, y = var_564_cast_fp16)[name = string("patch_indexes_cast_fp16")];
            tensor<int32, [3]> var_567 = const()[name = string("op_567"), val = tensor<int32, [3]>([4, -1, 2])];
            tensor<fp16, [4, 800, 2]> keypoints_9_cast_fp16 = reshape(shape = var_567, x = patch_indexes_cast_fp16)[name = string("keypoints_9_cast_fp16")];
            fp16 var_574_to_fp16 = const()[name = string("op_574_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [4, 800, 2]> var_575_cast_fp16 = add(x = keypoints_9_cast_fp16, y = var_574_to_fp16)[name = string("op_575_cast_fp16")];
            tensor<fp16, [2]> _inversed_keypoints_13_y_0_to_fp16 = const()[name = string("_inversed_keypoints_13_y_0_to_fp16"), val = tensor<fp16, [2]>([0x1.998p-8, 0x1.998p-8])];
            tensor<fp16, [4, 800, 2]> _inversed_keypoints_13_cast_fp16 = mul(x = var_575_cast_fp16, y = _inversed_keypoints_13_y_0_to_fp16)[name = string("_inversed_keypoints_13_cast_fp16")];
            tensor<fp16, [2]> const_18_promoted_to_fp16 = const()[name = string("const_18_promoted_to_fp16"), val = tensor<fp16, [2]>([0x1.4p+7, 0x1.4p+7])];
            tensor<fp16, [4, 800, 2]> var_577_cast_fp16 = mul(x = _inversed_keypoints_13_cast_fp16, y = const_18_promoted_to_fp16)[name = string("op_577_cast_fp16")];
            fp16 var_578_to_fp16 = const()[name = string("op_578_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [4, 800, 2]> keypoints_15_cast_fp16 = sub(x = var_577_cast_fp16, y = var_578_to_fp16)[name = string("keypoints_15_cast_fp16")];
            fp16 var_580_promoted_to_fp16 = const()[name = string("op_580_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [4, 800, 2]> var_581_cast_fp16 = mul(x = keypoints_15_cast_fp16, y = var_580_promoted_to_fp16)[name = string("op_581_cast_fp16")];
            tensor<fp16, [2]> _inversed_584_y_0_to_fp16 = const()[name = string("_inversed_584_y_0_to_fp16"), val = tensor<fp16, [2]>([0x1.9c4p-8, 0x1.9c4p-8])];
            tensor<fp16, [4, 800, 2]> _inversed_584_cast_fp16 = mul(x = var_581_cast_fp16, y = _inversed_584_y_0_to_fp16)[name = string("_inversed_584_cast_fp16")];
            fp16 var_585_to_fp16 = const()[name = string("op_585_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [4, 800, 2]> keypoints_xy_norm_cast_fp16 = sub(x = _inversed_584_cast_fp16, y = var_585_to_fp16)[name = string("keypoints_xy_norm_cast_fp16")];
            tensor<int32, [4]> var_588 = const()[name = string("op_588"), val = tensor<int32, [4]>([4, 1, -1, 2])];
            tensor<fp16, [4, 1, 800, 2]> grid_3_cast_fp16 = reshape(shape = var_588, x = keypoints_xy_norm_cast_fp16)[name = string("grid_3_cast_fp16")];
            string var_590_sampling_mode_0 = const()[name = string("op_590_sampling_mode_0"), val = string("bilinear")];
            string var_590_padding_mode_0 = const()[name = string("op_590_padding_mode_0"), val = string("constant")];
            string var_590_coordinates_mode_0 = const()[name = string("op_590_coordinates_mode_0"), val = string("normalized_minus_one_to_one")];
            bool var_590_align_corners_0 = const()[name = string("op_590_align_corners_0"), val = bool(true)];
            fp16 var_590_padding_value_0_to_fp16 = const()[name = string("op_590_padding_value_0_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [4, 1, 1, 800]> var_590_cast_fp16 = resample(align_corners = var_590_align_corners_0, coordinates = grid_3_cast_fp16, coordinates_format = string("xy"), coordinates_mode = var_590_coordinates_mode_0, coordinates_type = string("absolute"), padding_mode = var_590_padding_mode_0, padding_value = var_590_padding_value_0_to_fp16, sampling_mode = var_590_sampling_mode_0, x = scores_map_cast_fp16)[name = string("op_590_cast_fp16")];
            tensor<int32, [4]> var_592_begin_0 = const()[name = string("op_592_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_592_end_0 = const()[name = string("op_592_end_0"), val = tensor<int32, [4]>([4, 1, 1, 800])];
            tensor<bool, [4]> var_592_end_mask_0 = const()[name = string("op_592_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<bool, [4]> var_592_squeeze_mask_0 = const()[name = string("op_592_squeeze_mask_0"), val = tensor<bool, [4]>([false, true, false, false])];
            tensor<fp16, [4, 1, 800]> var_592_cast_fp16 = slice_by_index(begin = var_592_begin_0, end = var_592_end_0, end_mask = var_592_end_mask_0, squeeze_mask = var_592_squeeze_mask_0, x = var_590_cast_fp16)[name = string("op_592_cast_fp16")];
            tensor<int32, [3]> var_593_begin_0 = const()[name = string("op_593_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_593_end_0 = const()[name = string("op_593_end_0"), val = tensor<int32, [3]>([4, 1, 800])];
            tensor<bool, [3]> var_593_end_mask_0 = const()[name = string("op_593_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_593_squeeze_mask_0 = const()[name = string("op_593_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [4, 800]> var_593_cast_fp16 = slice_by_index(begin = var_593_begin_0, end = var_593_end_0, end_mask = var_593_end_mask_0, squeeze_mask = var_593_squeeze_mask_0, x = var_592_cast_fp16)[name = string("op_593_cast_fp16")];
            tensor<int32, [3]> var_595 = const()[name = string("op_595"), val = tensor<int32, [3]>([4, 32, -1])];
            tensor<fp16, [4, 32, 25]> patch_scores_cast_fp16 = reshape(shape = var_595, x = var_593_cast_fp16)[name = string("patch_scores_cast_fp16")];
            tensor<int32, [1]> reduce_max_0_axes_0 = const()[name = string("reduce_max_0_axes_0"), val = tensor<int32, [1]>([2])];
            bool reduce_max_0_keep_dims_0 = const()[name = string("reduce_max_0_keep_dims_0"), val = bool(false)];
            tensor<fp16, [4, 32]> reduce_max_0_cast_fp16 = reduce_max(axes = reduce_max_0_axes_0, keep_dims = reduce_max_0_keep_dims_0, x = patch_scores_cast_fp16)[name = string("reduce_max_0_cast_fp16")];
            tensor<int32, [1]> max_v_axes_0 = const()[name = string("max_v_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 32, 1]> max_v_cast_fp16 = expand_dims(axes = max_v_axes_0, x = reduce_max_0_cast_fp16)[name = string("max_v_cast_fp16")];
            tensor<fp16, [4, 32, 25]> var_601_cast_fp16 = sub(x = patch_scores_cast_fp16, y = max_v_cast_fp16)[name = string("op_601_cast_fp16")];
            fp16 _inversed_x_1_y_0_to_fp16 = const()[name = string("_inversed_x_1_y_0_to_fp16"), val = fp16(0x1.4p+3)];
            tensor<fp16, [4, 32, 25]> _inversed_x_1_cast_fp16 = mul(x = var_601_cast_fp16, y = _inversed_x_1_y_0_to_fp16)[name = string("_inversed_x_1_cast_fp16")];
            tensor<fp16, [4, 32, 25]> exp_x_cast_fp16 = exp(x = _inversed_x_1_cast_fp16)[name = string("exp_x_cast_fp16")];
            tensor<int32, [1]> exp_x_sum_axes_0 = const()[name = string("exp_x_sum_axes_0"), val = tensor<int32, [1]>([2])];
            bool exp_x_sum_keep_dims_0 = const()[name = string("exp_x_sum_keep_dims_0"), val = bool(false)];
            tensor<fp16, [4, 32]> exp_x_sum_cast_fp16 = reduce_sum(axes = exp_x_sum_axes_0, keep_dims = exp_x_sum_keep_dims_0, x = exp_x_cast_fp16)[name = string("exp_x_sum_cast_fp16")];
            tensor<fp16, [2, 25]> transpose_0_to_fp16 = const()[name = string("transpose_0_to_fp16"), val = tensor<fp16, [2, 25]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424000)))];
            tensor<fp16, [2]> var_607_bias_0_to_fp16 = const()[name = string("op_607_bias_0_to_fp16"), val = tensor<fp16, [2]>([0x0p+0, 0x0p+0])];
            tensor<fp16, [4, 32, 2]> var_607_cast_fp16 = linear(bias = var_607_bias_0_to_fp16, weight = transpose_0_to_fp16, x = exp_x_cast_fp16)[name = string("op_607_cast_fp16")];
            tensor<int32, [1]> var_608_axes_0 = const()[name = string("op_608_axes_0"), val = tensor<int32, [1]>([2])];
            tensor<fp16, [4, 32, 1]> var_608_cast_fp16 = expand_dims(axes = var_608_axes_0, x = exp_x_sum_cast_fp16)[name = string("op_608_cast_fp16")];
            tensor<fp16, [4, 32, 2]> xy_residual_cast_fp16 = real_div(x = var_607_cast_fp16, y = var_608_cast_fp16)[name = string("xy_residual_cast_fp16")];
            tensor<fp16, [4, 32, 2]> keypoints_xy_1_cast_fp16 = add(x = keypoints_1_cast_fp16, y = xy_residual_cast_fp16)[name = string("keypoints_xy_1_cast_fp16")];
            tensor<fp16, [4, 1, 2]> var_617_to_fp16 = const()[name = string("op_617_to_fp16"), val = tensor<fp16, [4, 1, 2]>([[[0x0p+0, 0x0p+0]], [[0x1.4p+7, 0x0p+0]], [[0x0p+0, 0x1.4p+7]], [[0x1.4p+7, 0x1.4p+7]]])];
            tensor<fp16, [4, 32, 2]> keypoints_xy_3_cast_fp16 = add(x = keypoints_xy_1_cast_fp16, y = var_617_to_fp16)[name = string("keypoints_xy_3_cast_fp16")];
            tensor<int32, [3]> var_620 = const()[name = string("op_620"), val = tensor<int32, [3]>([1, -1, 2])];
            tensor<fp16, [1, 128, 2]> keypoints_xy_cast_fp16 = reshape(shape = var_620, x = keypoints_xy_3_cast_fp16)[name = string("keypoints_xy_cast_fp16")];
            tensor<int32, [3]> keypoints_17_begin_0 = const()[name = string("keypoints_17_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> keypoints_17_end_0 = const()[name = string("keypoints_17_end_0"), val = tensor<int32, [3]>([1, 128, 2])];
            tensor<bool, [3]> keypoints_17_end_mask_0 = const()[name = string("keypoints_17_end_mask_0"), val = tensor<bool, [3]>([false, true, true])];
            tensor<bool, [3]> keypoints_17_squeeze_mask_0 = const()[name = string("keypoints_17_squeeze_mask_0"), val = tensor<bool, [3]>([true, false, false])];
            tensor<fp16, [128, 2]> keypoints_17_cast_fp16 = slice_by_index(begin = keypoints_17_begin_0, end = keypoints_17_end_0, end_mask = keypoints_17_end_mask_0, squeeze_mask = keypoints_17_squeeze_mask_0, x = keypoints_xy_cast_fp16)[name = string("keypoints_17_cast_fp16")];
            tensor<int32, [4]> var_623_begin_0 = const()[name = string("op_623_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_623_end_0 = const()[name = string("op_623_end_0"), val = tensor<int32, [4]>([1, 44, 160, 160])];
            tensor<bool, [4]> var_623_end_mask_0 = const()[name = string("op_623_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_623_squeeze_mask_0 = const()[name = string("op_623_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [44, 160, 160]> var_623_cast_fp16 = slice_by_index(begin = var_623_begin_0, end = var_623_end_0, end_mask = var_623_end_mask_0, squeeze_mask = var_623_squeeze_mask_0, x = input_103_cast_fp16)[name = string("op_623_cast_fp16")];
            tensor<int32, [1]> input_127_axes_0 = const()[name = string("input_127_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 44, 160, 160]> input_127_cast_fp16 = expand_dims(axes = input_127_axes_0, x = var_623_cast_fp16)[name = string("input_127_cast_fp16")];
            fp16 var_630_to_fp16 = const()[name = string("op_630_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [128, 2]> var_631_cast_fp16 = add(x = keypoints_17_cast_fp16, y = var_630_to_fp16)[name = string("op_631_cast_fp16")];
            tensor<fp16, [2]> _inversed_keypoints_21_y_0_to_fp16 = const()[name = string("_inversed_keypoints_21_y_0_to_fp16"), val = tensor<fp16, [2]>([0x1.998p-9, 0x1.998p-9])];
            tensor<fp16, [128, 2]> _inversed_keypoints_21_cast_fp16 = mul(x = var_631_cast_fp16, y = _inversed_keypoints_21_y_0_to_fp16)[name = string("_inversed_keypoints_21_cast_fp16")];
            tensor<fp16, [2]> const_23_promoted_to_fp16 = const()[name = string("const_23_promoted_to_fp16"), val = tensor<fp16, [2]>([0x1.4p+7, 0x1.4p+7])];
            tensor<fp16, [128, 2]> var_633_cast_fp16 = mul(x = _inversed_keypoints_21_cast_fp16, y = const_23_promoted_to_fp16)[name = string("op_633_cast_fp16")];
            fp16 var_634_to_fp16 = const()[name = string("op_634_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [128, 2]> keypoints_23_cast_fp16 = sub(x = var_633_cast_fp16, y = var_634_to_fp16)[name = string("keypoints_23_cast_fp16")];
            fp16 var_636_promoted_to_fp16 = const()[name = string("op_636_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [128, 2]> var_637_cast_fp16 = mul(x = keypoints_23_cast_fp16, y = var_636_promoted_to_fp16)[name = string("op_637_cast_fp16")];
            tensor<fp16, [2]> _inversed_640_y_0_to_fp16 = const()[name = string("_inversed_640_y_0_to_fp16"), val = tensor<fp16, [2]>([0x1.9c4p-8, 0x1.9c4p-8])];
            tensor<fp16, [128, 2]> _inversed_640_cast_fp16 = mul(x = var_637_cast_fp16, y = _inversed_640_y_0_to_fp16)[name = string("_inversed_640_cast_fp16")];
            fp16 var_641_to_fp16 = const()[name = string("op_641_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [128, 2]> keypoints_i_norm_1_cast_fp16 = sub(x = _inversed_640_cast_fp16, y = var_641_to_fp16)[name = string("keypoints_i_norm_1_cast_fp16")];
            tensor<int32, [4]> var_643 = const()[name = string("op_643"), val = tensor<int32, [4]>([1, 1, -1, 2])];
            tensor<fp16, [1, 1, 128, 2]> grid_5_cast_fp16 = reshape(shape = var_643, x = keypoints_i_norm_1_cast_fp16)[name = string("grid_5_cast_fp16")];
            string var_645_sampling_mode_0 = const()[name = string("op_645_sampling_mode_0"), val = string("bilinear")];
            string var_645_padding_mode_0 = const()[name = string("op_645_padding_mode_0"), val = string("border")];
            string var_645_coordinates_mode_0 = const()[name = string("op_645_coordinates_mode_0"), val = string("normalized_minus_one_to_one")];
            bool var_645_align_corners_0 = const()[name = string("op_645_align_corners_0"), val = bool(true)];
            fp16 var_645_padding_value_0_to_fp16 = const()[name = string("op_645_padding_value_0_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 44, 1, 128]> var_645_cast_fp16 = resample(align_corners = var_645_align_corners_0, coordinates = grid_5_cast_fp16, coordinates_format = string("xy"), coordinates_mode = var_645_coordinates_mode_0, coordinates_type = string("absolute"), padding_mode = var_645_padding_mode_0, padding_value = var_645_padding_value_0_to_fp16, sampling_mode = var_645_sampling_mode_0, x = input_127_cast_fp16)[name = string("op_645_cast_fp16")];
            tensor<int32, [4]> var_646_begin_0 = const()[name = string("op_646_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_646_end_0 = const()[name = string("op_646_end_0"), val = tensor<int32, [4]>([1, 44, 1, 128])];
            tensor<bool, [4]> var_646_end_mask_0 = const()[name = string("op_646_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_646_squeeze_mask_0 = const()[name = string("op_646_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [44, 1, 128]> var_646_cast_fp16 = slice_by_index(begin = var_646_begin_0, end = var_646_end_0, end_mask = var_646_end_mask_0, squeeze_mask = var_646_squeeze_mask_0, x = var_645_cast_fp16)[name = string("op_646_cast_fp16")];
            tensor<int32, [3]> var_648_begin_0 = const()[name = string("op_648_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_648_end_0 = const()[name = string("op_648_end_0"), val = tensor<int32, [3]>([44, 1, 128])];
            tensor<bool, [3]> var_648_end_mask_0 = const()[name = string("op_648_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_648_squeeze_mask_0 = const()[name = string("op_648_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [44, 128]> var_648_cast_fp16 = slice_by_index(begin = var_648_begin_0, end = var_648_end_0, end_mask = var_648_end_mask_0, squeeze_mask = var_648_squeeze_mask_0, x = var_646_cast_fp16)[name = string("op_648_cast_fp16")];
            tensor<int32, [4]> var_651_begin_0 = const()[name = string("op_651_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_651_end_0 = const()[name = string("op_651_end_0"), val = tensor<int32, [4]>([1, 42, 80, 80])];
            tensor<bool, [4]> var_651_end_mask_0 = const()[name = string("op_651_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_651_squeeze_mask_0 = const()[name = string("op_651_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [42, 80, 80]> var_651_cast_fp16 = slice_by_index(begin = var_651_begin_0, end = var_651_end_0, end_mask = var_651_end_mask_0, squeeze_mask = var_651_squeeze_mask_0, x = input_105_cast_fp16)[name = string("op_651_cast_fp16")];
            tensor<int32, [1]> input_129_axes_0 = const()[name = string("input_129_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 42, 80, 80]> input_129_cast_fp16 = expand_dims(axes = input_129_axes_0, x = var_651_cast_fp16)[name = string("input_129_cast_fp16")];
            tensor<fp16, [2]> const_25_promoted_to_fp16 = const()[name = string("const_25_promoted_to_fp16"), val = tensor<fp16, [2]>([0x1.4p+6, 0x1.4p+6])];
            tensor<fp16, [128, 2]> var_661_cast_fp16 = mul(x = _inversed_keypoints_21_cast_fp16, y = const_25_promoted_to_fp16)[name = string("op_661_cast_fp16")];
            fp16 var_662_to_fp16 = const()[name = string("op_662_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [128, 2]> keypoints_29_cast_fp16 = sub(x = var_661_cast_fp16, y = var_662_to_fp16)[name = string("keypoints_29_cast_fp16")];
            fp16 var_664_promoted_to_fp16 = const()[name = string("op_664_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [128, 2]> var_665_cast_fp16 = mul(x = keypoints_29_cast_fp16, y = var_664_promoted_to_fp16)[name = string("op_665_cast_fp16")];
            tensor<fp16, [2]> _inversed_668_y_0_to_fp16 = const()[name = string("_inversed_668_y_0_to_fp16"), val = tensor<fp16, [2]>([0x1.9ecp-7, 0x1.9ecp-7])];
            tensor<fp16, [128, 2]> _inversed_668_cast_fp16 = mul(x = var_665_cast_fp16, y = _inversed_668_y_0_to_fp16)[name = string("_inversed_668_cast_fp16")];
            fp16 var_669_to_fp16 = const()[name = string("op_669_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [128, 2]> keypoints_i_norm_3_cast_fp16 = sub(x = _inversed_668_cast_fp16, y = var_669_to_fp16)[name = string("keypoints_i_norm_3_cast_fp16")];
            tensor<int32, [4]> var_671 = const()[name = string("op_671"), val = tensor<int32, [4]>([1, 1, -1, 2])];
            tensor<fp16, [1, 1, 128, 2]> grid_7_cast_fp16 = reshape(shape = var_671, x = keypoints_i_norm_3_cast_fp16)[name = string("grid_7_cast_fp16")];
            string var_673_sampling_mode_0 = const()[name = string("op_673_sampling_mode_0"), val = string("bilinear")];
            string var_673_padding_mode_0 = const()[name = string("op_673_padding_mode_0"), val = string("border")];
            string var_673_coordinates_mode_0 = const()[name = string("op_673_coordinates_mode_0"), val = string("normalized_minus_one_to_one")];
            bool var_673_align_corners_0 = const()[name = string("op_673_align_corners_0"), val = bool(true)];
            fp16 var_673_padding_value_0_to_fp16 = const()[name = string("op_673_padding_value_0_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 42, 1, 128]> var_673_cast_fp16 = resample(align_corners = var_673_align_corners_0, coordinates = grid_7_cast_fp16, coordinates_format = string("xy"), coordinates_mode = var_673_coordinates_mode_0, coordinates_type = string("absolute"), padding_mode = var_673_padding_mode_0, padding_value = var_673_padding_value_0_to_fp16, sampling_mode = var_673_sampling_mode_0, x = input_129_cast_fp16)[name = string("op_673_cast_fp16")];
            tensor<int32, [4]> var_674_begin_0 = const()[name = string("op_674_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_674_end_0 = const()[name = string("op_674_end_0"), val = tensor<int32, [4]>([1, 42, 1, 128])];
            tensor<bool, [4]> var_674_end_mask_0 = const()[name = string("op_674_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_674_squeeze_mask_0 = const()[name = string("op_674_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [42, 1, 128]> var_674_cast_fp16 = slice_by_index(begin = var_674_begin_0, end = var_674_end_0, end_mask = var_674_end_mask_0, squeeze_mask = var_674_squeeze_mask_0, x = var_673_cast_fp16)[name = string("op_674_cast_fp16")];
            tensor<int32, [3]> var_676_begin_0 = const()[name = string("op_676_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_676_end_0 = const()[name = string("op_676_end_0"), val = tensor<int32, [3]>([42, 1, 128])];
            tensor<bool, [3]> var_676_end_mask_0 = const()[name = string("op_676_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_676_squeeze_mask_0 = const()[name = string("op_676_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [42, 128]> var_676_cast_fp16 = slice_by_index(begin = var_676_begin_0, end = var_676_end_0, end_mask = var_676_end_mask_0, squeeze_mask = var_676_squeeze_mask_0, x = var_674_cast_fp16)[name = string("op_676_cast_fp16")];
            tensor<int32, [4]> var_679_begin_0 = const()[name = string("op_679_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_679_end_0 = const()[name = string("op_679_end_0"), val = tensor<int32, [4]>([1, 42, 40, 40])];
            tensor<bool, [4]> var_679_end_mask_0 = const()[name = string("op_679_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_679_squeeze_mask_0 = const()[name = string("op_679_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [42, 40, 40]> var_679_cast_fp16 = slice_by_index(begin = var_679_begin_0, end = var_679_end_0, end_mask = var_679_end_mask_0, squeeze_mask = var_679_squeeze_mask_0, x = input_107_cast_fp16)[name = string("op_679_cast_fp16")];
            tensor<int32, [1]> input_131_axes_0 = const()[name = string("input_131_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 42, 40, 40]> input_131_cast_fp16 = expand_dims(axes = input_131_axes_0, x = var_679_cast_fp16)[name = string("input_131_cast_fp16")];
            tensor<fp16, [2]> const_27_promoted_to_fp16 = const()[name = string("const_27_promoted_to_fp16"), val = tensor<fp16, [2]>([0x1.4p+5, 0x1.4p+5])];
            tensor<fp16, [128, 2]> var_689_cast_fp16 = mul(x = _inversed_keypoints_21_cast_fp16, y = const_27_promoted_to_fp16)[name = string("op_689_cast_fp16")];
            fp16 var_690_to_fp16 = const()[name = string("op_690_to_fp16"), val = fp16(0x1p-1)];
            tensor<fp16, [128, 2]> keypoints_cast_fp16 = sub(x = var_689_cast_fp16, y = var_690_to_fp16)[name = string("keypoints_cast_fp16")];
            fp16 var_692_promoted_to_fp16 = const()[name = string("op_692_promoted_to_fp16"), val = fp16(0x1p+1)];
            tensor<fp16, [128, 2]> var_693_cast_fp16 = mul(x = keypoints_cast_fp16, y = var_692_promoted_to_fp16)[name = string("op_693_cast_fp16")];
            tensor<fp16, [2]> _inversed_696_y_0_to_fp16 = const()[name = string("_inversed_696_y_0_to_fp16"), val = tensor<fp16, [2]>([0x1.a4p-6, 0x1.a4p-6])];
            tensor<fp16, [128, 2]> _inversed_696_cast_fp16 = mul(x = var_693_cast_fp16, y = _inversed_696_y_0_to_fp16)[name = string("_inversed_696_cast_fp16")];
            fp16 var_697_to_fp16 = const()[name = string("op_697_to_fp16"), val = fp16(0x1p+0)];
            tensor<fp16, [128, 2]> keypoints_i_norm_cast_fp16 = sub(x = _inversed_696_cast_fp16, y = var_697_to_fp16)[name = string("keypoints_i_norm_cast_fp16")];
            tensor<int32, [4]> var_699 = const()[name = string("op_699"), val = tensor<int32, [4]>([1, 1, -1, 2])];
            tensor<fp16, [1, 1, 128, 2]> grid_cast_fp16 = reshape(shape = var_699, x = keypoints_i_norm_cast_fp16)[name = string("grid_cast_fp16")];
            string var_701_sampling_mode_0 = const()[name = string("op_701_sampling_mode_0"), val = string("bilinear")];
            string var_701_padding_mode_0 = const()[name = string("op_701_padding_mode_0"), val = string("border")];
            string var_701_coordinates_mode_0 = const()[name = string("op_701_coordinates_mode_0"), val = string("normalized_minus_one_to_one")];
            bool var_701_align_corners_0 = const()[name = string("op_701_align_corners_0"), val = bool(true)];
            fp16 var_701_padding_value_0_to_fp16 = const()[name = string("op_701_padding_value_0_to_fp16"), val = fp16(0x0p+0)];
            tensor<fp16, [1, 42, 1, 128]> var_701_cast_fp16 = resample(align_corners = var_701_align_corners_0, coordinates = grid_cast_fp16, coordinates_format = string("xy"), coordinates_mode = var_701_coordinates_mode_0, coordinates_type = string("absolute"), padding_mode = var_701_padding_mode_0, padding_value = var_701_padding_value_0_to_fp16, sampling_mode = var_701_sampling_mode_0, x = input_131_cast_fp16)[name = string("op_701_cast_fp16")];
            tensor<int32, [4]> var_702_begin_0 = const()[name = string("op_702_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_702_end_0 = const()[name = string("op_702_end_0"), val = tensor<int32, [4]>([1, 42, 1, 128])];
            tensor<bool, [4]> var_702_end_mask_0 = const()[name = string("op_702_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_702_squeeze_mask_0 = const()[name = string("op_702_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [42, 1, 128]> var_702_cast_fp16 = slice_by_index(begin = var_702_begin_0, end = var_702_end_0, end_mask = var_702_end_mask_0, squeeze_mask = var_702_squeeze_mask_0, x = var_701_cast_fp16)[name = string("op_702_cast_fp16")];
            tensor<int32, [3]> var_704_begin_0 = const()[name = string("op_704_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_704_end_0 = const()[name = string("op_704_end_0"), val = tensor<int32, [3]>([42, 1, 128])];
            tensor<bool, [3]> var_704_end_mask_0 = const()[name = string("op_704_end_mask_0"), val = tensor<bool, [3]>([true, false, true])];
            tensor<bool, [3]> var_704_squeeze_mask_0 = const()[name = string("op_704_squeeze_mask_0"), val = tensor<bool, [3]>([false, true, false])];
            tensor<fp16, [42, 128]> var_704_cast_fp16 = slice_by_index(begin = var_704_begin_0, end = var_704_end_0, end_mask = var_704_end_mask_0, squeeze_mask = var_704_squeeze_mask_0, x = var_702_cast_fp16)[name = string("op_704_cast_fp16")];
            tensor<int32, [1]> var_716_axes_0 = const()[name = string("op_716_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 44, 128]> var_716_cast_fp16 = expand_dims(axes = var_716_axes_0, x = var_648_cast_fp16)[name = string("op_716_cast_fp16")];
            tensor<int32, [1]> var_718_axes_0 = const()[name = string("op_718_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 44, 128, 1]> var_718_cast_fp16 = expand_dims(axes = var_718_axes_0, x = var_716_cast_fp16)[name = string("op_718_cast_fp16")];
            tensor<int32, [1]> var_723_axes_0 = const()[name = string("op_723_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 42, 128]> var_723_cast_fp16 = expand_dims(axes = var_723_axes_0, x = var_676_cast_fp16)[name = string("op_723_cast_fp16")];
            tensor<int32, [1]> var_725_axes_0 = const()[name = string("op_725_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 42, 128, 1]> var_725_cast_fp16 = expand_dims(axes = var_725_axes_0, x = var_723_cast_fp16)[name = string("op_725_cast_fp16")];
            tensor<int32, [1]> var_730_axes_0 = const()[name = string("op_730_axes_0"), val = tensor<int32, [1]>([0])];
            tensor<fp16, [1, 42, 128]> var_730_cast_fp16 = expand_dims(axes = var_730_axes_0, x = var_704_cast_fp16)[name = string("op_730_cast_fp16")];
            tensor<int32, [1]> var_732_axes_0 = const()[name = string("op_732_axes_0"), val = tensor<int32, [1]>([3])];
            tensor<fp16, [1, 42, 128, 1]> var_732_cast_fp16 = expand_dims(axes = var_732_axes_0, x = var_730_cast_fp16)[name = string("op_732_cast_fp16")];
            int32 var_734 = const()[name = string("op_734"), val = int32(1)];
            bool input_133_interleave_0 = const()[name = string("input_133_interleave_0"), val = bool(false)];
            tensor<fp16, [1, 128, 128, 1]> input_133_cast_fp16 = concat(axis = var_734, interleave = input_133_interleave_0, values = (var_718_cast_fp16, var_725_cast_fp16, var_732_cast_fp16))[name = string("input_133_cast_fp16")];
            string var_746_pad_type_0 = const()[name = string("op_746_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> var_746_strides_0 = const()[name = string("op_746_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_746_pad_0 = const()[name = string("op_746_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_746_dilations_0 = const()[name = string("op_746_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 var_746_groups_0 = const()[name = string("op_746_groups_0"), val = int32(1)];
            tensor<fp16, [129, 128, 1, 1]> densefeat_df_decoder_convhead2_weight_to_fp16 = const()[name = string("densefeat_df_decoder_convhead2_weight_to_fp16"), val = tensor<fp16, [129, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(424192)))];
            tensor<fp16, [129]> densefeat_df_decoder_convhead2_bias_to_fp16 = const()[name = string("densefeat_df_decoder_convhead2_bias_to_fp16"), val = tensor<fp16, [129]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457280)))];
            tensor<fp16, [1, 129, 128, 1]> var_746_cast_fp16 = conv(bias = densefeat_df_decoder_convhead2_bias_to_fp16, dilations = var_746_dilations_0, groups = var_746_groups_0, pad = var_746_pad_0, pad_type = var_746_pad_type_0, strides = var_746_strides_0, weight = densefeat_df_decoder_convhead2_weight_to_fp16, x = input_133_cast_fp16)[name = string("op_746_cast_fp16")];
            tensor<int32, [4]> var_756_begin_0 = const()[name = string("op_756_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_756_end_0 = const()[name = string("op_756_end_0"), val = tensor<int32, [4]>([1, 128, 128, 1])];
            tensor<bool, [4]> var_756_end_mask_0 = const()[name = string("op_756_end_mask_0"), val = tensor<bool, [4]>([true, false, true, true])];
            tensor<fp16, [1, 128, 128, 1]> var_756_cast_fp16 = slice_by_index(begin = var_756_begin_0, end = var_756_end_0, end_mask = var_756_end_mask_0, x = var_746_cast_fp16)[name = string("op_756_cast_fp16")];
            tensor<int32, [1]> var_769 = const()[name = string("op_769"), val = tensor<int32, [1]>([1])];
            bool var_770 = const()[name = string("op_770"), val = bool(true)];
            tensor<fp16, [1, 1, 128, 1]> var_772_cast_fp16 = reduce_l2_norm(axes = var_769, keep_dims = var_770, x = var_756_cast_fp16)[name = string("op_772_cast_fp16")];
            fp16 var_773_to_fp16 = const()[name = string("op_773_to_fp16"), val = fp16(0x1p-24)];
            tensor<fp16, [1, 1, 128, 1]> var_774_cast_fp16 = maximum(x = var_772_cast_fp16, y = var_773_to_fp16)[name = string("op_774_cast_fp16")];
            tensor<int32, [4]> denom_reps_0 = const()[name = string("denom_reps_0"), val = tensor<int32, [4]>([1, 128, 1, 1])];
            tensor<fp16, [1, 128, 128, 1]> denom_cast_fp16 = tile(reps = denom_reps_0, x = var_774_cast_fp16)[name = string("denom_cast_fp16")];
            tensor<fp16, [1, 128, 128, 1]> descriptor_i_cast_fp16 = real_div(x = var_756_cast_fp16, y = denom_cast_fp16)[name = string("descriptor_i_cast_fp16")];
            tensor<int32, [4]> var_779_begin_0 = const()[name = string("op_779_begin_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [4]> var_779_end_0 = const()[name = string("op_779_end_0"), val = tensor<int32, [4]>([1, 128, 128, 1])];
            tensor<bool, [4]> var_779_end_mask_0 = const()[name = string("op_779_end_mask_0"), val = tensor<bool, [4]>([false, true, true, true])];
            tensor<bool, [4]> var_779_squeeze_mask_0 = const()[name = string("op_779_squeeze_mask_0"), val = tensor<bool, [4]>([true, false, false, false])];
            tensor<fp16, [128, 128, 1]> var_779_cast_fp16 = slice_by_index(begin = var_779_begin_0, end = var_779_end_0, end_mask = var_779_end_mask_0, squeeze_mask = var_779_squeeze_mask_0, x = descriptor_i_cast_fp16)[name = string("op_779_cast_fp16")];
            tensor<int32, [3]> var_782_begin_0 = const()[name = string("op_782_begin_0"), val = tensor<int32, [3]>([0, 0, 0])];
            tensor<int32, [3]> var_782_end_0 = const()[name = string("op_782_end_0"), val = tensor<int32, [3]>([128, 128, 1])];
            tensor<bool, [3]> var_782_end_mask_0 = const()[name = string("op_782_end_mask_0"), val = tensor<bool, [3]>([true, true, false])];
            tensor<bool, [3]> var_782_squeeze_mask_0 = const()[name = string("op_782_squeeze_mask_0"), val = tensor<bool, [3]>([false, false, true])];
            tensor<fp16, [128, 128]> var_782_cast_fp16 = slice_by_index(begin = var_782_begin_0, end = var_782_end_0, end_mask = var_782_end_mask_0, squeeze_mask = var_782_squeeze_mask_0, x = var_779_cast_fp16)[name = string("op_782_cast_fp16")];
            tensor<int32, [2]> x_3_perm_0 = const()[name = string("x_3_perm_0"), val = tensor<int32, [2]>([1, 0])];
            tensor<int32, [1]> _x_3_axes_0 = const()[name = string("_x_3_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [128, 128]> x_3_cast_fp16 = transpose(perm = x_3_perm_0, x = var_782_cast_fp16)[name = string("transpose_1")];
            tensor<fp16, [128, 128, 1]> _x_3_cast_fp16 = expand_dims(axes = _x_3_axes_0, x = x_3_cast_fp16)[name = string("_x_3_cast_fp16")];
            tensor<int32, [1]> _x_5_axes_0 = const()[name = string("_x_5_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<fp16, [128, 128, 1, 1]> _x_5_cast_fp16 = expand_dims(axes = _x_5_axes_0, x = _x_3_cast_fp16)[name = string("_x_5_cast_fp16")];
            string _x_7_pad_type_0 = const()[name = string("_x_7_pad_type_0"), val = string("valid")];
            tensor<int32, [2]> _x_7_strides_0 = const()[name = string("_x_7_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> _x_7_pad_0 = const()[name = string("_x_7_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> _x_7_dilations_0 = const()[name = string("_x_7_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            int32 _x_7_groups_0 = const()[name = string("_x_7_groups_0"), val = int32(1)];
            tensor<fp16, [512, 128, 1, 1]> lsh_binarization_conv_weight_to_fp16 = const()[name = string("lsh_binarization_conv_weight_to_fp16"), val = tensor<fp16, [512, 128, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(457664)))];
            tensor<fp16, [128, 512, 1, 1]> _x_7_cast_fp16 = conv(dilations = _x_7_dilations_0, groups = _x_7_groups_0, pad = _x_7_pad_0, pad_type = _x_7_pad_type_0, strides = _x_7_strides_0, weight = lsh_binarization_conv_weight_to_fp16, x = _x_5_cast_fp16)[name = string("_x_7_cast_fp16")];
            tensor<fp16, [128, 512, 1, 1]> var_805_cast_fp16 = relu(x = _x_7_cast_fp16)[name = string("op_805_cast_fp16")];
            tensor<fp16, [128, 512, 1, 1]> x_cast_fp16 = sign(x = var_805_cast_fp16)[name = string("x_cast_fp16")];
            string var_811_pad_type_0 = const()[name = string("op_811_pad_type_0"), val = string("valid")];
            int32 var_811_groups_0 = const()[name = string("op_811_groups_0"), val = int32(64)];
            tensor<int32, [2]> var_811_strides_0 = const()[name = string("op_811_strides_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<int32, [4]> var_811_pad_0 = const()[name = string("op_811_pad_0"), val = tensor<int32, [4]>([0, 0, 0, 0])];
            tensor<int32, [2]> var_811_dilations_0 = const()[name = string("op_811_dilations_0"), val = tensor<int32, [2]>([1, 1])];
            tensor<fp16, [64, 8, 1, 1]> lsh_binarization_bitpacker_kernel_to_fp16 = const()[name = string("lsh_binarization_bitpacker_kernel_to_fp16"), val = tensor<fp16, [64, 8, 1, 1]>(BLOBFILE(path = string("@model_path/weights/weight.bin"), offset = uint64(588800)))];
            tensor<fp16, [128, 64, 1, 1]> var_811_cast_fp16 = conv(dilations = var_811_dilations_0, groups = var_811_groups_0, pad = var_811_pad_0, pad_type = var_811_pad_type_0, strides = var_811_strides_0, weight = lsh_binarization_bitpacker_kernel_to_fp16, x = x_cast_fp16)[name = string("op_811_cast_fp16")];
            string _x_9_dtype_0 = const()[name = string("_x_9_dtype_0"), val = string("uint8")];
            tensor<int32, [1]> _x_11_axes_0 = const()[name = string("_x_11_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<uint8, [128, 64, 1, 1]> var_811_cast_fp16_to_uint8 = cast(dtype = _x_9_dtype_0, x = var_811_cast_fp16)[name = string("cast_40")];
            tensor<uint8, [128, 64, 1]> _x_11 = squeeze(axes = _x_11_axes_0, x = var_811_cast_fp16_to_uint8)[name = string("_x_11")];
            tensor<int32, [1]> _x_axes_0 = const()[name = string("_x_axes_0"), val = tensor<int32, [1]>([-1])];
            tensor<uint8, [128, 64]> _x = squeeze(axes = _x_axes_0, x = _x_11)[name = string("_x")];
            tensor<int32, [4]> var_820 = const()[name = string("op_820"), val = tensor<int32, [4]>([1, 1, 128, 2])];
            tensor<fp16, [1, 1, 128, 2]> keypoints_reshape = reshape(shape = var_820, x = keypoints_17_cast_fp16)[name = string("op_821_cast_fp16")];
            tensor<int32, [4]> var_826 = const()[name = string("op_826"), val = tensor<int32, [4]>([128, 1, 1, 64])];
            tensor<uint8, [128, 1, 1, 64]> descriptors_reshape = reshape(shape = var_826, x = _x)[name = string("op_827")];
            tensor_buffer<fp16, shape=[1, 1, 128, 2], strides=[8192, 8192, 64, 1], interleave_factors=[1,1,1, 1]> keypoints = tensor_to_tensor_buffer(input = keypoints_reshape, interleave_factors = tensor<uint8, [4]>([1,1, 1, 1]), strides = tensor<int64, [4]>([8192, 8192, 64, 1]))[name = string("keypoints_output")];
            tensor_buffer<uint8, shape=[128, 1, 1, 64], strides=[64, 64, 64, 1], interleave_factors=[1,1,1, 1]> descriptors = tensor_to_tensor_buffer(input = descriptors_reshape, interleave_factors = tensor<uint8, [4]>([1, 1, 1, 1]), strides = tensor<int64, [4]>([64, 64, 64, 1]))[name = string("descriptors_output")];
        } -> (keypoints, descriptors);
}