onnx和pytorch模型转换rknn模型，输出结果shape不一样

如题，部署设备是rk1808，问题出现的测试demo:

import torch
import torch.nn as nn

class Tmodel(nn.Module):
    def __init__(self, length=28):
        super().__init__()
        self.values = [self._norm(d) for d in [length, length]]

    def _norm(self, length):
        first = -(length - 1.0) / length
        values = torch.arange(length, device='cpu') * (2.0 / length) + first
        return values

    def forward(self, probs):
        assert(len(self.values) == probs.ndimension() - 2)
        expectation = []
        marg1 = probs.sum(3, keepdim=False)
        marg2 = probs.sum(2, keepdim=False)
        marg1 = self.values[0] * marg1
        marg2 = self.values[1] * marg2
        expectation.append(marg1.sum(marg1.ndim-1, keepdim=False))
        expectation.append(marg2.sum(marg2.ndim-1, keepdim=False))
        return torch.stack(expectation, -1)

net = Tmodel()
net.eval()
x = torch.rand(1, 1, 28, 28)
torch.onnx.export(net,
                      x,
                      './tmodel.onnx',
                      verbose=False,
                      opset_version=11,
                      do_constant_folding=True,
                      input_names=['input'],
                      output_names=['output'])
print('export success.')

在转换到rknn时，出现如下错误： 查看转换成的rknn模型，发现输出的shape与onnx不相同， 转换日志如下： D Using CPPUTILS: True I Start importing onnx... I Current ONNX Model use ir_version 6 opset_version 11 I Call RKNN onnx optimize fail, skip optimize D Calc tensor ReduceSum_2 (1, 1, 28) D Calc tensor ReduceSum_1 (1, 1, 28) D Calc tensor Initializer_5 (28,) D Calc tensor Mul_6 (1, 1, 28) D Calc tensor ReduceSum_8 (1, 1) D Calc tensor Unsqueeze_10 (1, 1, 1) D Calc tensor Initializer_3 (28,) D Calc tensor Mul_4 (1, 1, 28) D Calc tensor ReduceSum_7 (1, 1) D Calc tensor Unsqueeze_9 (1, 1, 1) D Calc tensor Concat_output (1, 1, 2) D import clients finished I build output layer attach_Concat_Concat_10:out0 I Try match layer: Concat_Concat_10:out0 I Match concat_2, convert ONNX layer: ['Concat_Concat_10'](OPs: ['Concat']) to ['concat'] I Try match layer: Unsqueeze_Unsqueeze_8:out0 I Match r_unsqueeze, convert ONNX layer: ['Unsqueeze_Unsqueeze_8'](OPs: ['Unsqueeze']) to ['reshape'] I Try match layer: Unsqueeze_Unsqueeze_9:out0 I Match r_unsqueeze, convert ONNX layer: ['Unsqueeze_Unsqueeze_9'](OPs: ['Unsqueeze']) to ['reshape'] I Try match layer: ReduceSum_ReduceSum_6:out0 I Match r_reducesum, convert ONNX layer: ['ReduceSum_ReduceSum_6'](OPs: ['ReduceSum']) to ['reducesum'] I Try match layer: ReduceSum_ReduceSum_7:out0 I Match r_reducesum, convert ONNX layer: ['ReduceSum_ReduceSum_7'](OPs: ['ReduceSum']) to ['reducesum'] I Try match layer: Mul_Mul_3:out0 I Match r_mul, convert ONNX layer: ['Mul_Mul_3'](OPs: ['Mul']) to ['multiply'] I Try match layer: Mul_Mul_5:out0 I Match r_mul, convert ONNX layer: ['Mul_Mul_5'](OPs: ['Mul']) to ['multiply'] I Try match layer: Initializer_3:out0 I Match r_variable, convert ONNX layer: ['Initializer_3'](OPs: ['Constant']) to ['variable'] I Try match layer: ReduceSum_ReduceSum_0:out0 I Match r_reducesum, convert ONNX layer: ['ReduceSum_ReduceSum_0'](OPs: ['ReduceSum']) to ['reducesum'] I Try match layer: Initializer_5:out0 I Match r_variable, convert ONNX layer: ['Initializer_5'](OPs: ['Constant']) to ['variable'] I Try match layer: ReduceSum_ReduceSum_1:out0 I Match r_reducesum, convert ONNX layer: ['ReduceSum_ReduceSum_1'](OPs: ['ReduceSum']) to ['reducesum'] W Input channel 1 less then 3, please update NPU Driver to v1.3.0 or later. I build input layer input:out0 D connect Unsqueeze_Unsqueeze_8_2 0 ~ Concat_Concat_10_1 0 D connect Unsqueeze_Unsqueeze_9_3 0 ~ Concat_Concat_10_1 1 D connect ReduceSum_ReduceSum_6_4 0 ~ Unsqueeze_Unsqueeze_8_2 0 D connect ReduceSum_ReduceSum_7_5 0 ~ Unsqueeze_Unsqueeze_9_3 0 D connect Mul_Mul_3_6 0 ~ ReduceSum_ReduceSum_6_4 0 D connect Mul_Mul_5_7 0 ~ ReduceSum_ReduceSum_7_5 0 D connect Initializer_3_8 0 ~ Mul_Mul_3_6 0 D connect ReduceSum_ReduceSum_0_9 0 ~ Mul_Mul_3_6 1 D connect Initializer_5_10 0 ~ Mul_Mul_5_7 0 D connect ReduceSum_ReduceSum_1_11 0 ~ Mul_Mul_5_7 1 D connect input_12 0 ~ ReduceSum_ReduceSum_0_9 0 D connect input_12 0 ~ ReduceSum_ReduceSum_1_11 0 D connect Concat_Concat_10_1 0 ~ attach_Concat_Concat_10/out0_0 0 D Process Initializer_3_8 ... D RKNN output shape(variable): (28) D Process input_12 ... D RKNN output shape(input): (1 28 28 1) D Tensor @input_12:out0 type: float32 D Process ReduceSum_ReduceSum_0_9 ... D RKNN output shape(reducesum): (1 1 28) D Tensor @ReduceSum_ReduceSum_0_9:out0 type: float32 D Process Mul_Mul_3_6 ... D RKNN output shape(multiply): (1 1 28) D Tensor @Mul_Mul_3_6:out0 type: float32 D Process ReduceSum_ReduceSum_6_4 ... W Warning: Axis may need to be adjusted according to original model shape. D RKNN output shape(reducesum): (1 1) D Tensor @ReduceSum_ReduceSum_6_4:out0 type: float32 D Process Unsqueeze_Unsqueeze_8_2 ... D RKNN output shape(reshape): (1 1) D Tensor @Unsqueeze_Unsqueeze_8_2:out0 type: float32 D Process Initializer_5_10 ... D RKNN output shape(variable): (28) D Process ReduceSum_ReduceSum_1_11 ... D RKNN output shape(reducesum): (1 1 28) D Tensor @ReduceSum_ReduceSum_1_11:out0 type: float32 D Process Mul_Mul_5_7 ... D RKNN output shape(multiply): (1 1 28) D Tensor @Mul_Mul_5_7:out0 type: float32 D Process ReduceSum_ReduceSum_7_5 ... W Warning: Axis may need to be adjusted according to original model shape. D RKNN output shape(reducesum): (1 1) D Tensor @ReduceSum_ReduceSum_7_5:out0 type: float32 D Process Unsqueeze_Unsqueeze_9_3 ... D RKNN output shape(reshape): (1 1) D Tensor @Unsqueeze_Unsqueeze_9_3:out0 type: float32 D Process Concat_Concat_10_1 ... W Warning: Axis may need to be adjusted according to original model shape. D RKNN output shape(concat): (1 2) D Tensor @Concat_Concat_10_1:out0 type: float32 D Process attach_Concat_Concat_10/out0_0 ... D RKNN output shape(output): (1 2) D Tensor @attach_Concat_Concat_10/out0_0:out0 type: float32 I Build torch-jit-export complete. I Start C2T Switcher... D Optimizing network with broadcast_op D Align variable Initializer_3_8 shape [28] to [1, 1, 28] D Align variable Initializer_5_10 shape [28] to [1, 1, 28] D convert ReduceSum_ReduceSum_0_9(reducesum) axes [3] to [2] D convert ReduceSum_ReduceSum_1_11(reducesum) axes [2] to [1] D convert Mul_Mul_3_6(multiply) axis 1 to 2 D convert Mul_Mul_5_7(multiply) axis 1 to 2 D insert permute ReduceSum_ReduceSum_6_4_RKNN_mark_perm_13 before ReduceSum_ReduceSum_6_4 D insert permute ReduceSum_ReduceSum_7_5_RKNN_mark_perm_14 before ReduceSum_ReduceSum_7_5 I End C2T Switcher... D Optimizing network with force_1d_tensor, swapper, merge_duplicate_quantize_dequantize, merge_layer, auto_fill_bn, auto_fill_l2normalizescale, auto_fill_instancenormalize, resize_nearest_transformer, eltwise_transform, auto_fill_multiply, auto_fill_zero_bias, proposal_opt_import, special_add_to_conv2d I End importing onnx... W The target_platform is not set in config, using default target platform rk1808. I Generate input meta ... I Load input meta I Generate input meta ... D import clients finished I Load net... I Load data... I Load input meta D Process Initializer_3_8 ... D RKNN output shape(variable): (1 28 1) D Process input_12 ... D RKNN output shape(input): (1 28 28 1) D Tensor @input_12:out0 type: float32 D Process ReduceSum_ReduceSum_0_9 ... D RKNN output shape(reducesum): (1 1 28) D Tensor @ReduceSum_ReduceSum_0_9:out0 type: float32 D Process Mul_Mul_3_6 ... D RKNN output shape(multiply): (1 28 28) D Tensor @Mul_Mul_3_6:out0 type: float32 D Process ReduceSum_ReduceSum_6_4_RKNN_mark_perm_13 ... D RKNN output shape(permute): (1 28 28) D Tensor @ReduceSum_ReduceSum_6_4_RKNN_mark_perm_13:out0 type: float32 D Process ReduceSum_ReduceSum_6_4 ... D RKNN output shape(reducesum): (1 28) D Tensor @ReduceSum_ReduceSum_6_4:out0 type: float32 D Process Unsqueeze_Unsqueeze_8_2 ... D RKNN output shape(reshape): (28 1) D Tensor @Unsqueeze_Unsqueeze_8_2:out0 type: float32 D Process Initializer_5_10 ... D RKNN output shape(variable): (1 28 1) D Process ReduceSum_ReduceSum_1_11 ... D RKNN output shape(reducesum): (1 1 28) D Tensor @ReduceSum_ReduceSum_1_11:out0 type: float32 D Process Mul_Mul_5_7 ... D RKNN output shape(multiply): (1 28 28) D Tensor @Mul_Mul_5_7:out0 type: float32 D Process ReduceSum_ReduceSum_7_5_RKNN_mark_perm_14 ... D RKNN output shape(permute): (1 28 28) D Tensor @ReduceSum_ReduceSum_7_5_RKNN_mark_perm_14:out0 type: float32 D Process ReduceSum_ReduceSum_7_5 ... D RKNN output shape(reducesum): (1 28) D Tensor @ReduceSum_ReduceSum_7_5:out0 type: float32 D Process Unsqueeze_Unsqueeze_9_3 ... D RKNN output shape(reshape): (28 1) D Tensor @Unsqueeze_Unsqueeze_9_3:out0 type: float32 D Process Concat_Concat_10_1 ... D RKNN output shape(concat): (28 2) D Tensor @Concat_Concat_10_1:out0 type: float32 D Process attach_Concat_Concat_10/out0_0 ... D RKNN output shape(output): (28 2) D Tensor @attach_Concat_Concat_10/out0_0:out0 type: float32 I Build torch-jit-export complete. I Initialzing network optimizer by /root/anaconda3/envs/torch/lib/python3.6/site-packages/rknn/config/npu_config/LION ... D Optimizing network with merge_ximum, qnt_adjust_coef, multiply_transform, add_extra_io, format_input_ops, auto_fill_zero_bias, conv_kernel_transform, twod_op_transform, strip_op, extend_unstack_split, merge_layer, transform_layer, broadcast_op, strip_op, auto_fill_reshape_zero, adjust_output_attrs, insert_dtype_converter I Start T2C Switcher... D Optimizing network with broadcast_op, t2c_fc D convert ReduceSum_ReduceSum_0_9(reducesum) axes [2] to [3] D convert ReduceSum_ReduceSum_1_11(reducesum) axes [1] to [2] D convert Mul_Mul_3_6(multiply) axis 2 to 1 D convert Mul_Mul_5_7(multiply) axis 2 to 1 D insert permute ReduceSum_ReduceSum_6_4_RKNN_mark_perm_13_RKNN_mark_perm_15 before ReduceSum_ReduceSum_6_4_RKNN_mark_perm_13 D insert permute ReduceSum_ReduceSum_7_5_RKNN_mark_perm_14_RKNN_mark_perm_16 before ReduceSum_ReduceSum_7_5_RKNN_mark_perm_14 D remove permute ReduceSum_ReduceSum_6_4_RKNN_mark_perm_13_RKNN_mark_perm_15 D remove permute ReduceSum_ReduceSum_6_4_RKNN_mark_perm_13 D remove permute ReduceSum_ReduceSum_7_5_RKNN_mark_perm_14_RKNN_mark_perm_16 D remove permute ReduceSum_ReduceSum_7_5_RKNN_mark_perm_14 I End T2C Switcher... D Optimizing network with merge_ximum, qnt_adjust_coef, multiply_transform, add_extra_io, format_input_ops, auto_fill_zero_bias, conv_kernel_transform, twod_op_transform, strip_op, extend_unstack_split, merge_layer, transform_layer, broadcast_op, strip_op, auto_fill_reshape_zero, adjust_output_attrs, insert_dtype_converter D Optimizing network with conv_1xn_transform, proposal_opt, c2drv_convert_axis, c2drv_convert_shape, c2drv_convert_array, c2drv_cast_dtype, c2drv_trans_data I Building data ... I Packing data ... D Packing Initializer_3_8 ... D Packing Initializer_5_10 ... D output tensor id = 0, name = Concat_Concat_10/out0_0 D input tensor id = 1, name = input_12 I Build config finished. D target set by user is: None D Starting ntp or adb, target is None, host is None I When we run model on simulator, we do not need set perf_debug to get layer performance. E Catch exception when init runtime! E Traceback (most recent call last): E File "rknn/api/rknn_base.py", line 897, in rknn.api.rknn_base.RKNNBase.init_runtime E File "rknn/api/rknn_runtime.py", line 341, in rknn.api.rknn_runtime.RKNNRuntime.build_graph E Exception: RKNN init failed. error code: RKNN_ERR_MODEL_INVALID E Please feedback the detailed log file <log_feedback_to_the_rknn_toolkit_dev_team.log> to the RKNN Toolkit development team. E You can also check github issues: https://github.com/rockchip-linux/rknn-toolkit/issues