Keras TimeDistributed Layer seems not supported

Hi, I am using a TimeDistributed Layer in model which has an input shape of {10, 150, 180, 3} (10 frames of (w:150, h:180, colors : 3 ) ).

Since my device support a maximum of 4 dimensions, I was hoping that it should work, but after converting the Model from Keras to ONNX and then trying to convert to RKNN, I am getting this error :

E build:   File "/home/mac/miniconda3/envs/ai/lib/python3.11/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 462, in _create_inference_session
E build:     sess = C.InferenceSession(session_options, self._model_bytes, False, self._read_config_from_model)
E build:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
E build: onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Node (sequential_21_1/time_distributed_20_1/transpose) Op (Transpose) [TypeInferenceError] Invalid attribute perm {1, 0, 2, 3, 4}, input shape = {10, 150, 180, 3}

Checking the ONNX model show indeed a first transpose layer with those attributs {1, 0, 2, 3, 4} it seems it comming from this Keras line code Keras TimeDistribued Layer :

        def time_distributed_transpose(data):
            """Swaps the timestep and batch dimensions of a tensor."""
            axes = [1, 0, *range(2, len(data.shape))]
            return ops.transpose(data, axes=axes)

So is there any hope that I could run a TimeDistribued Layer ( which by the way take an arrays of frames as DataSet and not images ) ?

Is it supported by the RKNN or not ? (The Keras TimeDistributed Layer which is converted using Transpose Layer in the ONNX)

Is there a workaround ?

Having change the opset to 18 while exporting to ONNX fix the problem nut now another issue is reported :

--> Loading model
W __init__: rknn-toolkit2 version: 1.6.0+81f21f4d
W load_onnx: If you don't need to crop the model, don't set 'inputs'/'input_size_list'/'outputs'!
W load_onnx: It is recommended onnx opset 19, but your onnx model opset is 18!
Loading : 100%|████████████████████████████████████████████████| 211/211 [00:00<00:00, 55730.36it/s]
W load_onnx: The config.mean_values is None, zeros will be set for input 0!
W load_onnx: The config.std_values is None, ones will be set for input 0!
done
--> Building model
W build: found outlier value, this may affect quantization accuracy
const name                                                                                                        abs_mean    abs_std     outlier value
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block4a_dwconv2/depthwise_weights_fused_bn      3.54        5.94        41.139      
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block4b_dwconv2/depthwise_weights_fused_bn      1.24        2.33        24.630      
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block4c_dwconv2/depthwise_weights_fused_bn      1.02        1.88        21.949      
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block5a_dwconv2/depthwise_weights_fused_bn      1.31        2.94        -31.833     
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block5b_dwconv2/depthwise_weights_fused_bn      1.28        1.91        -20.992     
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block5c_dwconv2/depthwise_weights_fused_bn      1.10        1.58        16.670      
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block5d_dwconv2/depthwise_weights_fused_bn      1.03        1.50        17.634      
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block5e_dwconv2/depthwise_weights_fused_bn      1.03        1.41        -20.423     
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block6a_dwconv2/depthwise_weights_fused_bn      1.38        1.12        -18.975     
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block6b_dwconv2/depthwise_weights_fused_bn      1.49        1.76        -29.987     
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block6c_dwconv2/depthwise_weights_fused_bn      1.55        1.77        -23.349     
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block6e_dwconv2/depthwise_weights_fused_bn      1.48        1.32        -17.528, 16.979
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block6f_dwconv2/depthwise_weights_fused_bn      1.49        1.23        -24.013     
sequential_23/time_distributed_16/sequential_22/efficientnetv2-b0/block6g_dwconv2/depthwise_weights_fused_bn      1.54        1.11        -15.516     
GraphPreparing : 100%|███████████████████████████████████████████| 257/257 [00:00<00:00, 849.90it/s]
Quantizating :   0%|                                                        | 0/257 [00:00<?, ?it/s]E build: The channel of r_shape must be 3!
W build: ===================== WARN(6) =====================
E rknn-toolkit2 version: 1.6.0+81f21f4d
Quantizating :   0%|                                                        | 0/257 [00:00<?, ?it/s]
E build: Catch exception when building RKNN model!
E build: Traceback (most recent call last):
E build:   File "rknn/api/rknn_base.py", line 1996, in rknn.api.rknn_base.RKNNBase.build
E build:   File "rknn/api/rknn_base.py", line 148, in rknn.api.rknn_base.RKNNBase._quantize
E build:   File "rknn/api/quantizer.py", line 1289, in rknn.api.quantizer.Quantizer.run
E build:   File "rknn/api/quantizer.py", line 821, in rknn.api.quantizer.Quantizer._get_layer_range
E build:   File "rknn/api/rknn_utils.py", line 176, in rknn.api.rknn_utils.get_input_img
E build:   File "rknn/api/rknn_log.py", line 92, in rknn.api.rknn_log.RKNNLog.e
E build: ValueError: The channel of r_shape must be 3!

rockchip-linux / rknn-toolkit2

Keras TimeDistributed Layer seems not supported #340