InvalidArgumentError: Default MaxPoolingOp only supports NHWC on device type CPU. Why this happen after onnx2keras?

Hi again,

I have done this steps:

onnx_model = onnx.load(FILE_PATH+"mnist_test.onnx")
k_model_onnx = onnx_to_keras(onnx_model, ['input_1'], name_policy="short")
k_model_onnx.summary()

Model: "model"
__________________________________________________________________________________________________
Layer (type)                    Output Shape         Param #     Connected to                     
==================================================================================================
input_1 (InputLayer)            [(None, 28, 28, 1)]  0                                            
__________________________________________________________________________________________________
adjusted (Permute)              (None, 1, 28, 28)    0           input_1[0][0]                    
__________________________________________________________________________________________________
convolut (Conv2D)               (None, 32, 26, 26)   320         adjusted[0][0]                   
__________________________________________________________________________________________________
conv2d/I (Activation)           (None, 32, 26, 26)   0           convolut[0][0]                   
__________________________________________________________________________________________________
convolut_1 (Conv2D)             (None, 64, 24, 24)   18496       conv2d/I[0][0]                   
__________________________________________________________________________________________________
conv2d_1 (Activation)           (None, 64, 24, 24)   0           convolut_1[0][0]                 
__________________________________________________________________________________________________
conv2d_1_1_pad (ZeroPadding2D)  (None, 64, 24, 24)   0           conv2d_1[0][0]                   
__________________________________________________________________________________________________
conv2d_1_1 (MaxPooling2D)       (None, 64, 12, 12)   0           conv2d_1_1_pad[0][0]             
__________________________________________________________________________________________________
conv2d_1_2 (Permute)            (None, 12, 12, 64)   0           conv2d_1_1[0][0]                 
__________________________________________________________________________________________________
flatten/ (Reshape)              (None, None)         0           conv2d_1_2[0][0]                 
__________________________________________________________________________________________________
transfor_reshape (Reshape)      (None, 9216)         0           flatten/[0][0]                   
__________________________________________________________________________________________________
transfor (Dense)                (None, 128)          1179648     transfor_reshape[0][0]           
__________________________________________________________________________________________________
biased_t_const2 (Lambda)        (128,)               0           input_1[0][0]                    
__________________________________________________________________________________________________
biased_t (Lambda)               (None, 128)          0           transfor[0][0]                   
                                                                 biased_t_const2[0][0]            
__________________________________________________________________________________________________
dense/Id (Activation)           (None, 128)          0           biased_t[0][0]                   
__________________________________________________________________________________________________
transfor_1 (Dense)              (None, 10)           1280        dense/Id[0][0]                   
__________________________________________________________________________________________________
biased_t_1_const2 (Lambda)      (10,)                0           input_1[0][0]                    
__________________________________________________________________________________________________
biased_t_1 (Lambda)             (None, 10)           0           transfor_1[0][0]                 
                                                                 biased_t_1_const2[0][0]          
__________________________________________________________________________________________________
dense_1/ (Activation)           (None, 10)           0           biased_t_1[0][0]                 
==================================================================================================
Total params: 1,199,744
Trainable params: 1,199,744
Non-trainable params: 0
__________________________________________________________________________________________________

y_pred_onnx = k_model_onnx.predict(x_test)

Result :

Tensor("model/transfor/MatMul:0", shape=(None, 128), dtype=float32) Tensor("model/biased_t_const2/Const:0", shape=(128,), dtype=float32)
Tensor("model/transfor_1/MatMul:0", shape=(None, 10), dtype=float32) Tensor("model/biased_t_1_const2/Const:0", shape=(10,), dtype=float32)
---------------------------------------------------------------------------
InvalidArgumentError                      Traceback (most recent call last)
<ipython-input-16-c199c87d14d2> in <module>()
----> 1 y_pred_onnx = k_model_onnx.predict(x_test)

7 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/eager/execute.py in quick_execute(op_name, num_outputs, inputs, attrs, ctx, name)
     58     ctx.ensure_initialized()
     59     tensors = pywrap_tfe.TFE_Py_Execute(ctx._handle, device_name, op_name,
---> 60                                         inputs, attrs, num_outputs)
     61   except core._NotOkStatusException as e:
     62     if name is not None:

InvalidArgumentError:  Default MaxPoolingOp only supports NHWC on device type CPU
	 [[node model/conv2d_1_1/MaxPool (defined at <ipython-input-16-c199c87d14d2>:1) ]] [Op:__inference_predict_function_2104]

Function call stack:
predict_function

I have no idea what this mean and why it has happen. Thanks and sorry!

[M-Tonin]

But after this, the shape doesn’t match. I have create the model in keras after I have saved model.h5 and model.onnx. In another project I have load the model.h5 and predict x_test and thats ok and now I am trying to load model.onnx and convert to keras model (change_ordering=True, name_policy=“short”) and when I try to predict using this model I have shape does not match.

WARNING:tensorflow:Model was constructed with shape (None, 28, 1, 28) for input Tensor("input_1_55:0", shape=(None, 28, 1, 28), dtype=float32), but it was called on an input with incompatible shape (None, 28, 28).
WARNING:tensorflow:Model was constructed with shape (None, 28, 1, 28) for input Tensor("input_1_55:0", shape=(None, 28, 1, 28), dtype=float32), but it was called on an input with incompatible shape (None, 28, 28).
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
<ipython-input-72-b08c625d5e00> in <module>()
----> 1 y_predas = k_model_onnx.predict(x_test)

10 frames
/usr/local/lib/python3.6/dist-packages/tensorflow/python/framework/func_graph.py in wrapper(*args, **kwargs)
    966           except Exception as e:  # pylint:disable=broad-except
    967             if hasattr(e, "ag_error_metadata"):
--> 968               raise e.ag_error_metadata.to_exception(e)
    969             else:
    970               raise

ValueError: in user code:

    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:1150 predict_function  *
        outputs = self.distribute_strategy.run(
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:951 run  **
        return self._extended.call_for_each_replica(fn, args=args, kwargs=kwargs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:2290 call_for_each_replica
        return self._call_for_each_replica(fn, args, kwargs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/distribute/distribute_lib.py:2649 _call_for_each_replica
        return fn(*args, **kwargs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/training.py:1125 predict_step  **
        return self(x, training=False)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/base_layer.py:927 __call__
        outputs = call_fn(cast_inputs, *args, **kwargs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/network.py:719 call
        convert_kwargs_to_constants=base_layer_utils.call_context().saving)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/network.py:888 _run_internal_graph
        output_tensors = layer(computed_tensors, **kwargs)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/base_layer.py:886 __call__
        self.name)
    /usr/local/lib/python3.6/dist-packages/tensorflow/python/keras/engine/input_spec.py:180 assert_input_compatibility
        str(x.shape.as_list()))

    ValueError: Input 0 of layer adjusted is incompatible with the layer: expected ndim=4, found ndim=3. Full shape received: [None, 28, 28]

I tryed to use the argument input_shape=, but not sucessfull

[gmalivenko]

Hello @M-Tonin.

InvalidArgumentError:  Default MaxPoolingOp only supports NHWC on device type CPU

There are 2 options to fix:

1. You can try save your model and then load using CPU device
2. You can try to pass change_ordering=True argument to the converter call (it’s experimental and may fail)