Tensorflowjs Operands could not be broadcast together with shapes

I have a working model in Tensorflow that works fine with Python. It is now a saved model and attempted to convert to TensorflowJS.

A converted model appears to not work correctly. The mean value from the TensorMap appears to be 0 and has a shape of [0], despite the input being [1,96,192,3].

The conversion was done as follows....

tensorflowjs_converter --input_format=tf_saved_model ~/Projects/models/model_0 ~/Projects/modelsjs/model_0

This works fine and it also appears to load ok. However, when it comes to predicting, errors are thrown and any advice would be appreciated.

<script>
    async function handleButtonClick(){

       for(var i=0; i<1;i++)
       {
           var t0 = performance.now();
           console.log("Loading - model_"+i);
           var inputTensor = tf.tensor4d(input);
           var model = await tf.loadGraphModel('/modelsjs/model_'+i+'/model.json');
           var poutput = model.predict(inputTensor);
       }
</script>

The error appears as follows.

graph_model.ts:213 Uncaught (in promise) Error: The model contains control flow or dynamic shape ops, please use executeAsync method
    at t.execute_ (graph_model.ts:213)
    at t.predict (graph_model.ts:169)
    at (index):157
    at engine.ts:156
    at t.scopedRun (engine.ts:167)
    at t.tidy (engine.ts:153)
    at Object.t.tidy (environment.ts:186)
    at handleButtonClick ((index):156)

As per the above error, the prediction was attempted using executeAsync but produces the error relevant to this question.

<script>
    async function handleButtonClick(){

       for(var i=0; i<1;i++)
       {
           var t0 = performance.now();
           console.log("Loading - model_"+i);
           var inputTensor = tf.tensor4d(input);
           var model = await tf.loadGraphModel('/modelsjs/model_'+i+'/model.json');
           console.log("Load model_" + i + "took " + (t1 - t0) + " milliseconds.");
           const res = await model.executeAsync(inputTensor);
       }
</script>

The Error appears as follows. And appears to be related to the $mean value from the Tensormap. This value is [0]

broadcast_util.ts:81 Uncaught (in promise) Error: Operands could not be broadcast together with shapes 1,12,24,64 and 0.
    at un (broadcast_util.ts:81)
    at new kn (batchnorm_packed_gpu.ts:32)
    at t.batchNormalization (backend_webgl.ts:869)
    at Bt.engine.runKernel.$x (batchnorm.ts:344)
    at engine.ts:206
    at t.scopedRun (engine.ts:167)
    at t.runKernel (engine.ts:202)
    at $a (batchnorm.ts:343)
    at batchNorm (operation.ts:46)
    at xy (normalization_executor.ts:31)

With some digging in developer tools, it looks like the error begins here.....

Setting 'strict' to true, the model still loads and does not throw and error.

var model = await 
        tf.loadGraphModel('/modelsjs/model_0/model.json', {onProgress:onProgressCallback, strict:true});

Sadly, i am unable to share the model as it is proprietary.

The model is wrong.

Originally a Keras model which was converted to a Tensorflow SavedModel. Converting this model to Tensorflowjs did not work.

However, converting the original model from Keras to Tensorflowjs worked.

Lesson learned, don't mix your models so much!