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.