.tflite Model Dimensions Change After Using BYOM

Report bugs
#1

I’m trying to use the BYOM feature and upload a .tflite model into the EI Studio. The model is designed for audio classification using spectrograms. The model input shape is (1, 99, 43, 1) when I get the input details with TensorFlow. However, when I upload it into the studio, it says my input shape is (99, 43, 1). Edge Impulse says it wants one dimension for audio classification. The Job ID where I uploaded my pre-trained model is 14894009 (my file was converted_model.tflite)

Let me know if you need any more files or have any questions!

#2

Hi @jeanelled,

TensorFlow includes the batch dimension (the first “1”) whereas Edge Impulse drops that dimension. As a result, you will need to remove that dimension from input data if you are using a model trained on Edge Impulse. For example, you could use Numpy’s squeeze (numpy.squeeze — NumPy v1.26 Manual) if you are working in Python.