I have a simple keyword recognition model, which I have trained on colab using keras. My goal is to finally convert this model into a tflite model and deploy it on a target board.
The limitation of the board is that it doesn’t support ExpandDims layer. While conversion from keras to tflite, the process adds ExpandDims layer explicitly.
The code used for conversion is as follows:
converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
with open(‘mfcc_cnn model.tflite’, ‘wb’) as f:
The model flow from netron for both keras and tflite model is attached here. Can anyone help me out in how to avoid the expand_dims layer during conversion. Quantized models are also not supported by the target board.