Using convert_weights_to_tf_lite.py did not produce the same results as the pre-training model.

[FrozenEelFanGirl]

I used this project to retrain on DNS-challenge dataset. But When I finished the training model, I tried to convert the model using convert_weights_to_tf_lite.py, but found that the model I converted(_1.tflite 372k, _2.tflite 641k) did not match the pre-training model(model_quant_1.tflite 369k, model_quant_2.tflite 635k). So I tried to convert pre-training model(model.h5 and other two h5 models) to tflite, they did not match the pre-training model either. Is there any suggestion for me? I wonder why I could not convert pre-training model to get consistent results. I used tf 2.10.0 which should be the last version to be supported on windows and just tried cpu to convert.

[FrozenEelFanGirl]

I tried to check the model information of pretrained quant_model and converted quant_model, and they're totally different:
There is pretrained quant_model:
Input details:
[{'name': 'input_2', 'index': 0, 'shape': array([ 1, 1, 257]), 'shape_signature': array([ 1, 1, 257]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}, {'name': 'input_3', 'index': 1, 'shape': array([ 1, 2, 128, 2]), 'shape_signature': array([ 1, 2, 128, 2]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}]
Output details:
[{'name': 'Identity', 'index': 66, 'shape': array([ 1, 1, 257]), 'shape_signature': array([ 1, 1, 257]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}, {'name': 'Identity_1', 'index': 69, 'shape': array([ 1, 2, 128, 2]), 'shape_signature': array([ 1, 2, 128, 2]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}]
Model details:
Name: input_2
Type: <class 'numpy.float32'>
Shape: [ 1 1 257]
Quantization Parameters: (0.0, 0)
..... and more

There is converted quant_model:
nput details:
[{'name': 'serving_default_input_13:0', 'index': 0, 'shape': array([ 1, 2, 128, 2]), 'shape_signature': array([ 1, 2, 128, 2]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}, {'name': 'serving_default_input_12:0', 'index': 1, 'shape': array([ 1, 1, 257]), 'shape_signature': array([ 1, 1, 257]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}]
Output details:
[{'name': 'StatefulPartitionedCall:0', 'index': 64, 'shape': array([ 1, 1, 257]), 'shape_signature': array([ 1, 1, 257]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}, {'name': 'StatefulPartitionedCall:1', 'index': 69, 'shape': array([ 1, 2, 128, 2]), 'shape_signature': array([ 1, 2, 128, 2]), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}]
Model details:
Name: serving_default_input_13:0
Type: <class 'numpy.float32'>
Shape: [ 1 2 128 2]
Quantization Parameters: (0.0, 0)
..... and more

[StuartIanNaylor]

#52 (comment)

[FrozenEelFanGirl]

Thanks for your help, I will try to fix this.