Differences between neuralnetwork and mlprogram

[119458]

❓Question

The 160-size data set is trained with 8, 16 and 32 times downsampling. Under pt weight, the target can be classified and identified by adding post-processing to the model. When neuralnetwork conversion mode is adopted, the target can be recognized normally, but when mlprogram is adopted, the target cannot be recognized normally and a bunch of detection boxes will be given. Almost all the detection images are covered.

neuralnetwork == coremltools==6.0 mlprogram == coremltools==7.0

[daikankan]

same issue not resolved: https://github.com/apple/coremltools/issues/2301

[junpeiz]

Hello @119458, could you provide your steps to re-produce the inconsistent results? Thanks!

[daikankan]

@junpeiz I have checked my cases, and found for some models, ct.convert(convert_to="mlprogram", compute_precision=ct.precision.FLOAT16) have big problem of precision, while ct.convert(convert_to="mlprogram", compute_precision=ct.precision.FLOAT32) is correct. also ct.convert(convert_to="neuralnetwork") along with quantization_utils.quantize_weights(model_ct, nbits=16) is correct.