FrozenGene commented on issue #4828: [QNN][TFLite] TFLite rounding mode support URL: https://github.com/apache/incubator-tvm/pull/4828#issuecomment-596615575 > > > > CI's LLVM is LLVM 4.0. I think we should supply `quantized add` so that we avoid it is just a lucky. > > > > > > > > > Just tested, it's a lucky case indeed. > > > BTW, is it possible to follow tensorflow to have a specific llvm version for each tvm release to avoid > > > certain problems like this? > > > > > > To avoid lucky case, when we support TFLite rounding, we could change the `data = get_real_image(224, 224)` to > > ```python > > np.random.seed(0) > > data = np.random.random_integers(low=0, high=128, size=(1, 224, 224, 3)).astype('uint8') > > ``` > > > > > > back. > > Update：I made a mistake for the unit test that I still compare the labels instead of numerical predictions, so rtol and atol set to zero does not make any sense. Now that I changed to compare numerical prediction vector, the results begin to differ for mobilenetV2. And further, changing qnn.add rounding mode to TFLITE does not lead to bit exact execution. I have to reimplement the whole algorithm again in TFLITE add op to see it works. But I'm bothered by the segfault problem to get any reasonable test results. > > > For LLVM version, I think we can not at least currently. We support different LLVM versions on TVM. We have the check in `FindLLVM.cmake` to make sure the minimal requirement is 4.0. > > Could you dig it more? If we can make sure it is the bug of LLVM, we then can discuss how to deal with this case. > > I digged a little bit on this, but it seems to be not related with llvm but rather with tflite conversion from tensorflow graph to tflite (not even related to tvm). I'm using tf 1.15 for my test. And I find it's related to this test `_test_forward_elemwise(partial(_test_add, fused_activation_function="RELU6"))`, specifically `RELU6` will cause the error, if I change it to `RELU`, no segfault will be triggered. @FrozenGene Please help , or maybe I should just for now switch back to an older version of tensorflow. don’t worry. Let us break down the problem. Firstly, CI’s tflite version is 1.13. We could keep your tf version as the same firstly. Second, as I suggested before, you could use tflite_convert tool to convert quantized model layer by layer `tflite_convert --input_format=TENSORFLOW_GRAPHDEF --graph_def_file="xx.pb" --output_file=xx.tflite --output_format=TFLITE --input_arrays=input --input_shapes=1,224,224,3 --std_dev_values=127 --mean_values=127 --inference_type=QUANTIZED_UINT8 --inference_input_type=QUANTIZED_UINT8 --default_ranges_min=0 --default_ranges_max=6 --output_arrays=xx` replace the output arrays args to your target layer name. Like the first layer name / second layer name until you find the issue layer.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
