Why the result of kpu is different from the result of my computer?



  • As the topic, I have designed a cnn model, and successfully transform it to the .kmodel format by the ncc. Then I use the same pic in the k210 and my computer to calculate, but there are too much different in this two result, and the computer's calculation was correct by evaluation. I post the two calculation results below, and the normal range of the output is 0 to 1, the kpu's calculation results can be said to be messy. How can I solve this or where's the question?alt text my model result's size is 144 in chw format, so the kpu output is a uint_8t[16] the result I haved multiple the scale and add the bias.


  • Staff

    @EdVince

    1. Tensorflow uses NHWC format, while K210 uses NCHW.
    2. The dataset in --dataset option when coverting your model with nncase should have a similar distribution with the dataset used in training process. For example you can use a part of the validation dataset as the option.


  • 0_1549681085982_123.png