{
  "ticketNumber" : "698104",
  "reporterName" : "vijay m36",
  "rankPoints" : "90",
  "resolutionStatus" : "",
  "ticketName" : "TDA2: TIDL Host emulation tool convolution output issues",
  "rankName" : "Prodigy",
  "replies" : "",
  "views" : "",
  "queryText" : "Part Number: TDA2 Issue with the convolutional output in the function : TIDL_refConv2d_i8u_032s_cn in tidl_conv2d_base.c First, the test setup: A simple convolutional network with 3x3 kernel and bias with 2 output channels. The first kernel has all zeros and the second kernel has all ones. Bias for first channel is 1 and second channel is 2. Now all the variables involved: input raw rgb = 255 (all 1s) inQuantFactor = 65280 weightsQ = 65536 biasQ = 4194304 biasQFact = 255 rgb size input = 3x3x3 num of output channels = 2 In foating point: input = 1.0 weights = [0, 1] (1st kernel all weights are zero, and second kernel all weights are 1) bias = [ 1, 2] The convolved output of the first channel should be 1. Just bias, because all weights are zero. The convolved output of the second channel should be : 1.0 * (3x3x3) *1 + 2 = 29. Ti fixed point: weights having values[0,1] becomes approximately: [0,1]*weightsQ/255 = [0, 255]. biases [1, 2] becomes approximately: [1, 2]*biasQ/255 = [16384,32767]. Now checking the convolution in TIDL_refConv2d_i8u_032s_cn function: There, pInChannel contains inputs which is uint8, and all the values are 255. Bias is multiplied by the biasQFact (255). I understand this is because of the inQuantFactor, and the bias needs to be scaled. convolution happens as: input * weights + bias*biasQFact for first channel, output is = bias*biasQFact = 4177920 for second channel: 255* (3*3*3)*255 + 32767*255 = 10111260 So floating point 1 is represented as 4177920 and 29 is represented as 10111260 which is wrong. Since the bias and weights are not represented by the same Q factor, the output will match only when the input also has a QFactor, With inout Q factor, second channel should be: 255*64 * (3*3*3)*255 + 32767*255 = 120718785, which should be the output. What am i missing here? I have attached the model and inputs to reproduce this problem. ti.zip",
  "imageList" : null,
  "partNumber" : "NA",
  "allResponseList" : [ {
    "contentId" : "",
    "userName" : "Praveen Eppa1",
    "rankPoints" : "17580",
    "rankName" : "TI__Genius",
    "date" : "",
    "userId" : "/members/6019814",
    "content" : "Hi Vijay, Please refer to below e2e thread, where we had one simple cifar net example to understand the tidl layers. I would recommend to start with this to better understand the input,weights and output Q formats. Then on top of it you try your experiments to get the better results. e2e.ti.com/.../689876 Thanks, Praveen",
    "imageList" : null
  }, {
    "contentId" : "",
    "userName" : "vijay m36",
    "rankPoints" : "90",
    "rankName" : "Prodigy",
    "date" : "",
    "userId" : "/members/5160008",
    "content" : "Hi Praveen, The above example is created in such a way that one can figure out what the issue is without doing into the code to debug. Anyway, I went and fixed the TIDL_importQuantLayerParams function in tidl_tfimport.cpp to make the above toy example work. The issue arises when the min - max range of weights and bias are not same. By the way, the cifar net example also didnt match (TIDL .01.01). I loaded the model shared in the above link and obtained outputs from each of the layers to compare with the host emulation output. Outputs doesnt match. I have attached all the tools required to reproduce the issue, in case you guys want to try it. tidl_keras_example.zip",
    "imageList" : null
  }, {
    "contentId" : "",
    "userName" : "Praveen Eppa1",
    "rankPoints" : "17580",
    "rankName" : "TI__Genius",
    "date" : "",
    "userId" : "/members/6019814",
    "content" : "Hi Vijay, Thanks for pointing this. I am closing this thread as you could fix the issue and you can feel free to open this thread or create a new thread if you face any issues. Regards, Praveen",
    "imageList" : null
  } ],
  "tags" : [ ],
  "fourmType" : "processors-forum"
}