• State TI Thinks Resolved
  • Locked Locked
  • Replies 5 replies
  • Answers 1 answer
  • Subscribers 96 subscribers
  • Views 473 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

TDA2SX: Emotion recognition - TIDL import

Sasa Mihajlica
Intellectual 860 points

Part Number: TDA2SX
Other Parts Discussed in Thread: TDA2

Hi to all,

I am trying to import emotion recognition model and I am facing some difficulties. Model have 7 output classes.  For every input image I have same output index (same class) and same output for every input image. I looked at deploy.prototxt and added flatten layers before every InnerProduct layer. Also, I am using sparse convolution, so with kernel_size: 5 I think there shouldn't be a problem for convolution layers, should it? I don't see any other limitation that could affect me, but it will be useful if some expert take a look at model and import file to check if I missed something.https://e2e.ti.com/cfs-file/__key/communityserver-discussions-components-files/791/emotion_2D00_recognition.7z

I am using TIDL 01.01.01.00 with PSDK 03.04.00.00. There are model files in attachment.

Thank you in advance and regards,
Sasa

  • 0
    Intellectual 860 points

    https://e2e.ti.com/cfs-file/__key/communityserver-discussions-components-files/791/8637.emotion_2D00_recognition.7z

  • 0 in reply to Sasa Mihajlica
    TI__Genius 17580 points

    Hi Sasa,

    Form the "deploy.prototxt" it looks like TIDL can support all the layers, but in the deploy.prototxt, can you use different output names for different flatten layers and try?

    Also, Did out check that outputs correct with caffe inference on these input images and your model?

    Thanks,

    Praveen 

  • 0 in reply to Praveen Eppa1
    TI__Genius 17580 points

    Hi Sasa,

    Did you get a chance to try out above suggestions?

    Thanks,

    Praveen 

  • 0 in reply to Praveen Eppa1
    Intellectual 860 points

    Hi Pravenn,

    I only managed to try renaming flatten layers to have different names and there was no success. I have usually class 4 detected and for some of images class 5.

    layer {
  name: "flatten1"
  type: "Flatten"
  bottom: "pool_3"
  top: "flatten1"
  flatten_param {
    axis: 1
  }
}
layer {
        type: "InnerProduct"
        name: "hidden_1"
        inner_product_param {
                num_output: 200
        }
        bottom: "flatten1"
        top: "hidden_1"
}

    These are changes made. I did the same with different names for other flatten and innerProduct pairs of layers.

    Also I have a question about checking correctness. This same caffe model is tried on another platform and worked correctly (same model and same inputs) . I wanted to run it on TDA2 to have comparison, so I can confirm that model is correct.

  • 0 in reply to Sasa Mihajlica
    TI__Genius 17580 points

    Hi Sasa,

    Ideally we expect this to work properly, but not sure where the problem is. So I would suggest to check the layer level outputs from import tool against the caffe reference outputs to identify which layer is causing the issue, please refer to section 3.8 (Matching TIDL inference result) in the TIDL user guide on how to do this.

    Thanks,

    Praveen