Mobilenet-ssd caffe model gives NaN as output

idata · ‎11-24-2017

Hi everyone, I am using mobilenet-ssd caffe model for detection. Using caffemodel and prototxt file from here : https://github.com/chuanqi305/MobileNet-SSD

Using the caffemodel and prototxt I was able to generate graph file without any error using mvNCCompile

But whenever I load my output after detection I see NAN in the output.

        input_image = cv2.resize(input_image,(300,300), cv2.INTER_LINEAR)
        display_image = input_image
        height_crop = display_image.shape[0]
        width_crop = display_image.shape[1]
        input_image = input_image.astype(np.float32)
        input_image = np.subtract(input_image,127.5)
        input_image = np.multiply(input_image,0.007843)
        input_image = input_image.astype(np.float32)
        graph.LoadTensor(input_image, 'user object')
        out, userobj = graph.GetResult()

Any directions are appreciated , thank you .

idata · ‎11-27-2017

@hrsht.sarma I'm attempting to reproduce your issue, but I was unable to generate the graph file from the links you provided. Did you modify the prototxt file in any way? I tried removing the training and data layers but this did not work for me. If you performed any special steps, please let me know and I will try it again. Thanks.

idata · ‎11-28-2017

@Tome_at_Intel isn't mobilenet-ssd supported in current release? I even can not compile the graph file.

idata · ‎11-28-2017

I have following files :

prototxt: https://drive.google.com/file/d/1O11sw2AvG4RlG_JnzQSFodT9WTH7PKaQ/view?usp=sharing

caffemodel : https://drive.google.com/file/d/1d4Bfey6knTGCJtH5P_Q3WXT2J6pdCH6y/view?usp=sharing

I can make a graph file with following warning :

/usr/local/bin/ncsdk/Controllers/FileIO.py:52: UserWarning: You are using a large type. Consider reducing your data sizes for best performance

"Consider reducing your data sizes for best performance\033[0m")

idata · ‎11-29-2017

@hrsht.sarma As I see you are using float32. Input tensor uses half precision (fp16) values.

I also had problem with generating the graph. Can you please share the command line you used with mvNCCompile?

idata · ‎11-29-2017

mvNCCompile MobileNet_deploy.prototxt -w MobileNet_deploy.caffemodel -s 12

idata · ‎11-29-2017

@georgievm_cms Using fp16 half precision , i get output of length 1001 but I dont understand how to interpret this output , thanks for your help. Do you have any hints ?

This is how my output looks like : [ -324.25,520.0,82.812,-1400.0,-425.5,-407.5,-705.5,-1011.5,-351.5,-555.5,1123.0,264.0,923.0,331.0,-677.5………]

idata · ‎12-11-2017

@hrsht.sarma, why I got output of length 707 with the same caffe model you used. You only change the type to fp16 in the script above to get output of length 1001?

idata · ‎12-11-2017

@hrsht.sarma, the weight and prototxt you used is not mobilenet ssd, but mobilenet. That is why you got 1001 results.

prototxt: https://drive.google.com/file/d/1O11sw2AvG4RlG_JnzQSFodT9WTH7PKaQ/view?usp=sharing

caffemodel : https://drive.google.com/file/d/1d4Bfey6knTGCJtH5P_Q3WXT2J6pdCH6y/view?usp=sharing

idata · ‎12-14-2017

@xhuan28 @hrsht.sarma @Tome_at_Intel @georgievm_cms

There are now a few examples for SSD MobileNet in the ncappzoo

Look in the apps/video_object and caffe/SSD_MobileNet projects

idata · ‎12-15-2017

@hrsht.sarma, To address your original question, if you look at the run.py code here: https://github.com/movidius/ncappzoo/tree/master/caffe/SSD_MobileNet you can see how to handle the NaN values in some of the network results.

Neal

idata · ‎12-18-2017

Thanks , I am running mobilenet SSD now.

idata · ‎02-06-2018

@neal_at_intel , @Tome_at_Intel But mobilenet-ssd‘s some layers output is nan when use the size of image larger than 300×300.

Issue: https://ncsforum.movidius.com/discussion/comment/1618/#Comment_1618