VIM3 NPU using Tengine SDK for MobileNet

@Frank I am not sure if you got my question right. My question was about, why are the values of scale and mean different in different files, even though both were classifications. And what parameters for mean and scale are to be set for activation functions having output [0,1].
I am sorry, If my questionn is not clear! :frowning:

@Akkisony Because they use different parameters in the model when they are training

1 Like

@Frank I want to use MobileNet model for checking the inference time of my model. But during creation of graph using Tengine SDK, I got an error - Create graph failed!
During that time, you suggested me to use Caffe SSD instead of MobileNet v1. Is it because MobileNetv1 (trained with tf framework) is not compatible for tengine sdk?

@Akkisony The SSD model is transferred from caffe to tensorflow. The operators used to achieve the same function are different. After converting to tensorflow, these operators all support conversion. I think you should check the documentation to confirm whether your tf-SSD exists. Operators and layers that do not support conversion

1 Like

@Frank Thank you for your time :slight_smile:

Did you mean SSD model is transferred from tensorflow to Caffe? Hence it is Caffe SSD model.

Please correct me if I am wrong. Here MobileNet (Similarly VGG) is the base network and SSD is the detection network right?

I will now have to train my model something like this right? GitHub - weiliu89/caffe at ssd
And then convert it to tengine format.

from caffe to tf


I am not familiar with the caffe framework, you need to explore it yourself

1 Like

Thanks alot @Frank for the clarity. I am not specific to Caffe framework either. My main goal is to monitor the performance of Google Coral Edge TPU vs VIM3 NPU.

NPU is also compatible with Tensorflow - MobileNet SSD model aswell right? If yes, I will train my model with Tensorflow framework. :slightly_smiling_face:

I jumped to Caffe only after reading your suggestion to use “Caffe SSD”.

@Akkisony I have never used tf-ssd, if you check the document, all layers are supported in the document, then it can be converted

1 Like

@Frank Do you have any suggestions for models, which can be used as a baseline for checking the performance of VIM3 and Coral TPU?
I worked with Yolov3 on VIM3 and got an inference time of about 80ms. This cannot be used on Coral TPU as it supports only tensorflow models.
Do you have any suggestion? Just want to get some input from an expert like you in this platform. :wink:

@Akkisony The tensorflow community has a tf version of yolov3, maybe you can find it through google search


@Frank Can you please tell me how Do i measure the power consumption of VIM3 while runnning the model on NPU through Tengine SDK?