graph output of uff file for TensorRT

[JadBatmobile]

Hey Andres,

I believe i understand whats going on, but id like to be sure. the UFF file for deploying the model in tensorrt outputs the unnormalized logits layer from the neural network, as opposed to the mask?

[tano297]

Correct!

[JadBatmobile]

Great, I noticed that you do not do a softmax on the logits, you just take the argmax immediately, which i realize is equivalent. I also see in the c++ deploy package, in netTRT.cpp, line 160: _sizeof_out = num_classes * _size_in_pix * sizeof(int);

i am curious why you used sizeof(int), instead of sizeof(float), since the output logits are float32 types. Now, i recognize that sizeof(int) is actually = to sizeof(float), but im curious why you explicitly stated it as int

Thank you!

[JadBatmobile]

Hey Andres, i have another question, this is unrelated and more theoretical, perhaps this is the wrong place, but id be down to email you instead.

When preprocessing images for segmentation learning with CNN's, there are multiple options, (1) take the mean for a pixel across the entire data set, and the std dev for the pixel across the data set, then subtract that mean from that pixel in each image and divide by the std dev. do this for all pixels. (2) for each image, find the mean value and std deviation of all your pixels, and normalize each pixel with those. Or (3) for each image, for each channel, find the mean value and deviation and normalize each pixel in each channel, then merge the channels.

I am wondering what your thoughts on those options are