Hi, I’m trying to write a report about GPU/CPU performance analysis.
I’ve used CIFAR-10 dataset, and I’ve got some data from profiling tool (VTune Amplifier, Nvidia Nsight Compute), while executing ResNet18.
I’m looking for bigger images. (CIFAR-10 is only 32x32) I’ve heard about ImageNet, but I think it is too heavy for me. I’m a beginner in deep learning. Are there any popular image datasets used in analyzing computing performance? Also I wanna get recommendation about the models. Maybe I’ll use one or more CNN models for it.
The last question is, for performance analysis, should I use training code or just do inference?
Imagenette tries to fill the gap of “small dataset with imagenet-like images”.
You would have to decide what performance you want to analyse. If you plan to deploy a model into production, inference might be most important, otherwise training matters most.
Thank you very much. The size and the data amount are desirable for me.
I have one question more. I’ve checked Imagenette-320 and I found that the images have various size. But all have the form of (320, xxx) or (xxx, 320). As far as I know, if the network has fully connected layer, this different size may be issue. But I’m a newbie in the world of neural network model, so may I ask you for some reference of the way to transform this kind of various image files to tensor? Or the reference of the code using some famous cnn models dealing with this kind of image data?