albanie / convnet-burden Goto Github PK

Hi, albanie.
How to compute the model Flops, are there some tools?

I see two memories listed here: parameter memory and feature memory. When we plan hardware for inference work, should we put them together as a requirement to memory size of GPU card? thanks

For present 18:
Flops: 2 GFLOPs, feature memory=3GB

For mcn-mobilenet:
Flops: 579 MFLOPs, feature memory=5GB

And the question is why mobilenet get smaller flops, but with larger feature memory. What's the relationship between FLOPS and feature memory?

Please tell me what is mean by res5c_relu? and why
Estimates for a single full pass of model at input size 224 x 224 takes 8GFLOPs
while, res5c_relu features with the network 224 x 224 input sizes using a batch size of 128 takes 974 GFLOPs?

Hi, albanie,
Can I have an example of your matlab model file?
Thank you very much.

FLOPs = Floating-point operations.
FMAs = Floating-point multiplication-adds.
Based on the values you give, I think they should be FMAs instead of FLOPs, and the number of FLOPs is approximately 2 * FMAs. Could you please clarify? Thanks.

Hi @albanie , the flops you estimated is only for inference step, without Back Propagation. Am I right?

thanks

You've calculated the FLOPS per respective CNN architecture, what execution performance can be expected from particular device (GPU or CPU)? means what would be estimated inference time per single image where FLOPS performance of device is known?

How do you calculate the FLOPs for each model? Are there relevant papers or learning materials available?
And is this FLOPs defined as the time it takes to run a forward propagation during testing or the time it takes to complete forward and backward propagation completely during training?