I am highly interested in innovative ideas and performance as well.
However, I have a few questions regarding BitOPs. It seems that the formula for calculating BitOPs in Table 2 of the CABM paper is EDSR-CABM_x4 BitOPs (3.75G) = 114.23G x (5.80 / 32) x (5.80 / 32).
114.23G for 32bit model
5.80 / 32 for activation
5.80 / 32 for weight
However, the paper and code mention the use of fixed-precision for weights.
Could you please explain how the BitOPs are calculated?
When I run the code, I encounter this below error:
with open(f"{args.model}-[4]-['div2k_valid'].json", "r") as f:
FileNotFoundError: [Errno 2] No such file or directory: "EDSR-[4]-['div2k_valid'].json"
I'm looking forward for your reply.
It's a nice work! Can I just use the student and teacher networks edsr_Γ4_pams_w8a8.pt which provided in CADyQ from Google Cloud to train the edsrbaseline_cabm_simple?