Hi everyone!
I’m experimenting with the baseline Depth Anything V2 model and just wanted to practice the submission pipeline on the SYNS-Patches val set. However, my zipped .npz file easily exceeds 300 MB. I’ve tried a few tricks:
float16 results in an OpenCV error (cv2.resize doesn’t support float16),
uint16 and uint8 cause PyTorch’s evaluator to crash because it calls torch.finfo() on the predictions, expecting a floating dtype,
If I stick to float32 raw, I end up with ~1 GB.
I’m currently exploring per-image normalization to [0..1] in float32, then np.savez_compressed, which helps, but I’m still hovering near 300+ MB.
Has anyone found a good approach that keeps file size under 300 MB and avoids the grader errors? Do you do a specific float quantization, or do you downsample, or something else? I’d love any advice before moving on to my custom-trained models!
Thanks in advance!
Posted by: Theonewhomadethings @ Feb. 16, 2025, 6:37 a.m.I am having the same issue. Any tips on this would be greatly appreciated...
Posted by: goowfd @ Feb. 18, 2025, 7:14 p.m.