search

techmn / satmae_pp

Official repository for "Rethinking Transformers Pre-training for Multi-Spectral Satellite Imagery" (CVPR 2024)
Apache License 2.0
89 stars 5 forks source link

Environment details? #1

Closed RichardScottOZ closed 8 months ago

RichardScottOZ commented 8 months ago

Thanks very much for making this available - this is excellent.

I am considering adapting.

Any chance you could share your environment details?

Always tricky getting deep learning and geospatial things working together, so would be nice to know.

Thanks again,

Richard

techmn commented 8 months ago

Hi, environment details are:

python 3.8
pytorch 1.10
cuda 11.1
timm 0.4.12

hope it helps ....

RichardScottOZ commented 8 months ago

Thanks! Will see how I go.

RichardScottOZ commented 8 months ago

A default timm install got me 0.9.16 via pip anyway, so will need to redo.

Appreciate it.

RichardScottOZ commented 8 months ago

FYI, ran a 3 epoch test with 1 warmup on around 20 odd images - e.g. a subset of _1_1 and _0_0 for testing one gpu - which I just went with your disable telling it to distribute and let it fall through your branch there

[12:22:25.163846] Epoch: [2]  [0/2]  eta: 0:00:12  lr: 0.000025  loss: 1.6014 (1.6014)  time: 6.0662  data: 5.3526  max mem: 8379
[12:22:25.595520] Epoch: [2]  [1/2]  eta: 0:00:03  lr: 0.000025  loss: 1.5285 (1.5649)  time: 3.2484  data: 2.6763  max mem: 8379
[12:22:26.272235] Epoch: [2] Total time: 0:00:07 (3.5873 s / it)
[12:22:26.272235] Averaged stats: lr: 0.000025  loss: 1.5285 (1.5649)
[12:22:31.371837] Training time 0:00:34

have a checkpoint etc. - so that seems good as far as framework goes?