search

Lkruitwagen / deepsentinel

DeepSentinel: a sentinel-1 and -2 self-supervised sensor fusion model for general purpose semantic embedding
43 stars 12 forks source link

Pretraining: channel stat .json file or pretrain path's? #7

Open wtrainor opened 3 years ago

wtrainor commented 3 years ago

Starting new issue, as I'm trying to train now! But I need some guidance.

Having troubles with the pretrain step. I left the default options from your example ML_CONFIG.yaml file. And got an error about the channel stats file: _No such file or directory: '/home/jupyter/deepsentinel/data/channel_stats/MIsland_v1junk.json'​ Is this not generated?

So I tried removing pretraining step by keeping pretrain: blank in ML_CONFIG.yaml but still got an error _No such file or directory: '/home/jupyter/deepsentinel/experiments/sacred/4/pretrainedmodel.pth'​

What should I be doing to either use or not use a pretrained model?

Lkruitwagen commented 3 years ago

Cool, glad you're moving on to the next step!

channel_stats is used to get the distributions of each band(/channel) in the imagery -> important for normalising the data prior to model ingestion. You'll need to use bin/channel_stats.py to first generate the means and standard deviations of your data that will be used to normalise. I think I wrote a cli for it, so give it a go.

For using pretraining or not, hmm, I'm not sure what a blank behaviour should give you. If you want to skip pretraining try setting it to none. (I think in yaml none is lower case, but double-check).

I've actually just finished writing up a bit of a paper on this work which might help. I've submitted it to arXiv but might take a day or so to come through. If you email me lucas.kruitwagen@gmail.com I can send it to you directly.

wtrainor commented 3 years ago

Cool! Got the stats calculated for my data! Thanks. I for the other error: I just needed to create a subdirectory 'tmp' for the .pth files. doh!