search

Accioy / SAR2optical

A simple deeplearning model for translating SAR images to optical images.
30 stars 3 forks source link

pre-training weight #2

Open nh945 opened 2 years ago

nh945 commented 2 years ago

Thank you for your sharing. Could you provide the pre-training weight? It cannot run perfectly because of computer reasons,Looking forward to your reply！

Accioy commented 2 years ago

Sorry but I've been out of this field for a long time and I didn't preserve the weights.

nh945 commented 2 years ago

Sorry but I've been out of this field for a long time and I didn't preserve the weights.

Thank you for your reply! I am training this code. I have used 400,000 pictures but trained 24 epochs, but the results have been fluctuating. The network conversion result is not clear, because the training cost is high, I am considering whether to continue training. Can you give me some advice? Like what code to change? Or wait for a long training session? Sincerely wish you do better in other field！Looking forward to your reply.

Accioy commented 2 years ago

I forget how many pictures I used but I got the best result with 150 epochs. The loss should converge stably. Maybe you can try to use less pictures and train more epochs, then check the predictions in both trainset and testset.