You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Hi !
I tried integrating your model into my own pipeline, and seem to have run into some issues. I tried porting your loss implementation into proper modules, but for some reason the loss keeps going into the negative ( e.g -1.51 ) which in turn makes optimization difficult...
Could you maybe take a look if i made a mistake or if there is some bug in the original code ?
Hello there! Have you managed to acheieve any results using this model? I have been trying multiple times now to get the same results as the provided "Block0_skip_model_100", but I think there is something wrong with the model given in the repo. Any model that is trained further than epoch 1 or 2 only results in a model that produces an image where all values are the same. The predicted images have values like 0, 65534 or 2 for or all values when testing predictions.
I am training using the training portion of the "osiDataset", with learning rate of 0.0001 and weight decay of 0.001.
I have tried predictions and training both with and without the provided encoder model in the "pretrained_model" folder.
Below you can see the tensorboard plot of the loss:
I haved used a different loss function ( either RSME or my own, very simple absolute height diff ) I get some results, but they are not very good after 50 or so epochs. However the results are definitly not all the same.
I am also trying to predict building height from satellite images, but i am using a different dataset than in the examples.
I have had moderate success with a modified UNET model, and tried this model for better results, however so far this seems not to work.
Hi !
I tried integrating your model into my own pipeline, and seem to have run into some issues. I tried porting your loss implementation into proper modules, but for some reason the loss keeps going into the negative ( e.g -1.51 ) which in turn makes optimization difficult...
Could you maybe take a look if i made a mistake or if there is some bug in the original code ?
The text was updated successfully, but these errors were encountered: