Open muthukumaranR opened 4 years ago
naturally, The snapshot of the model with the lowest validation loss is stored for evaluation and further use.
model overfitting, reducing param # and re-running
Some sample predictions on test set:
since The predictions look good, the model just needs to be trained with a lower learning rate for a smoother graph. It just learns too fast, especially with the augmentations included.
@muthukumaranR are we getting these results from the same model posted above the images?
yes, @xhagrg
hmm, let's try with some other test samples for this. see if we get better/worse results. (skeptical because of the graph)
predictions can be found here:https://drive.google.com/drive/u/1/folders/15FfMvlFUL9qajKrbrim_bhOfjrl3Ucld
any false positives? let's run it for a day's worth of data.
Updated loss plot with a slower learning rate
@muthukumaranR noice!
Testing the model on random NON-HLD images for a day's worth of data from worldview, the model is susceptible to false positives. Taking these example and adding them to the training/validation cases to reduce false positives.
some false positive cases from the model:
The false positives were due to not scaling the images. once the images are scaled, there were no false positives.
initial Model Summary: 3 layer standard Unet
Run #1: precision: 0.79 recall: 0.70 accuracy: 0.83, f1_score: 0.749