Training, Testing and Results
Closed this issue · 5 comments
How do we get metrics on what the performance of the model is, since, I don't think we can use the Loss function for it.
#36
I would Standardize the ground truth and the predicted depth (make both have mean of 0 and standard deviation of 1). Then we could take mean square error of both images. Then find the mean error among the dataset.
We could also use SSIM loss, but then we have to agree on the same hyper parameters we use for evaluation. I suggest we use
tf.image.ssim(
img1, img2, max_val=10, filter_size=11, filter_sigma=1.5, k1=0.01, k2=0.03
)
Where img1 and img2 are standardized.
that sounds good to me as long as we can display some metrics and graph showing the performance.
Is there a standard test to run for these to compare with other available networks?
#37 Added python script to run MobileNet
...