ML model experiments

[ebgoldstein]

• [ ] architecture changes?
• [ ] skip connections (N.B.- it's sequential API now)?
• [x] sedinet architecture
• [ ] finetune from sedinet?
• [x] grayscale vs RGB?
• [x] best LR — loss curve..
• [x] weighted loss
• [x] Quant aware training

[ebgoldstein]

ok, I've looked at a bunch of these...

[ebgoldstein]

Something interesting i noticed. The model is output is is a list of grain sizes that correspond to specific values of the cumulative distribution, from D2 to D98.. so from the grain size representing 2% and 98%... What is interesting is that these values should be sequentially larger, but the model output is not always that case...

here is an example.. the observed values and the model predicted values are printed above the photo... it would be cool to figure out how to enforce this somehow

:

[ebgoldstein]

best Adam learning rate was 1e-2, but w/ that LR the model has higher loss compared to just Adam w/ default and a LR scheduler as a callback...

[ebgoldstein]

I am using a modified sedinet architecture now.. stacks of Separable convs + BN + Pool.. and increasing #s of filters.. works well

[ebgoldstein]

I'm going to close this... we have been experimenting and have a reasonable first model