Work Log Day 86

The model continues to train, using both more data and more batch normalization in place of Spatial Dropout. With the addition of more data, the model has dipped down to 79 MSE. Not a big improvement, it does show that targeting pull times that are less represented in the data does have some impact on model performance. I have also added a third grinder into the mix, the Vario-W, to see if flat burrs grind differently enough for the model to perform poorly. So far that doesn’t seem to be the case, and will continue to use it occasionally for a larger sample of grinders. However the focus is likely to switch away from the primary model to the puck preparation methods.

I finally sat down and spent some time with the initial data I had collected on different methods preparing the coffee grounds. The notebook is here. I initially pulled 10 shots per method expecting a normal distribution of pull times. However looking at the data, it is closer to uniformly distributed. Though it definitely seems like 10 shots is not enough data to make the type of conclusions I would like. Looking at the box plots it seems that the OCD at a depth of 9 seems to produce the ‘best’ range besides no prep. However it seems surprising to see WDT shift so much. I am not convinced that there is something off about the coffee to vary by almost 20 seconds. Now there is an argument to be made that I just don’t know what I am doing when it comes to WDT, and I will consider myself a beginner for these methods. I would much rather use the OCD and get it ‘right’ every time than mess up RDT sometimes when I am not properly decaffeinated.

I collected humidity and temperature data while pulling the shots, as I had over heard baristas lament the impact of ambient temperature and humdity mess with shot times. Plotting the values doesn’t seem to indicate any meaningful relationship, though the space where the shots are pulled does not see a ton of variance in the humdity or temperature.