Skip to main content

Group Bucket Test

testing more explicit group subsampling
Created on January 6|Last edited on January 24

Bidirectional RNN Sweeps




Computing group metrics from first 20 groups
02468epoch9696.59797.59898.599
  1. Visualize only "layer count" tab below--observe that the "acc" values in the table do not match the final accuracies in the top plot on the left, and the plot lines will change substantially for no reason if you sort by a different config variable in the table below. Also note, you should no longer see fractional epochs on the x-axis.

  2. Click "edit" on the top left chart, go to the advanced menu, increase the "Max runs" setting to the higher value (total runs visible in the table--also has an absolute global max of 1000).

  3. This should fix the stochasticity of the accuracy values based on table sort and give a more accurate result that matches (modulo a rounding error) the acc in the table.

  4. Now select "Hidden size" instead of "Layer count" as the only checked tab in the table. In the bottom chart, repeat this process--you will then see the error for "Showing first 3 groups", which you can correct separately.

Computing group metrics from first 3 groups
02468epoch96979899
Layer count
315
Batch size
210
Hidden size
210
All Bi-RNN runs
315


TL;DR: Logging basic PyTorch models




Run set
854


01: Effect of hidden layer size on basic feedforward net




Vary Hidden Layer Size
11


02: RNNs: Balance Generalization with Overfitting




1 Layer Count
27
2 Batch Size
27
3 Hidden Size
27


02: Overfitting in RNNs/LSTMs (dupe)




RNN by Layer Count
27


CNN Layer Size Combinations




Run set 1
7
Run set 2
14


RNN Experiments




All RNN Experiments
32