Infer.NET user guide : Learners : Bayes Point Machine classifiers : Command-line runners
Convergence diagnosis
The DiagnoseTrain module is used to assess the convergence of the message-passing algorithms used to train the Bayes Point Machine classifiers. It is available for both binary and multi-class classification. The DiagnoseTrain module reads a training set and produces a CSV file containing the differences in the posterior weight distributions, computed for a specified number of consecutive iterations of the training algorithm. When the training algorithm is converging, the differences will become smaller and smaller (see sample output below).
The DiagnoseTrain module has the following command-line arguments:
Required arguments
training-set: The file with training data containing ground truth labels and features in the format described earlier.
Optional arguments
results: The CSV file to which the weight distribution differences will be saved.model: The file to which the trained Bayes Point Machine classifier will be saved.iterations: The number of training algorithm iterations (defaults to 30).batches: The number of batches into which the training data is split (defaults to 1).
For more information about the command-line arguments, see Settings. A more detailed explanation of training is available here.
Example
Learner Classifier BinaryBayesPointMachine DiagnoseTrain
--training-set training.dat --model trained-binary-bpm.bin
--results results.csv --iterations 500 --batches 1
Learner Classifier MulticlassBayesPointMachine DiagnoseTrain
--training-set training.dat --model trained-multiclass-bpm.bin
--results results.csv --iterations 500 --batches 1
Sample output
This is what gets written to the command-line window (console):
Data set contains 10000 instances, 2 classes and 5888 features.
[17:16:09] Starting training...
[17:16:10] Iteration 1 dp = 37641.0378952995 dt = 583ms
[17:16:10] Iteration 2 dp = 37616.6784075578 dt = 242ms
[17:16:10] Iteration 3 dp = 17.7014780451184 dt = 298ms
[17:16:10] Iteration 4 dp = 1.26924794894729 dt = 173ms
[17:16:11] Iteration 5 dp = 0.769664011372179 dt = 293ms
[17:16:11] Iteration 6 dp = 0.15651774182376 dt = 192ms
[17:16:11] Iteration 7 dp = 0.164699512014064 dt = 267ms
[17:16:11] Iteration 8 dp = 0.0835439679522345 dt = 228ms
[17:16:12] Iteration 9 dp = 0.068138911019533 dt = 288ms
[17:16:12] Iteration 10 dp = 0.0439520998092477 dt = 246ms
[17:16:12] Iteration 11 dp = 0.0322686108618166 dt = 341ms
[17:16:13] Iteration 12 dp = 0.0237455855736709 dt = 238ms
[17:16:13] Iteration 13 dp = 0.0175368057738852 dt = 282ms
[17:16:13] Iteration 14 dp = 0.0129868852183733 dt = 217ms
[17:16:13] Iteration 15 dp = 0.00964343919792277 dt = 292ms
[17:16:14] Iteration 16 dp = 0.0071772245476559 dt = 173ms
[17:16:14] Iteration 17 dp = 0.00535275621234321 dt = 233ms
.
.
.