Plot a run · plot-training-log

Draw the curves recorded by a training run, finished or still running:

adabmDCA plot-training-log model/history.csv

The argument can be history.csv, <label>_history.csv, or the run's log file, which points to its history. train --plot-training-logs does the same automatically when training ends.

Outputs

One PNG per recorded quantity, named <label>_<quantity>.png (training_pearson.png without a label), and an overview combining all of them (<label>_overview.png), in <label>_plots/ next to the history table (training_plots/ without a label; choose another folder with -o).

Quantity	Plots	Available for
pearson	Pearson and slope of the connected correlations, training and validation, with the target	all runs
loglikelihood	Log-likelihood per site, training and validation	PTT
entropy	Model entropy	PTT
density	Fraction of active couplings	sparse models
ladder	Active replicas and models saved along the trajectory	PTT
acceptance	Smallest swap acceptance between neighbouring models	PTT
learning_rates	Field and coupling learning rates	PTT adaptive
kl	Predicted KL divergence of each update	PTT adaptive
lag	Lag of the chains behind the model	PTT adaptive

When events.jsonl and training.json are present, the validation stop, lag pauses and recoveries are marked on the curves. History tables from older versions of the package are read too.

Training: reading the plots shows each figure on a real run and says what healthy and unhealthy curves look like.

In Python

import pandas as pd

history = pd.read_csv("model/history.csv")
ax = history.plot(x="step", y=["pearson", "pearson_val"])
ax.figure.savefig("pearson.png")

A TrainingResult returned by train_model gives the same table with result.history_dataframe().