Skip to content

③ Training tab

When you click ▶ Run Maxent at the bottom of the dock, focus shifts to the Training tab. A progress bar shows fold-by-fold progress and a log panel records what the model is doing in real time — useful for diagnosing slow runs, oversized rasters, or convergence issues.

Step by step

Step 1. Start the run and follow it

Click ▶ Run Maxent at the bottom of the dock. The dock switches to this tab.

Training tab while the model is being fitted

  1. The progress bar.
  2. The current step, for example the jackknife model being fitted.
  3. The log. Every step is written here as it happens.

Step 2. Read the summary

Training tab after the run, with the final summary

  1. The status line reports that the run is finished.
  2. The end of the log summarises the model and the evaluation (see Reading the log).
  3. The bottom line of the dock keeps the key numbers: the number of presence and background points, the training AUC and the cross-validated AUC.

Step 3. Keep a copy of the log

Clear log, Copy log and Save log as buttons

  1. Clear log empties the log window.
  2. Copy log copies the whole log to the clipboard.
  3. Save log as... saves the log as a text file, which is useful as a record of the run.

What happens when you click Run

QMaxent runs the following pipeline; each step writes a status line to the log:

flowchart TD
    A["Sample background points"] --> B["Extract raster covariates<br/>(presence + background)"]
    B --> C["Apply maxnet auto-rule<br/>(or your manual feature choice)"]
    C --> D["Fit full Maxent model"]
    D --> E["Compute training ROC + AUC"]
    E --> F["Run spatial K-Fold CV<br/>(repeat fit per fold)"]
    F --> G["Compute pooled CV AUC"]
    G --> H["(optional) Jackknife loop:<br/>2N additional fits"]
    H --> I["Save .pkl, model card, coefficients<br/>+ results.xlsx"]

The fit is performed by elapid's MaxentModel, a faithful Python port of the original Java MaxEnt (Phillips et al. 2006; Phillips et al. 2017) that delegates the regularized likelihood maximisation to scikit-learn (Pedregosa et al. 2011). The result is numerically very close to Java MaxEnt for identical inputs and identical hyperparameters — the residual difference comes mostly from how each implementation handles ties in the optimisation step.

Reading the log

A typical successful run produces a log like:

→ 10,000 background points sampled
Extracting raster covariates for presence points…
Extracting raster covariates for background points…
→ Presence: 116, Background: 10,104
→ Feature types: ['linear', 'quadratic', 'product', 'hinge']
Training MaxentModel…
→ Model training complete
→ Model saved: …/model.pkl
Computing ROC curve…
→ Training AUC = 0.9569
Running cross-validation…
  Fold 1: 22 test presences, AUC = 0.7453
  Fold 2: 21 test presences, AUC = 0.7839
  Fold 3: 39 test presences, AUC = 0.8097
  Fold 4: 26 test presences, AUC = 0.8614
  Fold 5:  8 test presences, AUC = 0.5903
→ CV AUC = 0.7436 ± 0.0750  (n=5 fold(s))

A few patterns to recognise:

  • Presences dropped (⚠ N of M presence points fall outside the rasters or on NoData cells and were excluded.): the log now says how many presence points were lost and why. A large share usually means a CRS problem or rasters that do not cover the study area.
  • Presences sharing a raster cell (ℹ N presence point(s) share a raster cell with another presence.): two or more records fall in the same cell of the first raster. maxent.jar removes such duplicates by default; QMaxent keeps them, so they count more than once in the fit. If this is not intended, thin the records spatially before modeling (Boria et al. 2014).

  • Background = 10,104 (background + presence-added rows) — points that fell on NoData cells were dropped. Normal and not a cause for concern unless the loss is large (>5 %).

  • Per-fold AUC well below the mean — for example Fold 5 above (AUC = 0.59 with only 8 presences). Geographic K-Fold deliberately produces uneven folds, and a small fold can land on an atypical region of the study area. The pooled mean smooths this out.
  • Train AUC ≫ CV AUC — the spatial-CV gap. A good sign in the sense that it tells you the model has not been silently over-fitted to the training presences (Roberts et al. 2017). A train AUC of 0.95 with a CV AUC of 0.55 is the warning sign — the CV is telling you the model does not generalise.

Completion

When the run finishes, you see the green Done state:

Training tab at 100% with the full log

The dock's footer status bar populates with train AUC=… CV AUC=… and the ④ Results tab unlocks.

Cancelling a run

Click Cancel to abort. QMaxent terminates the current fold cleanly, discards the in-memory model, and leaves disk outputs untouched. You can restart by clicking ▶ Run Maxent again.

Common warnings

Message Meaning Action
⚠ 12 of 116 presence points fall outside the rasters or on NoData cells and were excluded. Some points fell outside the rasters or on NoData cells Inspect the raster stack; may indicate the wrong CRS or a clipped raster
ℹ 3 presence point(s) share a raster cell with another presence. Duplicate records at the raster resolution are kept Thin the presences if the duplicates are not intended
⚠ Background points cap reduced from 10,000 to N Study area too small for the requested background size Reduce the background-points field on Data tab
⚠ Convergence not reached after 500 iterations Hard-to-fit model (often very few presences) Increase the regularization multiplier or simplify the feature set

Next

Move to the ④ Results tab to inspect response curves, Jackknife importance, and run a spatial projection.