Analysis results
Interactive loss chart
Training diagnostics
Experiment ranking
| Rank | Run | Monitor | Best loss | Best epoch | Final loss | Status |
|---|
Detailed result table
| Run | Epoch | Training | Validation | Test | Smoothed monitor | Change | Change % | Gap | Status |
|---|
Loss data and settings
Formula used
Absolute reduction: initial loss minus final loss. It shows total improvement. Larger positive values indicate stronger reduction.
Percentage reduction: absolute reduction divided by initial loss. The result is multiplied by one hundred. It compares progress across different scales.
Generalisation gap: validation loss minus training loss. Positive gaps indicate weaker unseen-data performance. Large growing gaps can indicate overfitting.
Recent slope: linear-regression slope across recent epochs. Values near zero indicate convergence. Positive slopes indicate increasing monitored loss.
Early stopping: stop after patience epochs without minimum improvement. The best monitored checkpoint is recorded. Restoring it can improve generalisation.
How to use
- Paste loss data or upload a CSV file.
- Select visible curves, smoothing, and monitoring options.
- Configure convergence and early-stopping thresholds.
- Adjust chart labels, scales, and reference lines.
- Select Analyze loss curves to calculate results.
- Review diagnostics, rankings, tables, and export options.
Example data patterns
| Pattern | Typical interpretation |
|---|---|
| Healthy convergence | Training and validation losses decline together. |
| Overfitting | Training loss falls while validation loss rises. |
| Underfitting | Both losses remain high and improve slowly. |
| Oscillation | Loss repeatedly rises and falls across epochs. |
| Exploding loss | Loss grows rapidly instead of decreasing. |
| Plateau | Loss changes become negligible for many epochs. |
| Early stopping | Validation improvement stops before training ends. |
| Model comparison | Several runs are ranked by validation loss. |
Frequently asked questions
What is a training loss curve?
A training loss curve shows how model error changes during training. Epochs usually appear horizontally. Lower loss generally indicates improved fitting.
Why compare training and validation loss?
The comparison reveals generalisation behavior. Falling training loss with rising validation loss may indicate overfitting. Similar downward curves usually indicate healthier learning.
Can loss values be negative?
Some custom objectives can produce negative values. Many common losses remain nonnegative. The calculator warns instead of rejecting every negative value.
What does a plateau mean?
A plateau means recent loss improvement is very small. Training may have converged. Learning-rate changes or early stopping may help.
How is the best epoch selected?
The calculator uses the minimum monitored loss. Validation loss is preferred when available. Training loss becomes the fallback metric.
What is early-stopping patience?
Patience counts tolerated epochs without meaningful improvement. Training stops when that count is exceeded. The best checkpoint remains available for restoration.
Does smoothing change the original data?
Smoothing only affects displayed or analysed series when selected. Raw values remain preserved. Exports can include both raw and smoothed values.
What indicates unstable optimisation?
Large repeated rises and falls suggest instability. Excessive learning rates often cause this pattern. Noisy batches can also produce oscillations.
Can I compare multiple experiments?
Yes. Add a run column to your data. Every run receives separate curves, statistics, rankings, and diagnostic summaries.
Which CSV columns are supported?
Use epoch, training, validation, test, and run columns. Column names are case insensitive. Missing optional loss columns are accepted.
How should logarithmic loss charts be used?
Logarithmic scaling helps compare large proportional changes. Every plotted loss must be positive. Zero or negative values cannot appear on logarithmic axes.