Learning-Rate Comparison Plot Calculator

Comparison results

Load data or generate a simulation to compare learning rates.

No active dataset

Recommended rate

—

Best metric

—

Fastest convergence

—

Most stable rate

—

The calculator will explain the strongest learning rate after analysis.

Interactive comparison plots

Hover, zoom, isolate curves, and inspect individual epochs.

Learning-rate ranking

Compare final quality, speed, stability, and diagnostic status.

Rank	Learning rate	Initial	Final	Best	Best epoch	Convergence	Improvement	Mean	Std. dev.	Stability	Gap	AUC	Status	Score
No analysis is available.

1. Data input

Data source

Metric to compare

Metric objective

Custom metric label

X-axis unit

Missing-value policy

CSV or tab-separated observations

learning_rate,epoch,train_loss,validation_loss,train_accuracy,validation_accuracy,gradient_norm,lr_value,custom_metric
0.0001,1,1.180,1.205,42.0,40.5,1.85,0.0001,0.31
0.0001,2,1.110,1.148,45.0,43.1,1.73,0.0001,0.34
0.0001,3,1.045,1.096,48.2,46.0,1.64,0.0001,0.38
0.0001,4,0.992,1.050,51.1,48.8,1.55,0.0001,0.41
0.0001,5,0.948,1.018,53.6,51.0,1.47,0.0001,0.44
0.001,1,1.120,1.160,45.0,43.0,1.80,0.001,0.36
0.001,2,0.920,0.985,55.0,51.9,1.52,0.001,0.49
0.001,3,0.745,0.835,63.0,58.5,1.30,0.001,0.59
0.001,4,0.610,0.725,70.0,64.1,1.10,0.001,0.68
0.001,5,0.510,0.648,76.0,69.0,0.92,0.001,0.75
0.01,1,1.050,1.110,48.0,45.0,2.10,0.01,0.39
0.01,2,0.680,0.760,68.0,63.0,1.48,0.01,0.64
0.01,3,0.460,0.560,80.0,73.0,1.06,0.01,0.78
0.01,4,0.350,0.500,86.0,76.0,0.88,0.01,0.84
0.01,5,0.300,0.535,89.0,74.0,1.15,0.01,0.86
0.1,1,1.000,1.080,50.0,47.0,3.20,0.1,0.41
0.1,2,0.720,0.890,66.0,58.0,4.60,0.1,0.62
0.1,3,0.910,1.160,58.0,49.0,6.20,0.1,0.51
0.1,4,1.280,1.590,45.0,38.0,8.40,0.1,0.37
0.1,5,1.710,2.040,34.0,29.0,10.10,0.1,0.26

Required columns are learning_rate, epoch, and the selected metric.

Manual observations

Rate	Epoch	Train loss	Validation loss	Train accuracy	Validation accuracy	Gradient norm	LR value	Custom	Action

Upload CSV file through PHP

2. Learning-rate schedule

Schedule type

Decay factor

Decay interval

Minimum rate

Warmup epochs

Plateau patience

Custom multipliers or absolute rates

Values repeat when the sequence ends.

3. Analysis controls

Convergence tolerance

Minimum improvement

Plateau window

Divergence multiplier

Oscillation sensitivity

Early-stopping patience

Minimum epochs

Ignore first epochs

Recommendation criterion

Epoch comparison

Generalisation gap

Balanced-score weights

Final quality

Convergence speed

Stability

Generalisation

4. Smoothing and plot options

Smoothing

Smoothing window

Primary chart

Secondary chart

Y-axis scale

X-axis scale

Line width

Line style

Marker style

Marker size

Legend position

Curve interpolation

Axis tick density

X minimum

X maximum

Y minimum

Y maximum

Custom x-axis title

Custom y-axis title

Show gridlines

Show raw curves too

Highlight recommended rate

Reverse y-axis

Export and sharing

Formula used

The calculator measures improvement from the first usable observation to the final observation. It respects whether lower or higher values are better. Percentage improvement uses the first absolute value as its scale.

Improvement % = 100 × (initial − final) ÷ |initial|, for lower-is-better metrics Improvement % = 100 × (final − initial) ÷ |initial|, for higher-is-better metrics Mean = Σ(metric values) ÷ number of observations Standard deviation = √[Σ(value − mean)² ÷ number of observations] AUC = Σ[(xᵢ₊₁ − xᵢ)(yᵢ + yᵢ₊₁) ÷ 2] Generalisation gap = validation loss − training loss Accuracy gap = training accuracy − validation accuracy

The balanced recommendation normalises quality, speed, stability, and generalisation. User-defined weights combine those components. Divergent runs receive an additional score penalty.

How to use

Paste CSV data, enter observations, or generate a simulation.
Select the metric and its optimisation direction.
Choose convergence, smoothing, schedule, and plotting controls.
Press Analyse learning rates to calculate rankings.
Inspect warnings, curves, summary tiles, and the comparison table.
Export plots, results, or a PDF report when finished.

Example data format

learning_rate	epoch	train_loss	validation_loss	train_accuracy	validation_accuracy
0.001	1	1.12	1.16	45	43
0.001	2	0.92	0.985	55	51.9
0.01	1	1.05	1.11	48	45
0.01	2	0.68	0.76	68	63

Frequently asked questions

What does a learning rate control?

It controls each optimisation step size. Small values learn slowly. Large values may overshoot useful parameter regions.

Which learning rate is automatically best?

No rate is universally best. The recommendation depends on quality, convergence, stability, and generalisation. Change the criterion when priorities differ.

Why can validation loss rise later?

The model may begin overfitting. Training loss can still decline. Early stopping may preserve better validation performance.

What indicates a learning rate is too large?

Strong oscillations are common evidence. Exploding loss is more serious. Large gradient norms can confirm instability.

What indicates a learning rate is too small?

Improvement remains consistently slow. The curve may look smooth but nearly flat. Convergence can require excessive training.

Should smoothed or raw curves guide decisions?

Use both when possible. Smoothing reveals trends. Raw curves preserve instability and unusual spikes.

Can different runs have unequal epochs?

Yes, unequal lengths are supported. Shared-range comparison remains optional. Missing points can be skipped or filled.

Why use a logarithmic axis?

It helps compare wide numeric ranges. Every plotted value must remain positive. Invalid values trigger warnings.

Does the calculator train a real model?

It analyses supplied or simulated histories. It does not execute model training. Import logs from your training system.

Comparison results

Interactive comparison plots

Learning-rate ranking

1. Data input

2. Learning-rate schedule

3. Analysis controls

Balanced-score weights

4. Smoothing and plot options

Export and sharing

Formula used

How to use

Example data format

Frequently asked questions

What does a learning rate control?

Which learning rate is automatically best?

Why can validation loss rise later?

What indicates a learning rate is too large?

What indicates a learning rate is too small?

Should smoothed or raw curves guide decisions?

Can different runs have unequal epochs?

Why use a logarithmic axis?

Does the calculator train a real model?

Related Calculators