Calculated results
Waiting for inputSoftmax probability plot
Temperature comparison
This view shows sharpening, smoothing, entropy changes, and possible class-rank movement.
Probability results table
Select a numeric cell to copy it.
| Sample | Index | Class | Logit | Shifted | Exponential | Probability | Percentage | Rank | Predicted | Threshold |
|---|
Sample comparison
| Sample | Prediction | Confidence | Margin | Entropy | Top-k share |
|---|
Average class probabilities
| Class | Average probability | Prediction frequency |
|---|
Calculation steps
Confidence statistics
Input and calculation controls
Softmax settings
Chart controls
Formula used
Pi = exp((zi − max(z)) / T) ÷ Σ exp((zj − max(z)) / T)
Each logit is shifted before exponentiation. This prevents avoidable overflow. Temperature controls the final distribution.
A lower temperature sharpens the probabilities. A higher temperature smooths them. Every probability should sum to one.
How to use
- Choose a preset or enter class names.
- Add sample labels and logits.
- Set temperature, precision, threshold, and top-k.
- Select the required chart style.
- Calculate and inspect every result section.
- Copy, print, or export the report.
Example data format
| sample | Cat | Dog | Bird |
|---|---|---|---|
| Image 1 | 2.4 | 1.2 | -0.3 |
| Image 2 | 0.5 | 2.1 | 0.8 |
Understanding softmax probabilities
What softmax does
Softmax converts unrestricted logits into comparable class probabilities. Larger logits receive larger shares. The total remains one.
Softmax and sigmoid
Softmax models mutually exclusive classes. Sigmoid scores classes independently. Multilabel problems usually use sigmoid outputs.
Confidence and correctness
High confidence does not guarantee a correct prediction. Models can become overconfident. Calibration should be checked separately.
Temperature scaling
Temperature changes probability sharpness without changing logits. It often supports calibration. Class rankings usually remain unchanged.
Softmax limitations
Softmax only compares supplied classes. Unknown classes remain unrepresented. Distribution shifts can reduce reliability.
Frequently asked questions
Why use stable softmax?
Stable softmax subtracts the maximum logit. This avoids numerical overflow. The final probabilities remain mathematically equivalent.
What does temperature equal to one mean?
A temperature of one uses ordinary softmax. No extra sharpening occurs. No extra smoothing occurs.
Can logits be negative?
Yes, logits may be negative. Softmax accepts unrestricted real values. Their relative differences determine probabilities.
Why do probabilities sum to one?
Every exponential value uses the same denominator. Each value becomes a share. All shares therefore total one.
Does softmax change class ranking?
Ordinary positive temperature scaling preserves ranking. Larger logits remain more probable. Ties may remain tied.
What is entropy?
Entropy measures uncertainty across classes. Higher entropy means a flatter distribution. Lower entropy indicates stronger concentration.
What is the confidence margin?
The margin is the top probability minus second place. Larger margins suggest separation. They do not prove correctness.
What does top-k cumulative probability show?
It sums the largest k probabilities. This shows concentrated probability mass. It helps review broader candidate sets.
Can CSV data contain multiple samples?
Yes, each CSV row becomes one sample. The first column supplies labels. Remaining columns supply class logits.