---
kind: concept
slug: focal-loss
title: Focal loss
summary: >-
  Cross-entropy scaled by (1 - p_t)^gamma, often with an alpha weight, so easy examples count less. Sixteen wiki cards use it across detection, heatmaps, tracking, segmentation, action localisation, scene graphs and set prediction. Few isolate its gain, and many leave alpha or gamma unstated.
tags: [loss-function, class-imbalance]
papers: ["1905.12365", "1910.07655", "2003.12962", "2011.04841", "2101.07511", "2103.01100", "2103.13137", "2108.06152", "2201.01501", "2202.13065", "2202.07925", "2205.12257", "2304.05170", "2305.11175", "2307.02694", "2408.15857"]
related: [anchor-free-detection, bipartite-matching, class-balanced-loss, clustered-federated-learning, convolutional-neural-network, dice-loss, long-tailed-recognition, multimodal-fusion, multimodal-language-model, object-detection, object-queries, scene-graph-generation, semantic-segmentation, set-prediction, visual-object-tracking]
status: draft
updated_at: 2026-11-01
---

## Overview

Plain cross-entropy gives every training example the same kind of push, so a
large number of easy examples can dominate the gradient when positives are
rare. Focal loss multiplies the per-example loss by a modulating factor
$(1 - p_t)^{\gamma}$, with $\gamma \ge 0$, where $p_t$ is the model's
probability for the true class. The factor is close to zero when $p_t$ is near
1, so well-classified examples are down-weighted [arxiv:2101.07511 eq. 2]. An
$\alpha_t$ weight is often added on top to balance classes
[arxiv:2101.07511 §3.2.2].

The wiki cards use the loss in several shapes: binary classification with
imbalanced classes [arxiv:2101.07511 §3.2.1], per-pixel heatmaps where
positives are single pixels [arxiv:2304.05170 §4.3] or object centres
[arxiv:2011.04841 §3.2], per-pixel depth distributions over bins
[arxiv:2103.01100 §3.5], per-step action classes in a video
[arxiv:2202.07925 §3.3], softmax classification of boundary candidates in a
temporal localiser [arxiv:2103.13137 §3.4], confidence values for 2D-3D matches
[arxiv:2205.12257 §3.2], classification over many set-prediction outputs
[arxiv:2202.13065 §3.4], a generalised form for continuous targets
[arxiv:2201.01501 §3.3], a per-node focusing parameter in a scene-graph loss
[arxiv:2003.12962 eq. 8], a per-anchor confidence in a one-stage 2D
detector [arxiv:1905.12365 §4.2], a survey's pixel-wise binary form with
separate foreground and background weights [arxiv:1910.07655 §4.3], a
classification term in a DETR-style decoder [arxiv:2108.06152 §3.5], a
category-agnostic term in a vision-language tokenizer [arxiv:2305.11175 app.
B], and a survey's generic definition with no default values
[arxiv:2307.02694 §5.1.6]. In every case the card gives the loss as its
authors define it, and the values for $\alpha$ and $\gamma$ differ from card to
card.

## Variants

### Alpha-balanced focal loss (binary classification)

The form in the COVID-19 federated paper. The loss is $-\alpha_t
(1 - p_t)^{\gamma} \log p_t$ [arxiv:2101.07511 eq. 2]. Setting $\gamma = 0$
leaves an $\alpha_t$-weighted cross-entropy. The card does not say how
$\alpha_t$ is set per class or per sample, and gives no value for $\alpha$ or
$\gamma$ [arxiv:2101.07511 §3.2.2]. Card: [Federated COVID-19 diagnosis from X-ray and ultrasound](/papers/2101.07511).

### Foreground and background weights (survey form)

The segmentation review writes a binary version with two weights. The
foreground term is multiplied by $\alpha$ and the background term by
$1 - \alpha$, both with the $(1 - p)^{\gamma}$ modulator, and the review states
that $\gamma = 0$ gives binary cross-entropy [arxiv:1910.07655 §4.3]. Here
$\alpha$ is not the class weight $\alpha_t$ of the variant above, so a value
copied from one card does not carry over to the other. The review's plots use
$\alpha = 0.25$ and $\gamma = 2$, which it chose for its own loss-behaviour
figures; these are not recommended settings [arxiv:1910.07655 §4.4]. The review
runs no training experiments of its own. Card: [Deep Semantic Segmentation of Natural and Medical Images: A Review](/papers/1910.07655).

### Penalty-reduced focal loss on heatmaps

The tracker uses a CornerNet-style loss over a centre heatmap. Pixels where
the target is exactly 1 get $(1 - \hat h)^{\gamma} \log \hat h$. All other
pixels get an extra $(1 - h)^{\beta}$ factor, so pixels near a centre are
down-weighted [arxiv:2304.05170 eq. 3]. The card sets $\gamma = 2$ and
$\beta = 4$ and says these follow CornerNet; the paper has no alpha term
[arxiv:2304.05170 §4.3]. Card: [SportsMOT: A Large Multi-Object Tracking Dataset in Multiple Sports Scenes](/papers/2304.05170).

The radar-camera detector trains its primary centre heatmap with the same
shape, written with exponents $\alpha$ and $\beta$. The paper states neither
value. Its code sets them to 2 and 4 [arxiv:2011.04841 §3.2] [code:
src/lib/model/losses.py#L51-L55]. These are the same numbers the tracker card
reports, but the radar paper's $\alpha$ is an exponent, not a class weight.
Card: [CenterFusion: Center-based Radar and Camera Fusion for 3D Object Detection](/papers/2011.04841).

### Focal loss over depth bins

The monocular 3D detector trains its depth-distribution network with a focal
loss on each feature-map pixel. The target is a one-hot depth bin, and the loss
is averaged over all pixels [arxiv:2103.01100 §3.5] [arxiv:2103.01100 eq. 3].
The card gives $\gamma = 2.0$ and an alpha of 3.25 for pixels inside a 2D
object box and 0.25 elsewhere [arxiv:2103.01100 §3.5]. The code uses a single
alpha of 0.25 and multiplies foreground and background losses by 13 and 1
[code: tools/cfgs/kitti_models/CaDDN.yaml#L58-L61]. The two give the same
effective per-pixel factors, but by different mechanisms (see Contradictions).
Card: [Categorical Depth Distribution Network for Monocular 3D Object Detection](/papers/2103.01100).

### Focal classification over time steps

The temporal action localiser sums a focal loss over action classes at every
time step and every pyramid level. The sum is divided by the number of positive
steps [arxiv:2202.07925 eq. 7]. The paper states neither $\alpha$ nor $\gamma$
[arxiv:2202.07925 §3.3]. The code uses $\alpha = 0.25$ and $\gamma = 2.0$
[code: libs/modeling/losses.py#L8-L9]. Card: [ActionFormer: Localizing Moments of Actions with Transformers](/papers/2202.07925).

### Softmax focal classification with a class-split alpha (boundary-based localiser)

The anchor-free action localiser applies a softmax focal loss to both its
coarse and its refined classification heads. Both terms enter the total loss
with a localisation term and a quality term [arxiv:2103.13137 §3.4] [arxiv:2103.13137 eq. 11]
[arxiv:2103.13137 eq. 12]. The paper states neither $\alpha$ nor $\gamma$.
The code uses $\gamma = 2$, and an $\alpha$ of 0.25 on the background class
and 0.75 on the action classes [code: AFSD/thumos14/multisegment_loss.py#L32]
[code: AFSD/thumos14/multisegment_loss.py#L45-L51]. Note the placement: the
review's $\alpha$ sits on the foreground, so this 0.25 weights the opposite
class (see Contradictions). Card: [Learning Salient Boundary Feature for Anchor-free Temporal Action Localization](/papers/2103.13137).

### Focal confidence loss for 2D-3D matches

The pose-estimation paper applies a focal loss to match confidences. Each
match is scored against a ground-truth label, and the loss is
$-(1 - C')^{\gamma} \log C'$, with $C'$ the match confidence when the pair is a
ground-truth match and $1 - C$ otherwise [arxiv:2205.12257 §3.2]. The paper
does not state $\gamma$. The code uses $\gamma = 2$, $\alpha = 0.5$ and class
weights of 0.5 for positives and negatives [arxiv:2205.12257 §3.2] [code:
configs/experiment/train_GATsSPG.yaml#L52-L55]. Card: [OnePose: One-Shot Object Pose Estimation without CAD Models](/papers/2205.12257).

### Focal classification in set prediction

The crowd-localisation transformer trains its classification head with a focal
loss over all predictions, and adds an L1 point term with weight 2.5
[arxiv:2202.13065 §3.4]. The card states neither $\alpha$ nor $\gamma$. Its
ablations vary transformer size, matching cost and the number of neighbours,
not the loss [arxiv:2202.13065 tab. 7] [arxiv:2202.13065 tab. 8]
[arxiv:2202.13065 tab. 9]. Card: [An End-to-End Transformer Model for Crowd Localization](/papers/2202.13065).

### Focal confidence in an anchor-based 2D head

The monocular 3D detector uses a RetinaNet-style 2D head. Each anchor's
confidence is trained with a focal loss, where the target is 1 if the anchor's
reference box has an IoU above 0.5 with a ground-truth box, and 0 otherwise
[arxiv:1905.12365 §4.2] [arxiv:1905.12365 §7.2]. The card gives the loss only
as a weighted form with $\alpha$ and $\gamma$, and states neither value
[arxiv:1905.12365 §4.2]. The box regression on positive anchors uses a signed
IoU loss, not focal loss [arxiv:1905.12365 eq. 1]. Card: [Disentangling Monocular 3D Object Detection](/papers/1905.12365).

### Unified focal loss for continuous targets

The multi-view stereo paper extends focal loss to targets between 0 and 1.
The generalised form scales by the target, and the unified form divides the
error by the positive target and bounds the factor with a sigmoid-like
function with base $b$ [arxiv:2201.01501 §3.3] [arxiv:2201.01501 eq. 6]
[arxiv:2201.01501 eq. 9]. Positives carry no $\alpha$, and negatives use a
$\alpha_-$ that the paper sets per stage. $\gamma$ is set per stage to 2, 1 and
0, from coarse to fine [arxiv:2201.01501 §4]. Card: [Rethinking Depth Estimation for Multi-View Stereo: A Unified Representation](/papers/2201.01501).

### Focal classification in a DETR-style decoder

The Conditional DETR card trains its per-query classifier with focal loss,
inherited from Deformable DETR, with 300 queries
[arxiv:2108.06152 §3.5]. The loss is a sigmoid classification term. The
code's focal values are $\alpha = 0.25$ and $\gamma = 2$, and the classification
term has weight 2 [code: main.py#L86-L89]
[code: models/conditional_detr.py#L150]. The paper does not state the weights
[arxiv:2108.06152 §3.5]. Card: [Conditional DETR for Fast Training Convergence](/papers/2108.06152).

### Node-priority focusing parameter (scene graphs)

The scene-graph generator replaces the fixed focusing parameter with one per
node. A node's priority is the share of ground-truth triplets that contain it
[arxiv:2003.12962 eq. 6]. That share sets $\gamma$ through a mapping capped at
2, with an exponent $\mu$ that controls how strongly priority matters
[arxiv:2003.12962 eq. 7]. The per-node loss is then a focal-style term with this
$\gamma$ [arxiv:2003.12962 eq. 8]. The paper's best $\mu$ is 4 [arxiv:2003.12962 §4.4]
[arxiv:2003.12962 tab. 6]. In its ablation, the $\mu = 4$ setting sits 0.3
points above a plain focal-loss row at SGCLS R@20 and R@100 (36.1 and 40.1
against 35.8 and 39.8) [arxiv:2003.12962 tab. 6]. Card: [GPS-Net: Graph Property Sensing Network for Scene Graph Generation](/papers/2003.12962).

### Category-agnostic focal term in a vision-language tokenizer

The VisionLLM card uses a focal loss on the scores of object queries in its
tokenizer, with no class labels, plus an L1 term on centre points
[arxiv:2305.11175 app. B]. The card leaves the focal parameters unstated and
does not say how queries are matched to objects [arxiv:2305.11175 app. B].
Card: [VisionLLM: Large Language Model is also an Open-Ended Decoder for Vision-Centric Tasks](/papers/2305.11175).

### Generic multi-class form in a loss survey

The loss survey gives $-\alpha_t (1 - p_t)^{\gamma} \log p_t$ and describes
$\alpha_t$ as class-dependent, often inversely proportional to class frequency,
with no formula for it. It says $\gamma$ is commonly between 1 and 5, with no
source and no value per task [arxiv:2307.02694 §5.1.6]. The survey gives no
default for either. Use it for the definition, not for values. Card: [Loss Functions and Metrics in Deep Learning](/papers/2307.02694).

### Named but not specified

The YOLOv8 review lists a focal term for classification in its training loss.
It gives no formula, weight or $\gamma$ [arxiv:2408.15857 §3.4.1]. Do not
implement this as a specific variant from that card. Card: [What is YOLOv8: An In-Depth Exploration of the Internal Features of the Next-Generation Object Detector](/papers/2408.15857).

## When to use which

| Situation | Use | Why |
|---|---|---|
| Binary task, strongly imbalanced classes, classifier trained from scratch | Alpha-balanced focal loss, with $\alpha$ and $\gamma$ chosen on validation data | The COVID-19 card uses it on two imbalanced binary sets with an 80/20 split [arxiv:2101.07511 §3.2.1]. It gives no values to copy and no comparison with cross-entropy [arxiv:2101.07511 §3.2.2]. |
| Dense heatmap with one exact positive per object | Penalty-reduced focal loss, starting from $\gamma = 2$, $\beta = 4$ | The tracker card uses these values and its code matches [arxiv:2304.05170 §4.3]. The radar-camera code uses the same pair, though its paper states neither value [arxiv:2011.04841 §3.2]. |
| Continuous target in (0, 1], such as a proximity or quality score | Unified focal loss, instead of plain BCE | On DTU overall error falls from 0.327 mm (BCE) to 0.325 (GFL) to 0.320 (UFL). The gain is small, and the card gives no spread across runs [arxiv:2201.01501 tab. 3]. |
| Dense per-pixel classification over bins, most pixels background | Focal loss with $\gamma = 2$; for $\alpha$, follow the code's 0.25 with 13 / 1 weights if you need to match the reported numbers | The card reports $\gamma = 2.0$ [arxiv:2103.01100 §3.5]. Its ablation uses alpha 0.25 for its first three rows [arxiv:2103.01100 §4.3], and the paper's per-pixel alpha wording may behave differently [arxiv:2103.01100 §3.5]. |
| Semantic segmentation with a small foreground | Cross-entropy combined with an overlap loss such as Dice, or focal loss with $\alpha$ and $\gamma$ tuned on validation data | The review lists focal loss among the segmentation losses [arxiv:1910.07655 §4] and suggests pairing overlap losses with cross-entropy, since they behave badly on small objects [arxiv:1910.07655 §10.3]. It gives no benchmark that ranks focal loss against these choices. |
| Per-step classification in a single-pass temporal localiser | Focal loss, with $\alpha = 0.25$ and $\gamma = 2$ as in the code | The paper states neither value [arxiv:2202.07925 §3.3]. The code values are the only ones on the wiki [code: libs/modeling/losses.py#L8-L9], so say they come from the code. |
| Two-stage boundary-based action localiser with coarse and refined softmax classifiers | Softmax focal classification as in the anchor-free card, with $\gamma = 2$ and the code's class-split $\alpha$ | The paper states neither value [arxiv:2103.13137 §3.4]. The code sets $\alpha$ to 0.25 for background and 0.75 for actions [code: AFSD/thumos14/multisegment_loss.py#L45-L51], so say the values come from the code. The card gives no ablation of the focal term alone. |
| Scene-graph generation with a long-tailed predicate distribution | Node-priority focal loss with $\mu = 4$ as the paper's best, or plain focal loss as the baseline | The gain over the plain focal row is 0.3 points at SGCLS R@20 and R@100 [arxiv:2003.12962 tab. 6]. The card calls the released code's priority definition different from the paper's [arxiv:2003.12962 §3.2], so ablation numbers from the code do not test the paper's definition. |
| Set prediction over many outputs, most of them background | Focal classification with an L1 term, as in crowd localisation | The card uses it with $\lambda = 2.5$ [arxiv:2202.13065 §3.4], but no ablation isolates the focal term [arxiv:2202.13065 tab. 7]. |
| DETR-style decoder with learned queries | Focal classification with $\alpha = 0.25$, $\gamma = 2$, as in the code | The card's ablation adds focal loss and 300 queries in the same step, so its gain is not isolated: 34.9 to 35.3 AP for focal alone, and 36.9 with offset regression [arxiv:2108.06152 tab. 4]. Its GFLOPs rise with the query change [arxiv:2108.06152 tab. 4]. |
| Anchor-based one-stage 2D detector, most anchors background | Focal confidence on anchors, with $\alpha$ and $\gamma$ tuned on validation data | The card names the loss for the confidence but gives no $\alpha$ or $\gamma$ [arxiv:1905.12365 §4.2]. Its 2D ablation does not isolate the loss [arxiv:1905.12365 tab. 1]. |
| Classes roughly balanced, or no validation sweep is possible | Plain cross-entropy | None of the cards shows focal loss beating cross-entropy on a balanced task. The imbalanced COVID-19 card has no comparison against cross-entropy [arxiv:2101.07511 §3.2.2]. |
| Exact reproduction of a card's loss | Use the code's values where the paper is silent, and say so | The paper and code differ on $\alpha$ and class weights [arxiv:2205.12257 §3.2] [code: src/losses/focal_loss.py#L14-L24]. |
| Detector loss as in YOLOv8 | Not specified in the wiki | The card names the term but gives no formula [arxiv:2408.15857 §3.4.1]. |

## Implementation sketch

Two shapes cover most cards. The first is the alpha-balanced form for one
prediction, following the paper's equation 2. The second is the
penalty-reduced form on a heatmap. Batch averaging and the choice of
$\alpha_t$ are left to the caller, since the COVID-19 card does not specify
them. The survey's foreground and background weighting is a third form; use
its card only if you need that exact convention.

```pseudocode
# Alpha-balanced focal loss for one sample  [arxiv:2101.07511 eq. 2]
# p_t: probability the model gives the true class, in (0, 1)
# alpha_t: class weight, rule not stated in the paper  [§3.2.2]
# gamma: focusing parameter, gamma >= 0, value not stated  [§3.2.2]

def focal_loss(p_t, alpha_t, gamma):
    # Modulating factor: near 0 for easy examples, near 1 for hard ones  [eq. 2]
    modulator = (1 - p_t) ** gamma
    # Cross-entropy term, scaled by the modulator and the class weight  [eq. 2]
    return -alpha_t * modulator * log(p_t)

# Mean over a batch, as a training loss  [§3.2.2]
def batch_loss(p_t_list, alpha_t_list, gamma):
    losses = [focal_loss(p, a, gamma) for p, a in zip(p_t_list, alpha_t_list)]
    return sum(losses) / len(losses)
```

```pseudocode
# Penalty-reduced focal loss on a heatmap  [arxiv:2304.05170 eq. 3]
# h: target heatmap, exactly 1 at each centre; p: predicted heatmap in (0, 1)
# gamma = 2 and beta = 4 in the card  [§4.3]

def heatmap_focal_loss(p, h, gamma=2, beta=4):
    # Positives: pixels whose target is exactly 1  [eq. 3]
    pos = h == 1
    pos_term = (1 - p[pos]) ** gamma * log(p[pos])
    # Negatives: extra (1 - h)^beta factor shrinks pixels near a centre  [eq. 3]
    neg = ~pos
    neg_term = (1 - h[neg]) ** beta * p[neg] ** gamma * log(1 - p[neg])
    # Negated sum, as in eq. 3; normalisation is not shown there  [eq. 3]
    return -(pos_term.sum() + neg_term.sum())
```

## Contradictions

None between papers' reported numbers, because no two cards report values for
the same setting. Paper-versus-code gaps are noted in the variants above. The
pose-estimation card's code adds an $\alpha$ and class weights the paper does
not state [arxiv:2205.12257 §3.2]. The crowd card's code uses a focal-style
matching cost where the paper subtracts the predicted confidence
[arxiv:2202.13065 §3.3] [code: matcher.py#L66-L70]. The depth-estimation card
states a per-pixel alpha of 3.25 and 0.25, while its code uses one alpha of
0.25 with 13 and 1 multipliers. The card notes the effective factors match,
but the paper does not say which form produced its results
[arxiv:2103.01100 §3.5] [code: tools/cfgs/kitti_models/CaDDN.yaml#L58-L61]
[code: pcdet/models/backbones_3d/ffe/ddn_loss/ddn_loss.py#L40].

The scene-graph card has two gaps between paper and code. The paper's best
exponent is $\mu = 4$, but both released code paths use 5
[arxiv:2003.12962 §4.4] [code: lib/nps_loss.py#L23]. The paper does not say
whether the focusing weight is differentiated. Both code paths detach it, so
only the $-\log p$ term receives gradient [arxiv:2003.12962 eq. 8]
[code: lib/nps_loss.py#L16]. A standard focal-loss implementation does not
detach it, so the two can train differently.

The survey gives γ as commonly between 1 and 5, without a source
[arxiv:2307.02694 §5.1.6]. The cards that state γ use 2 [arxiv:2304.05170 §4.3]
[arxiv:2103.01100 §3.5] [arxiv:2108.06152 §3.5], which lies in that range. The
unified form sets γ per stage to 2, 1 and 0 [arxiv:2201.01501 §4], and its
stage value of 0 lies outside the survey's range. The survey does not say
where its range comes from.

The weighting conventions also differ. The COVID-19 card's $\alpha_t$ is a
class weight on the whole term [arxiv:2101.07511 eq. 2]. The survey's $\alpha$
weights the foreground term and $1 - \alpha$ weights the background term
[arxiv:1910.07655 §4.3]. The anchor-free localiser's code puts 0.25 on the
background and 0.75 on the action classes [code: AFSD/thumos14/multisegment_loss.py#L45-L51],
so its smaller value sits on the opposite class to the survey's convention.
The same number therefore means different things in the cards. The radar-camera
card's $\alpha$ is a heatmap exponent [arxiv:2011.04841 §3.2], a third meaning
of the same symbol.

## Papers

- [Disentangling Monocular 3D Object Detection](/papers/1905.12365): focal confidence on 2D anchors, with alpha and gamma not stated.
- [Deep Semantic Segmentation of Natural and Medical Images: A Review](/papers/1910.07655): survey of the binary focal form with foreground and background weights; no training values, plots use alpha 0.25 and gamma 2 as the review's own choice.
- [GPS-Net: Graph Property Sensing Network for Scene Graph Generation](/papers/2003.12962): node-priority focusing parameter in a scene-graph loss, with mu 4 as the paper's best and the code using 5.
- [CenterFusion: Center-based Radar and Camera Fusion for 3D Object Detection](/papers/2011.04841): penalty-reduced focal loss on a centre heatmap, with exponents not stated in the paper and 2 and 4 in the code.
- [Federated COVID-19 diagnosis from X-ray and ultrasound](/papers/2101.07511): alpha-balanced focal loss for imbalanced binary COVID-19 classification, with no reported alpha or gamma.
- [Categorical Depth Distribution Network for Monocular 3D Object Detection](/papers/2103.01100): focal loss per pixel over depth bins, with gamma 2 and foreground/background alpha from the paper.
- [Learning Salient Boundary Feature for Anchor-free Temporal Action Localization](/papers/2103.13137): softmax focal classification on coarse and refined heads, with alpha and gamma taken from the code.
- [ActionFormer: Localizing Moments of Actions with Transformers](/papers/2202.07925): focal classification over time steps, with alpha and gamma taken from the code.
- [Rethinking Depth Estimation for Multi-View Stereo: A Unified Representation](/papers/2201.01501): unified focal loss for continuous targets, with per-stage gamma and alpha in an ablation.
- [An End-to-End Transformer Model for Crowd Localization](/papers/2202.13065): focal classification over all set-prediction outputs, with no loss ablation.
- [OnePose: One-Shot Object Pose Estimation without CAD Models](/papers/2205.12257): focal loss on 2D-3D match confidences, with gamma from the code.
- [SportsMOT: A Large Multi-Object Tracking Dataset in Multiple Sports Scenes](/papers/2304.05170): penalty-reduced focal loss on a centre heatmap, with gamma 2 and beta 4.
- [What is YOLOv8: An In-Depth Exploration of the Internal Features of the Next-Generation Object Detector](/papers/2408.15857): names a focal classification term with no formula.
- [Conditional DETR for Fast Training Convergence](/papers/2108.06152): focal classification in a DETR-style decoder, with alpha 0.25 and gamma 2 from the code; the ablation changes the loss and query count together.
- [VisionLLM: Large Language Model is also an Open-Ended Decoder for Vision-Centric Tasks](/papers/2305.11175): category-agnostic focal loss on tokenizer query scores, with its parameters not stated.
- [Loss Functions and Metrics in Deep Learning](/papers/2307.02694): survey of the focal loss definition, with gamma commonly 1 to 5 and no default values.