docs/en/guides/custom-trainer.md
The Ultralytics training pipeline is built around BaseTrainer and task-specific trainers like DetectionTrainer. These classes handle the training loop, validation, checkpointing, and logging out of the box. When you need more control — tracking custom metrics, adjusting loss weighting, or implementing learning rate schedules — you can subclass the trainer and override specific methods.
This guide walks through seven common customizations:
!!! tip "Prerequisites"
Before reading this guide, make sure you're familiar with the basics of [training YOLO models](../modes/train.md) and the [Advanced Customization](../usage/engine.md) page, which covers the `BaseTrainer` architecture.
The YOLO model class accepts a trainer parameter in the train() method. This allows you to pass your own trainer class that extends the default behavior:
from ultralytics import YOLO
from ultralytics.models.yolo.detect import DetectionTrainer
class CustomTrainer(DetectionTrainer):
"""A custom trainer that extends DetectionTrainer with additional functionality."""
# Add your customizations here
model = YOLO("yolo26n.pt")
model.train(data="coco8.yaml", epochs=10, trainer=CustomTrainer)
Your custom trainer inherits all functionality from DetectionTrainer, so you only need to override the specific methods you want to customize.
The validation step computes precision, recall, and mAP. If you need additional metrics like per-class F1 score, override validate():
import numpy as np
from ultralytics import YOLO
from ultralytics.models.yolo.detect import DetectionTrainer
from ultralytics.utils import LOGGER
class MetricsTrainer(DetectionTrainer):
"""Custom trainer that computes and logs F1 score at the end of each epoch."""
def validate(self):
"""Run validation and compute per-class F1 scores."""
metrics, fitness = super().validate()
if metrics is None:
return metrics, fitness
if hasattr(self.validator, "metrics") and hasattr(self.validator.metrics, "box"):
box = self.validator.metrics.box
f1_per_class = box.f1
class_indices = box.ap_class_index
names = self.validator.names
mean_f1 = float(np.mean(f1_per_class)) if len(f1_per_class) else 0.0
LOGGER.info(f"Mean F1 Score: {mean_f1:.4f}")
per_class_str = [f"{names[i]}: {f1_per_class[j]:.3f}" for j, i in enumerate(class_indices)]
LOGGER.info(f"Per-class F1: {per_class_str}")
return metrics, fitness
model = YOLO("yolo26n.pt")
model.train(data="coco8.yaml", epochs=5, trainer=MetricsTrainer)
This logs the mean F1 score across all classes represented in validation and a per-class breakdown after each validation run.
!!! note "Available Metrics"
The validator provides access to many metrics through `self.validator.metrics.box`:
| Attribute | Description |
|---|---|
| `f1` | F1 score per class |
| `image_metrics` | Per-image metrics dictionary with precision, recall, F1, TP, FP, and FN |
| `p` | Precision per class |
| `r` | Recall per class |
| `ap50` | AP at IoU 0.5 per class |
| `ap` | AP at IoU 0.5:0.95 per class |
| `mp`, `mr` | Mean precision and recall |
| `map50`, `map` | Mean AP metrics |
Set cls_pw between 0.0 and 1.0 to apply normalized inverse-frequency weights to the classification loss. Override the existing weight computation only when you need hand-picked ratios:
import numpy as np
from ultralytics import YOLO
from ultralytics.models.yolo.detect import DetectionTrainer
class WeightedTrainer(DetectionTrainer):
"""Detection trainer with hand-picked class-weight ratios."""
def compute_class_weights(self, class_counts):
"""Return custom per-class weights for the production loss owner."""
weights = np.ones_like(class_counts)
weights[0] = 2.0
weights[1] = 3.0
return weights
model = YOLO("yolo26n.pt")
model.train(data="custom.yaml", epochs=10, cls_pw=1.0, trainer=WeightedTrainer)
set_class_weights() normalizes these values to a mean of 1.0 and stores them on the model, where the existing detection loss applies them. The indices above require a dataset with at least two classes.
The trainer saves best.pt based on fitness, which for detection defaults to [email protected]:0.95 (weights [0.0, 0.0, 0.0, 1.0] for [P, R, [email protected], [email protected]:0.95]). To use a different metric (like [email protected] or recall), override validate() and return your chosen metric as the fitness value. The built-in save_model() will then use it automatically:
from ultralytics import YOLO
from ultralytics.models.yolo.detect import DetectionTrainer
class CustomSaveTrainer(DetectionTrainer):
"""Trainer that saves the best model based on [email protected] instead of default fitness."""
def validate(self):
"""Override fitness to use [email protected] for best model selection."""
previous_best = self.best_fitness
metrics, fitness = super().validate()
if metrics is None:
return metrics, fitness
fitness = metrics["metrics/mAP50(B)"]
self.best_fitness = fitness if previous_best is None else max(previous_best, fitness)
return metrics, fitness
model = YOLO("yolo26n.pt")
model.train(data="coco8.yaml", epochs=20, trainer=CustomSaveTrainer)
BaseTrainer.validate() updates best_fitness using the default metric, so capture its previous value before calling it.
!!! note "Available Metrics"
Common metrics available in `self.metrics` after validation include:
| Key | Description |
|---|---|
| `metrics/precision(B)` | Precision |
| `metrics/recall(B)` | Recall |
| `metrics/mAP50(B)` | mAP at IoU 0.5 |
| `metrics/mAP50-95(B)` | mAP at IoU 0.5:0.95 |
Transfer learning workflows often benefit from freezing the pretrained backbone for the first N epochs, allowing the detection head to adapt before fine-tuning the entire network. Ultralytics provides a freeze parameter to freeze layers at the start of training, and you can use a callback to unfreeze them after N epochs:
from ultralytics import YOLO
from ultralytics.models.yolo.detect import DetectionTrainer
from ultralytics.utils import LOGGER
FREEZE_EPOCHS = 5
def unfreeze_backbone(trainer):
"""Callback to unfreeze the user-requested layers after FREEZE_EPOCHS."""
if trainer.epoch == FREEZE_EPOCHS:
user_freeze = [x for x in trainer.freeze_layer_names if x not in {".dfl", "teacher_model."}]
LOGGER.info(f"Epoch {trainer.epoch}: Unfreezing requested layers for fine-tuning")
for name, param in trainer.model.named_parameters():
if (
not param.requires_grad
and ".dfl" not in name
and "teacher_model." not in name
and any(x in name for x in user_freeze)
):
param.requires_grad = True
LOGGER.info(f" Unfroze: {name}")
trainer.freeze_layer_names = [x for x in trainer.freeze_layer_names if x not in user_freeze]
class FreezingTrainer(DetectionTrainer):
"""Trainer with backbone freezing for first N epochs."""
def __init__(self, *args, **kwargs):
"""Initialize and register the unfreeze callback."""
super().__init__(*args, **kwargs)
self.add_callback("on_train_epoch_start", unfreeze_backbone)
model = YOLO("yolo26n.pt")
model.train(data="coco8.yaml", epochs=20, freeze=10, trainer=FreezingTrainer)
The freeze=10 parameter freezes the first 10 layers (indices 0-9) at training start, which covers most of the YOLO26 backbone. The backbone spans layers 0-10, so freeze=10 leaves the final C2PSA block (layer 10) trainable; use freeze=11 to freeze the entire backbone. The on_train_epoch_start callback fires at the beginning of each epoch and unfreezes those requested layers once the freeze period is complete, while preserving permanently frozen DFL and distillation-teacher parameters.
!!! tip "Choosing What to Freeze"
- `freeze=10` freezes the first 10 layers, indices 0-9 (most of the YOLO26 backbone; use `freeze=11` to include the final C2PSA block at layer 10)
- `freeze=[0, 1, 2, 3]` freezes specific layers by index
- Higher `FREEZE_EPOCHS` values give the head more time to adapt before the backbone changes
Different parts of the network can benefit from different learning rates. A common strategy is to use a lower learning rate for the pretrained backbone to preserve learned features, while allowing the detection head to adapt more quickly with a higher rate:
from ultralytics import YOLO
from ultralytics.models.yolo.detect import DetectionTrainer
from ultralytics.utils import LOGGER
from ultralytics.utils.torch_utils import unwrap_model
class PerLayerLRTrainer(DetectionTrainer):
"""Trainer with different learning rates for backbone and head."""
backbone_lr_ratio = 0.1
def build_optimizer(self, model, name="auto", lr=0.001, momentum=0.9, decay=1e-5, iterations=1e5):
"""Reuse the trainer optimizer and lower its backbone parameter-group rates."""
optimizer = super().build_optimizer(model, name, lr, momentum, decay, iterations)
unwrapped = unwrap_model(model)
backbone_len = len(unwrapped.yaml["backbone"])
backbone = {
id(p)
for name, p in unwrapped.named_parameters()
if any(name.startswith(f"model.{i}.") for i in range(backbone_len))
}
groups = []
for group in optimizer.param_groups:
head_params = [p for p in group["params"] if id(p) not in backbone]
backbone_params = [p for p in group["params"] if id(p) in backbone]
if head_params:
groups.append({**group, "params": head_params})
if backbone_params:
groups.append({**group, "params": backbone_params, "lr": group["lr"] * self.backbone_lr_ratio})
optimizer.param_groups = groups
LOGGER.info(f"PerLayerLR: {len(backbone)} backbone params at {self.backbone_lr_ratio}x the head rate")
return optimizer
model = YOLO("yolo26n.pt")
model.train(data="coco8.yaml", epochs=20, trainer=PerLayerLRTrainer)
For RT-DETR, use the same override with RTDETRTrainer as the parent and load the checkpoint with RTDETR("rtdetr-l.pt").
When training on multiple GPUs with DistributedDataParallel, the default BatchNorm2d layers compute statistics independently on each GPU. For RT-DETR fine-tuning and other recipes that use small per-GPU batch sizes, per-GPU batch statistics can be noisy. PyTorch's SyncBatchNorm synchronizes mean and variance across all ranks for a single global batch statistic, which often improves convergence at the cost of a small inter-GPU communication overhead.
The conversion has to happen after the model is on the GPU but before DDP wraps it. The cleanest hook for this is set_model_attributes(), which BaseTrainer calls in exactly that window:
from torch import nn
from ultralytics import RTDETR
from ultralytics.models.rtdetr.train import RTDETRTrainer
class SyncBNTrainer(RTDETRTrainer):
"""RT-DETR trainer that converts BatchNorm to SyncBatchNorm for multi-GPU training."""
def set_model_attributes(self):
"""Run the parent setup, then convert BN to SyncBatchNorm when training on multiple GPUs."""
super().set_model_attributes()
if self.world_size > 1:
self.model = nn.SyncBatchNorm.convert_sync_batchnorm(self.model)
model = RTDETR("rtdetr-l.pt")
model.train(data="coco8.yaml", epochs=20, device=[0, 1], trainer=SyncBNTrainer)
The world_size > 1 guard ensures the trainer is safe to use in single-GPU runs as well; on a single GPU the conversion is skipped and training proceeds with regular BatchNorm2d. The same pattern works for YOLO by switching the parent class to DetectionTrainer.
!!! tip "When to use SyncBatchNorm"
| Scenario | Recommendation |
| ---------------------------------------------- | ------------------------ |
| Multi-GPU training, small per-GPU batch (≤ 16) | Enable |
| Multi-GPU training, large per-GPU batch (≥ 32) | Optional; minor benefit |
| Single-GPU training | Not applicable (skipped) |
The default trainer clips gradients to max_norm=10.0 in optimizer_step(), a loose value tuned for YOLO models where gradients rarely exceed it. DETR-family detectors (RT-DETR, DEIM, DINO) typically use much tighter values such as 0.1 to stabilize the decoder's cross-attention layers, where gradient magnitudes can spike. To override the clip value, subclass the trainer and override optimizer_step():
import torch
from ultralytics import RTDETR
from ultralytics.models.rtdetr.train import RTDETRTrainer
from ultralytics.utils.torch_utils import TORCH_2_0
class CustomClipTrainer(RTDETRTrainer):
"""RT-DETR trainer with configurable gradient clipping."""
clip_grad_norm = 0.1 # max gradient norm; set to 0 to disable clipping
def optimizer_step(self):
"""Run an optimizer step with a configurable gradient-norm clip."""
self.scaler.unscale_(self.optimizer)
if self.clip_grad_norm > 0:
kwargs = {"foreach": False} if self.device.type == "npu" and TORCH_2_0 else {}
torch.nn.utils.clip_grad_norm_(self.model.parameters(), max_norm=self.clip_grad_norm, **kwargs)
self.scaler.step(self.optimizer)
self.scaler.update()
self.optimizer.zero_grad()
if self.ema:
self.ema.update(self.model)
model = RTDETR("rtdetr-l.pt")
model.train(data="coco8.yaml", epochs=20, trainer=CustomClipTrainer)
The same trainer works for YOLO by switching the parent class to DetectionTrainer (from ultralytics.models.yolo.detect import DetectionTrainer) and loading a YOLO checkpoint with YOLO("yolo26n.pt"). The optimizer_step body is unchanged.
!!! tip "Typical clip_grad_norm values"
| Architecture family | Typical `max_norm` |
| ---------------------------- | ------------------ |
| RT-DETR / DEIM / DETR family | `0.1` |
| YOLO (Ultralytics default) | `10.0` |
| Disable clipping | `0` |
Pass your custom trainer class (not an instance) to the trainer parameter in model.train():
from ultralytics import YOLO
from ultralytics.models.yolo.detect import DetectionTrainer
class MyCustomTrainer(DetectionTrainer):
"""A custom trainer that extends DetectionTrainer."""
model = YOLO("yolo26n.pt")
model.train(data="coco8.yaml", trainer=MyCustomTrainer)
The YOLO class handles trainer instantiation internally. See the Advanced Customization page for more details on the trainer architecture.
Key methods available for customization:
| Method | Purpose |
|---|---|
validate() | Run validation and return metrics |
build_optimizer() | Construct the optimizer |
save_model() | Save training checkpoints |
get_model() | Return the model instance |
get_validator() | Return the validator instance |
get_dataloader() | Build the dataloader |
preprocess_batch() | Preprocess input batch |
label_loss_items() | Format loss items for logging |
For the full API reference, see the BaseTrainer documentation.
Yes, for simpler customizations, callbacks are often sufficient. Available callback events include on_train_start, on_train_epoch_start, on_train_epoch_end, on_fit_epoch_end, and on_model_save. These allow you to hook into the training loop without subclassing. The backbone freezing example above demonstrates this approach.
If your change is simpler (such as adjusting loss gains), you can modify the hyperparameters directly:
from ultralytics import YOLO
model = YOLO("yolo26n.pt")
model.train(data="coco8.yaml", box=10.0, cls=1.5, dfl=2.0)
On YOLO26, dfl scales the logged l1_loss because its detection head uses reg_max: 1; on YOLOv8 and YOLO11 it scales dfl_loss.