Training and evaluation

Training

The main function to train a yolo model can be accessed either through the API or on CLI. To start a training, you can launch the following command:

yolo train classify|detect|obb|segment model=<path_to_model.pt> ...

Here is an explanation of the different key parameters:

Parameter Description Type Default

Parameter	Description	Type	Default
`model`	Specifies the model file for training. Accepts a path to either a `.pt` pretrained model or a `.yaml` configuration file.	`pathlib.Path`
`data`	Path to the `data.yml` file. file	`pathlib.Path`
`epochs`	Total epochs to train. on.	`int`	`100`
`time`	Early stopping based on training time.	`int`
`patience`	Early stopping based on consecutive epochs without improvement on validation loss.	`int`	`100`
`batch`	Batch size.	`int` or `float`	`16`
`imgsz`	Image size (in pixels) to resize the training images.	`int`	`640`
`device`	On which GPU to run training. You can set multiple GPU usage with `gpu_id1,gpu_id2,...`	`int`or `list(int)`
`project`	Useful to keep track of exactly where your training logs are created. By default, will create a `runs` folder.	`pathlib.Path`
`name`	Name of run folder in the log folder.	`str`
`single_cls`	If you want to train a class agnostic detector and do not care anymore of the distinction between classes.	`bool`	`False`
`dfl`	Distribution focal loss: weight of one of the losses of the Total loss.	`float`	`1.5`

model

Specifies the model file for training. Accepts a path to either a .pt pretrained model or a .yaml configuration file.

pathlib.Path

data

Path to the data.yml file. file

pathlib.Path

epochs

Total epochs to train. on.

int

100

time

Early stopping based on training time.

int

patience

Early stopping based on consecutive epochs without improvement on validation loss.

int

100

batch

Batch size.

int or float

16

imgsz

Image size (in pixels) to resize the training images.

int

640

device

On which GPU to run training. You can set multiple GPU usage with gpu_id1,gpu_id2,...

intor list(int)

project

Useful to keep track of exactly where your training logs are created. By default, will create a runs folder.

pathlib.Path

name

Name of run folder in the log folder.

str

single_cls

If you want to train a class agnostic detector and do not care anymore of the distinction between classes.

bool

False

dfl

Distribution focal loss: weight of one of the losses of the Total loss.

float

1.5

More details:

Using batch size=-1 will automatically use a batch size for 60% of GPU Memory usage. It can be a fraction (ex: 0.70) to use a fraction of GPU memory.
The resizing conserves aspect ratio: the longest side is set to the target size then padding is added (grey padding) to complete the square image. Must be a multiple of 32 or 64.
By increasing the dlf parameter, one can augment the importance given to classes on which the model struggles the most to classify correctly. Useful for underrepresented or easily confused classes.
For more parameters, please refer to the Ultralytics training documentation.

Evaluation

To start an evaluation, one can launch the following command:

yolo val model=<path_to_model.pt> ...

If only a model is given in the command, the model outputs the internal evaluation done on training and validation data stored in the model checkpoint.
The evaluation function outputs confusion matrices and images of some batches on which it ran its evaluation to qualitatively evaluate the model.

To have more flexibility in the evaluation, it can be interesting to build dedicated functions to evaluate with your specific metrics.
The predict() function in python outputs the results as a dictionary.
There are libraries, like fiftyone for example, that can take a trained model and evaluate on given data.