Parameters for Basic Detection Training and Model Evaluation

--scale

int

1~10

4

Anchor scale.

--scales_per_octave

int

1~4

2

Number of feature scales at each layer.

--aspect_ratios

String

The value is a float ranging from 0.05 to 20. The number of values cannot exceed 10.

"1.0, 2.0, 0.5"

Anchor aspect ratio.

--color_adjust

bool

True or False

True

Whether to use random colors.

--random_crop

bool

True or False

True

Whether to enable random cropping.

--random_flip

bool

True or False

True

Whether to enable random flipping.

--train_dataset_path

String

-

None

Path of the training dataset.

--epoch_size

int

[1, 10000]

40

Number of training epochs.

--batch_size

int

Less than or equal to the number of images in the dataset.

[1, 512]

2

Batch size for training.

--loss_scale

int

[1, 2048]

1024

Loss scale.

--input_width

String

A multiple of 128 within [1,10000)

"2048"

Model input width.

--input_height

String

A multiple of 128 within [1,10000)

"1536"

Model input height.

--init_lr

float

(0, 1)

0.001

Training learning rate.

--weight_decay

float

(0, 1)

0.00004

Weight decay.

--quick_train

bool

True or False

False

Fast training switch.

--pretrained_ckpt_path

String

-

"pre_trained_ckpt"

Path for storing the pre-trained model.

--device_id

int

[0,7]

0

ID of NPU used for training.

--device_num

int

Currently, only one device is supported, that is, the value can only be 1.

1

Number of NPUs used for training.

--filter_weight

bool

True or False

True

Whether to filter the head network weight parameter.

--save_checkpoint_epochs

int

-

10

Interval (by epoch) for saving the training CKPT.

--train_output_path

String

-

"train_output_path"

Training output path.

--save_best_ckpt

bool

True or False

True

Whether to save the CKPT with the best accuracy when training and validation are performed simultaneously. (This parameter will be deleted in later versions.) Currently, when run_eval is set to True, the CKPT with the best accuracy is saved by default.

--run_eval

bool

True or False

True

Whether to enable the evaluation function during training.

--eval_start_epoch

int

-

5

Sequence umber of epochs from which evaluation is enabled when training and evaluation are performed simultaneously.

--eval_interval

int

[1, 10000]

5

Model evaluation interval (by epoch) when training and evaluation are performed simultaneously.

--enable_modelarts

bool

True or False

False

Whether to use ModelArts for training.

--data_url

String

-

None

Dataset path for the ModelArts training.

--train_output_url

String

-

None

Output path for the ModelArts training.

--checkpoint_url

String

-

None

Path of the pre-trained model for the ModelArts training.

--device_id

int

[0,7]

0

ID of the NPU for evaluation.

--eval_ckpt_path

String

-

None

Path for obtaining the evaluation CKPT. You need to specify the training output path.

--eval_output_path

String

-

"eval_output_path"

Output path of the evaluation results.

--min_score

float

-

0.4

Confidence threshold.

--nms_threshold

float

-

0.2

IoU threshold during NMS processing.

--eval_dataset_path

String

-

None

Path of the evaluation dataset.

--enable_modelarts

bool

True or False

False

Whether to use ModelArts for training.

--data_url

String

-

None

Dataset path for evaluation on ModelArts.

--eval_output_url

String

-

None

Output path for evaluation on ModelArts.

--checkpoint_url

String

-

None

Path of the pre-trained model during evaluation on ModelArts.

--pretrained_ckpt_path

String

-

"pre_trained_ckpt"

Path of the pre-trained model.