Platform For AI:Object detection (easycv)

Prerequisites

OSS is activated, and Machine Learning Studio is authorized to access OSS. For more information, see Activate OSS and Grant required permissions to use Machine Learning Designer.

Limitations

• Available only in Designer.

• Only DLC engine supported.

Parameters

• Input

  Input port	Type	Upstream component	Required
  Training data	OSS	Read OSS data	No If not connected, configure Annotation file path for training set on the Field settings tab.
  Validation data	OSS	Read OSS data	No If not connected, configure Annotation file path for validation set on the Field settings tab.
  Class list file	OSS	Read OSS data	No If not connected, configure Class list file OSS path on the Field settings tab.

• Configuration

  Tab	Parameter	Required	Description	Default
  Field settings	Model type	Yes	Model type to use for training. Valid values: FCOS YOLOX YOLOv5 YOLOv7	YOLOX
  	OSS directory for training	No	OSS directory to store the trained model. Example: examplebucket.oss-cn-shanghai-internal.aliyuncs.com/test/ckpt/. If empty, uses default workspace path.	None
  	Annotation file path for training set	No	If data format is DetSourcePAI, provide a file with a .manifest extension. If data format is COCO, provide a file with a .json extension. If training data provided through an input port, leave this parameter empty. If both configured, the input port takes precedence.	None
  	Annotation file path for validation set	No	If data format is DetSourcePAI, provide a file with a .manifest extension. If data format is COCO, provide a file with a .json extension. If validation data provided through an input port, leave this parameter empty. If both configured, the input port takes precedence.	None
  	Class list file OSS path	No	Specify a label list file with a .txt extension. If class list file provided through an input port, leave this parameter empty. If both configured, the input port takes precedence.	None
  	Pre-trained model OSS path	No	To use a custom pre-trained model, set this parameter to its OSS path. If not specified, PAI uses a default pre-trained model.	None
  	Training data OSS path	Yes	Appears only when data format is COCO. Specify the path to training images in COCO format.	None
  	Validation data OSS path	Yes	Appears only when data format is COCO. Specify the path to evaluation images in COCO format.	None
  	Data format	Yes	Valid values: COCO DetSourcePAI (YOLOv5 and YOLOv7 models support only DetSourcePAI format.)	DetSourcePAI
  Parameters	YOLOX model structure	Yes	Appears only when model type is YOLOX. Select the desired model structure from the drop-down list. yolox-s yolox-m yolox-l yolox-x	yolox-s
  	Number of classes	Yes	Class labels in the dataset.	20
  	Image scale	Yes	Size of the image after resizing. Separate height and width with a space. Example: 320 320.	320 320
  	Optimizer	Yes	Configurable only when model type is YOLOX. Optimizer for model training. Valid values: momentum adam	momentum
  	Initial learning rate	Yes	Initial learning rate.	0.01
  	Training batch size	Yes	Samples used in a single training iteration.	8
  	Validation batch size	Yes	Samples used in a single evaluation iteration.	8
  	Number of epochs	Yes	Training epochs to run. Note For YOLOX models, total epochs must exceed the sum of warmup epochs and Last no augmented lr epochs.	20
  	Log loss frequency	No	Frequency, in training batches, to log the loss. A value of 200 logs loss every 200 batches.	200
  	Validation frequency	No	Validation interval in epochs. A value of 2 runs validation every two epochs.	2
  	Warmup epochs	No	Applies only to YOLOX models.	5
  	Last no augmented lr epochs	No	Applies only to YOLOX models.	5
  	Export model format	Yes	Format of the exported model. Valid values: raw jit (YOLOv5 and YOLOv7 models do not support this format.) onnx	raw
  	Checkpoint saving frequency	No	Frequency, in epochs, for saving model checkpoints. A value of 1 saves a checkpoint after every epoch.	1
  Tuning	GPU machine type	Yes	Requires a GPU instance to run.	4
  	Enable FP16	No	Enables FP16 half-precision training.	false
  	Execution mode	Yes	Supports only distributed training on DLC.	Distributed DLC

• Output

  Output port	Type	Downstream component
  Output model	OSS path specified for OSS directory for training on the Field settings tab. Stores the trained model in SavedModel format.	image prediction

Example

For FCOS, YOLOX, YOLOv5, and YOLOv7 models, use the Object detection (easycv) component to build a pipeline as shown below.

Configure the components by following these steps:

1. Use the iTAG module in PAI to label your images. For more information, see Create a dataset: for data labeling and Create a labeling job.

   1. After creating a dataset, select the General Template template under Object Detection when creating the labeling job.

   2. 

   3. For Object Detection Label Configuration, select Single-label. Under Label Configuration, enter all your label names to simplify labeling.

   4. 

   5. After completing labeling, export the results in manifest format to OSS for the component to use.

   6. 

2. Use the Read OSS data component to read the labeled result file xxx.manifest, and set OSS Data Path to the OSS path of the labeled result dataset. For example, oss://examplebucket.oss-cn-shanghai.aliyuncs.com/ev_demo/xxx.manifest.

3. Connect training data and validation data to the Object detection (easycv) component and configure its parameters. For details, see Configure component parameters.

For FCOS and YOLOX models, perform offline inference using the image prediction component. For more information, see image prediction.

Related documents

• After training an object detection model, connect an image prediction component downstream of the Object detection (easycv) component to run predictions and evaluate model performance. For more information, see image prediction.

• For more information about Designer components, see Designer overview.

• Designer offers a variety of algorithm components. Select the component appropriate for your data processing scenario. For more information, see Designer component reference.