What is the Object detection (easycv) component - Platform For AI

Train object detection models using YOLOX, FCOS, YOLOv5, and YOLOv7 to identify and locate entities in images.

Prerequisites

Before you begin, ensure that you have:

Activated OSS. See Activate OSS.
Authorized Machine Learning Studio to access OSS. See Grant required permissions to use Machine Learning Designer.

Limitations

Available only in Designer.
Only the DLC engine is 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

Field settings tab

Parameter	Required	Default	Description
Model type	Yes	YOLOX	Model architecture for training. Valid values: FCOS, YOLOX, YOLOv5, YOLOv7.
Data format	Yes	DetSourcePAI	Annotation format for your dataset. DetSourcePAI uses `.manifest` files (produced by PAI iTAG); COCO uses `.json` files (standard COCO annotation format). YOLOv5 and YOLOv7 support DetSourcePAI only.
Annotation file path for training set	No	None	Path to the training annotation file. Use a `.manifest` file for DetSourcePAI or a `.json` file for COCO. If training data is connected through the input port, leave this blank — the input port takes precedence.
Annotation file path for validation set	No	None	Path to the validation annotation file. Use a `.manifest` file for DetSourcePAI or a `.json` file for COCO. If validation data is connected through the input port, leave this blank — the input port takes precedence.
Training data OSS path	Yes (COCO only)	None	Path to training images in COCO format. Appears only when Data format is set to COCO.
Validation data OSS path	Yes (COCO only)	None	Path to evaluation images in COCO format. Appears only when Data format is set to COCO.
Class list file OSS path	No	None	Path to a label list file with a `.txt` extension. If a class list file is connected through the input port, leave this blank — the input port takes precedence.
Pre-trained model OSS path	No	None	OSS path to a custom pre-trained model. If not specified, PAI uses a default pre-trained model.
OSS directory for training	No	None	OSS directory where the trained model is saved. Example: `examplebucket.oss-cn-shanghai-internal.aliyuncs.com/test/ckpt/`. If left blank, the default workspace path is used.

Parameters tab

Parameter	Required	Default	Description
YOLOX model structure	Yes (YOLOX only)	yolox-s	Architecture variant for YOLOX. Valid values: yolox-s, yolox-m, yolox-l, yolox-x.
Number of classes	Yes	20	Number of object classes in your dataset.
Image scale	Yes	320 320	Image dimensions after resizing (height width). Example: `320 320`.
Optimizer	Yes (YOLOX only)	momentum	Optimizer for model training. Valid values: momentum, adam.
Initial learning rate	Yes	0.01	Learning rate at the start of training.
Training batch size	Yes	8	Number of samples per training iteration.
Validation batch size	Yes	8	Number of samples per evaluation iteration.
Number of epochs	Yes	20	Total training epochs. For YOLOX, this value must exceed the sum of Warmup epochs and Last no augmented lr epochs.
Log loss frequency	No	200	Frequency (in batches) for logging the training loss.
Validation frequency	No	2	Validation interval in epochs. A value of 2 runs validation every two epochs.
Warmup epochs	No	5	Applies to YOLOX only. Number of warm-up epochs at the start of training.
Last no augmented lr epochs	No	5	Applies to YOLOX only. Number of final epochs run without data augmentation.
Export model format	Yes	raw	Format of the exported model. JIT is not supported for YOLOv5 and YOLOv7. Valid values: raw, jit, onnx.
Checkpoint saving frequency	No	1	Frequency (in epochs) for saving model checkpoints. A value of 1 saves a checkpoint after every epoch.

Tuning tab

Parameter	Required	Default	Description
GPU machine type	Yes	4	GPU instance type for training. A GPU instance is required.
Enable FP16	No	false	Enables FP16 half-precision training to reduce memory usage and speed up training on supported GPUs.
Execution mode	Yes	Distributed DLC	Distributed training on DLC. Only distributed mode is supported.

Output

Output port	Type	Downstream component
Output model	SavedModel format, stored at the OSS path specified in OSS directory for training on the Field settings tab	image prediction

Build an object detection pipeline

Build an end-to-end object detection pipeline using FCOS, YOLOX, YOLOv5, or YOLOv7.

Step 1: Label your images

Use the iTAG module in PAI to label your training images.

Create a dataset and a labeling job. See Create a dataset: for data labeling and Create a labeling job.
When creating the labeling job, select the General Template template under Object Detection.
Under Object Detection Label Configuration, select Single-label. In the Label Configuration section, enter all your label names.
After labeling is complete, export the results in manifest format to OSS.

Step 2: Read the labeled data

Add a Read OSS data component to the pipeline. Set OSS Data Path to the OSS path of the exported manifest file. For example: oss://examplebucket.oss-cn-shanghai.aliyuncs.com/ev_demo/xxx.manifest.

Step 3: Configure the training component

Connect the Training data and Validation data input ports of the Object detection (easycv) component to the Read OSS data component, then configure the parameters as described in Configuration.

Step 4: Run inference (FCOS and YOLOX)

For FCOS and YOLOX models, add an image prediction component downstream to run offline inference. See image prediction.

What's next

Connect an image prediction component downstream to run predictions on new images and evaluate model performance after training. See image prediction.
Browse all available algorithm components in Designer. See Designer component reference.
For an overview of Designer, see Designer overview.