Platform For AI:Image metric learning training (raw)

Prerequisites

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

Limitations

Requires DLC computing engine.

Supported models

Supported backbones: resnet50, resnet18, resnet34, resnet101, swint_tiny, swint_small, swint_base, vit_tiny, vit_small, vit_base, xcit_tiny, xcit_small, and xcit_base.

Configuration

• Input ports

  Port (left to right)	Data type	Recommended upstream component	Required
  Training annotation file	OSS	Read OSS Data	No
  Evaluation annotation file	OSS	Read OSS Data	No

• Parameters

  Tab	Parameter	Required	Description	Default value
  Field Settings	Metric learning model type	Yes	Algorithm type for training. Valid values: Data-parallel metric learning Model-parallel metric learning	Data-parallel metric learning
  	OSS directory for training output	Yes	OSS directory where trained models are stored. Example: oss://examplebucket/yun****/designer_test	None
  	Training annotation file path	No	Configure if no training annotation file is connected to the input port. Note If configured in both places, input port takes precedence. OSS path where the training annotation file is stored. Example: oss://examplebucket/yourfolder****/data/imagenet/meta/train_labeled.txt Each line in train_labeled.txt must follow this format: absolute path/image_name.jpg label_id Important Separate image path and label_id with a space.	None
  	Validation annotation file path	No	Configure if no evaluation annotation file is connected to the input port. Note If configured in both places, input port takes precedence. OSS path where the validation annotation file is stored. Example: oss://examplebucket/yourfolder****/data/imagenet/meta/val_labeled.txt Each line in val_labeled.txt must follow this format: absolute path/image_name.jpg label_id Important Separate image path and label_id with a space.	None
  	File of class name list	No	Enter class names directly or specify OSS path to a file containing class names.	None
  	Data source format	Yes	Input data format. Valid values: ClsSourceImageList and ClsSourceItag.	ClsSourceImageList
  	OSS path of the pre-trained model	No	OSS path to a custom pre-trained model. If not set, PAI uses the default pre-trained model.	None
  Parameter Settings	Metric learning backbone	Yes	Backbone model for metric learning. Valid values: resnet_50 resnet_18 resnet_34 resnet_101 swin_transformer_tiny swin_transformer_small swin_transformer_base	resnet50
  	Image resize size	Yes	Image size in pixels after resizing.	224
  	Backbone output feature dimension	Yes	Output feature dimension of the backbone. Must be an integer.	2048
  	Feature output dimension	Yes	Output feature dimension of the Neck. Must be an integer.	1536
  	Number of training classes	Yes	Number of output dimensions for metric learning.	None
  	Loss function	Yes	Loss function that evaluates the difference between predicted and actual values. Valid values: AMSoftmax (recommended parameters: margin 0.4, scale 30) ArcFaceLoss (recommended parameters: margin 28.6, scale 64) CosFaceLoss (recommended parameters: margin 0.35, scale 64) LargeMarginSoftmaxLoss (recommended parameters: margin 4, scale 1) SphereFaceLoss (recommended parameters: margin 4, scale 1) Model-parallel AMSoftmax, whose classification limit can be extended with the number of GPUs Model-parallel Softmax, whose classification limit can be extended with the number of GPUs	AMSoftmax (recommended parameters: margin 0.4, scale 30)
  	Loss function scale parameter	Yes	Set based on the selected loss function.	30
  	Loss function margin parameter	Yes	Set based on the selected loss function.	0.4
  	Loss function weight	No	Loss function weight. Balances optimization between metric and classification.	1.0
  	Optimizer	Yes	Optimizer for training. Valid values: SGD AdamW	SGD
  	Initial learning rate	Yes	Initial learning rate for training. Must be a floating-point number.	0.03
  	Training batch_size	Yes	Number of samples trained per iteration.	None
  	Total training epochs	Yes	Total number of training rounds on all samples.	200
  	Checkpoint saving frequency	No	Frequency to save model checkpoints. A value of 1 saves after each epoch.	10
  Execution Tuning	Training data read threads	No	Number of processes for reading training data.	4
  	Enable half-precision	No	Enables half-precision for training to reduce memory usage.	None
  	Computing mode	Yes	Computing engine to run this component. Supported engines: Standalone DLC Distributed DLC	Standalone DLC
  	Workers	No	Configure when using Distributed DLC. Number of concurrent worker processes during training.	1
  	GPU model	Yes	GPU specification for training.	8vCPU+60GB Mem+1xp100-ecs.gn5-c8g1.2xlarge

Example