Classic image classification 2.5.0¶

Classify the entire image. Presets currently include ResNet50, ResNet18, ResNet101, MobileNetV3_small_x0_35 and MobileNetV3_large_x1_0; entries appear only when their bundled configurations exist.

Select one classification dataset folder. train_list.txt, val_list.txt and Label.txt take priority; without lists, train/class_name and val/class_name are recognized. Existing training/validation sets are retained without automatic resplitting. Verify class IDs and name order; training, inference and export must agree.

Size and optimization¶

The input is a square side length; 0 means Use configuration. A custom value updates training/validation crops, model input and export preprocessing. The Resize short edge follows the original classification ratio 256/224 rather than replacing all preprocessing with an uncropped stretch.

Match the input size to the details needed by your application. This page has neither separate height/width controls nor the classic detection validation-only size mode. Learning rate, optimizer, weight decay, warmup, AMP and workers follow the classic common parameters, but map to the classification engine's configuration structure.

Metrics¶

Displays training and validation accuracy, loss and other metrics actually emitted by the classification engine. These must not be interpreted as detection Precision/Recall/mAP. Available metrics vary by model configuration; no curves are invented for missing values.

Automatic export and deployment¶

After training, automatic export writes to the experiment's export directory. Typical files are:

inference.pdmodel
inference.pdiparams
inference.pdiparams.info
ImageSize.txt
inference_cls.yaml
class_names.txt

ImageSize.txt, class mapping and inference_cls.yaml are generated from this run's effective configuration. Customers need not replace generic ImageNet settings manually. The deployment YAML uses relative paths: use the export folder as the working directory and configure the deployment program's device for the actual hardware.

Model inference validation uses the training checkpoint .pdparams and training configuration, rather than accepting an exported .pdmodel directly as this page's weight input. They serve different purposes.

See the classification parameter table for all controls.