Model Training - Imagery Analysis
Feature Description
Model training aims to use generated training data for neural network model training. The complete training procedure obtains optimized network models through multiple iterations (epochs). Hyperparameter tuning (learning rate, batch size, etc.) can enhance training efficiency and accuracy to derive usable neural network models.
Additionally, fine-tuning pretrained models with limited sample data can better adapt models to specific application scenarios while reducing training time.
Parameter Description
Parameter | Default | Description | Type |
---|---|---|---|
Training data path | Specifies the path of the generated training data | String | |
Model purpose | Object detection | Specifies the application purpose of the training data | Object |
Model algorithm | Select appropriate model algorithm based on application purpose | String | |
Training config file | Specifies the path of the training configuration file | String | |
Training times | 10 | Number of training epochs. Default: 10. Increased epochs improve model fitting but may cause overfitting. Maximum allowed: 300 epochs. Training duration is proportional to epoch count. | Integer |
Batch size | 1 | Number of images processed per training step. Higher batch sizes increase memory consumption but reduce training time within reasonable limits. | Integer |
Learning rate | Parameter update magnitude. Leave blank for automatic adjustment based on batch size and built-in baseline settings. | Double | |
Training log path | Specifies the save path for training logs | String | |
Load pretrained model | Enable to add pretrained model path for transfer learning | Boolean | |
Pretrained model | Specifies pretrained model file (*.sdm) for transfer learning | String | |
Fine-tuning method | Full fine-tuning | Segformer and Mask2Former support full/LoRA fine-tuning. Other models only support full fine-tuning. LoRA enables efficient transfer learning for large-parameter models. | String |
Processor type | GPU | Specifies processor type | String |
GPU index | 0 | Input GPU index(es). For multi-GPU training, use comma-separated values (e.g., "0,1,2"). | String |
Model save path | Specifies model storage path | String | |
Model name | saved_model | Specifies model name | String |
Output
Generated model files and training logs.