83 lines
7.5 KiB
Plaintext
83 lines
7.5 KiB
Plaintext
|
a
|
|||
|
|
ˆ’>j ( ã @ s~ d Z ddlmZmZmZmZ ddlmZ ddlm Z ddl
|
|||
|
|
Z
|
|||
|
|
ddl Z eG dd„ dƒƒZ G dd „ d ƒZ
|
|||
|
|
d
|
|||
|
|
d „ Zed krzeƒ dS )
|
|||
|
|
zC
|
|||
|
|
AI Training Pipeline
|
|||
|
|
Train YOLO and CLIP on infrastructure images
|
|||
|
|
é )ÚDictÚListÚOptionalÚTuple)Ú dataclass)ÚPathNc @ sj e Zd ZU dZeed< eed< dZeed< dZeed< dZ eed < d
|
|||
|
|
Z
|
|||
|
|
eed < d Z e ed
|
|||
|
|
< dZ
|
|||
|
|
eed< dS )ÚTrainingConfigzTraining configurationÚ dataset_pathÚ output_pathÚyolov8nÚ
|
|||
|
|
model_typeéd Úepochsé Ú
|
|||
|
|
batch_sizei€ Ú
|
|||
|
|
image_sizegü©ñÒMbP?Ú
|
|||
|
|
learning_rateÚcpuÚdeviceN)Ú__name__Ú
|
|||
|
|
__module__Ú __qualname__Ú__doc__ÚstrÚ__annotations__r r Úintr r r Úfloatr © r r úD/home/bernt/.openclaw/workspace/iom/ai_pipeline/training_pipeline.pyr
|
|||
|
|
s
|
|||
|
|
r c @ sž e Zd ZdZedœdd„Zedœdd„Zedœd d
|
|||
|
|
„Z eed œd d
|
|||
|
|
„Z
|
|||
|
|
edœdd„Z edœdd„Z edœdd„Z
|
|||
|
|
edœdd„Zedœdd„Zedœdd„ZdS )ÚTrainingPipelinea,
|
|||
|
|
Training pipeline for infrastructure AI models
|
|||
|
|
|
|||
|
|
Steps:
|
|||
|
|
1. Prepare dataset (images + annotations)
|
|||
|
|
2. Train YOLO for object detection
|
|||
|
|
3. Fine-tune CLIP for scene classification
|
|||
|
|
4. Train custom defect classifier
|
|||
|
|
5. Evaluate models
|
|||
|
|
6. Export to production format
|
|||
|
|
)Úconfigc C s || _ d | _d | _d | _d S )N)r ÚdatasetZ
|
|||
|
|
yolo_modelZ
|
|||
|
|
clip_model)Úselfr r r r Ú__init__' s zTrainingPipeline.__init__)Úreturnc C s’ t | jjƒ}| ¡ s4td|› �ƒ tdƒ | |¡ | |¡}tdƒ td|d › �ƒ td|d › �ƒ td|d › �ƒ td
|
|||
|
|
|d › �ƒ |S ) a
|
|||
|
|
Prepare dataset for training
|
|||
|
|
|
|||
|
|
Expected structure:
|
|||
|
|
dataset/
|
|||
|
|
images/
|
|||
|
|
train/
|
|||
|
|
val/
|
|||
|
|
test/
|
|||
|
|
labels/
|
|||
|
|
train/
|
|||
|
|
val/
|
|||
|
|
test/
|
|||
|
|
data.yaml
|
|||
|
|
zDataset not found at z$Creating sample dataset structure...zDataset prepared:z Train images: Ú train_imagesz Val images: Ú
|
|||
|
|
val_imagesz Test images: Ú test_imagesz Classes: Úclasses)r r r ÚexistsÚprintÚ_create_sample_datasetÚ_validate_dataset)r" r Ústatsr r r Úprepare_dataset- s
|
|||
|
|
|
|||
|
|
z TrainingPipeline.prepare_dataset)Úpathc C sî |d d j dd� |d d j dd� |d d j dd� |d d j dd� |d d j dd� |d d j dd� t| ¡ ƒdd d
|
|||
|
|
d g d ¢d
|
|||
|
|
œ}t|d dƒ� }tj||dd� W d ƒ n1 sÊ0 Y td|› �ƒ tdƒ dS )z Create sample dataset structureÚimagesÚtrainT)ÚparentsÚvalÚtestÚlabelsz images/trainz
|
|||
|
|
images/valz images/testé
|
|||
|
|
)
|
|||
|
|
Ú street_lightÚ traffic_signÚbenchÚ trash_canÚsidewalkÚroadÚbuildingÚbridgeÚtreeÚgraffiti)r/ r1 r3 r4 ÚncÚnamesú data.yamlÚwé ©ÚindentNzSample dataset created at z-Add your images and labels to train the model)Úmkdirr ÚabsoluteÚopenÚjsonÚdumpr* )r" r/ Z data_yamlÚfr r r r+ O s"
|
|||
|
|
ú.z'TrainingPipeline._create_sample_dataset)r/ r$ c C sš t t|d d d¡ƒƒ}t t|d d d¡ƒƒ}t t|d d d¡ƒƒ}t|d ƒ�}t |¡}W d ƒ n1 sz0 Y |||| dg ¡d œS )
|
|||
|
|
zValidate dataset structurer0 r1 Ú*r3 r4 rC NrB )r% r& r' r( )ÚlenÚlistÚglobrJ rK ÚloadÚget)r" r/ r% r& r' rM Údatar r r r, t s (
|
|||
|
|
üz"TrainingPipeline._validate_datasetc
|
|||
|
|
C s’ ddl m} tdƒ || jj› d�ƒ}|j| jj› d�| jj| jj| jj | jj
|
|||
|
|
| jj | jj dd�}| jj › d�}td ƒ td
|
|||
|
|
|› �ƒ ||j
|
|||
|
|
d œS ) zTrain YOLO modelr )ÚYOLOz
|
|||
|
|
=== Training YOLO ===z.ptz
|
|||
|
|
/data.yamlÚyolo_infrastructure)rT r ÚbatchÚimgszZlr0r ÚprojectÚnameú$/yolo_infrastructure/weights/best.ptzYOLO training completeú
|
|||
|
|
Model saved: )Ú
|
|||
|
|
model_pathÚmetrics)Ú ultralyticsrU r* r r r1 r r r r r r r
|
|||
|
|
Z results_dict)r" rU ÚmodelÚresultsr] r r r Ú
|
|||
|
|
train_yolo… s& ø þzTrainingPipeline.train_yoloc C sn ddl m}m}m}m} ddlm} tdƒ | d¡}| d¡}| j j
|
|||
|
|
› d�}tdƒ td|› �ƒ |d d
|
|||
|
|
œS ) z'Fine-tune CLIP for scene classificationr )Ú CLIPModelÚ
|
|||
|
|
CLIPProcessorÚTrainingArgumentsÚTrainer)Ú load_datasetz
|
|||
|
|
=== Fine-tuning CLIP ===z openai/clip-vit-base-patch32ú/clip_infrastructurezCLIP fine-tuning completer\ Útrained©r] Zstatus) Z transformersrc rd re rf Zdatasetsrg r* Zfrom_pretrainedr r
|
|||
|
|
) r" rc rd re rf rg r` Z processorr] r r r Ú
|
|||
|
|
finetune_clip¥ s
|
|||
|
|
|
|||
|
|
þz TrainingPipeline.finetune_clipc C s6 t dƒ | jj› d�}t dƒ t d|› �ƒ |ddœS )z Train custom defect classifierz#
|
|||
|
|
=== Training Defect Classifier ===ú/defect_classifier.pthz#Defect classifier training completer\ ri rj )r* r r
|
|||
|
|
)r" r] r r r Útrain_defect_classifierÐ s þz(TrainingPipeline.train_defect_classifierc C s€ t dƒ dddddœddd œd
|