During training and evaluation of a YOLO (You Only Look Once) object detection model, the console output provides key insights into performance and resource usage. Below is a breakdown of the typical metrics shown:

Training Log Example

Epoch    GPU_mem   box_loss   obj_loss   cls_loss  Instances       Size
  49/49      7.55G    0.03141    0.05182  0.0003582        104        640: 100%|██████████| 12/12 [00:03<00:00,  3.80it/s]

• Epoch 49/49: The model has completed its final training epoch (out of 49 total).
• GPU_mem 7.55G: Approximately 7.55 GB of GPU memory is in use.
• box_loss: Localization error between predicted and ground-truth bounding boxes (lower is better).
• obj_loss: Measures how well the model detects whether an object is present in a region.
• cls_loss: Classification error—how accurate the model assigns correct class labels.
• Instances 104: Total object instances processed in this epoch.
• Size 640: Input images are resized to 640×640 pixels.
• Progress bar: Indicates 12 batches were processed in ~3 seconds at ~3.8 iterations per second.

Validation Metrics

Class     Images  Instances          P          R      mAP50   mAP50-95: 100%|██████████| 2/2 [00:00<00:00,  2.62it/s]
  all         89       1006      0.966      0.935       0.98      0.655

• Images: 89 validation images used.
• Instances: 1006 annotated objects across those images.
• Precision (P) = 0.966: 96.6% of detected objects are true positives.
• Recall (R) = 0.935: The model detected 93.5% of all actual objects.
• mAP50 = 0.98: Mean Average Percision at IoU threshold ≥ 0.5.
• mAP50-95 = 0.655: Average mAP across IoU thresholds from 0.5 to 0.95 (in 0.05 increments).

Training Completion Summary

50 epochs completed in 0.059 hours.
Optimizer stripped from runs/train/exp8/weights/last.pt, 14.4MB
Optimizer stripped from runs/train/exp8/weights/best.pt, 14.4MB

The full training run finished in about 3.5 minutes. The optimizer state was removed from both last.pt (final epoch weights) and best.pt (weights with highest validation mAP), reducing file size for inference-only deployment.

Model Validation and Summary

Validating runs/train/exp8/weights/best.pt...
Fusing layers... 
Model summary: 157 layers, 7015519 parameters, 0 gradients, 15.8 GFLOPs

• Fusing layers: Optimization step that merges batch normalization into convolutional layers for faster inference.
• 157 layers: Total number of layers in the network architecture.
• 7,015,519 parameters: Total trainable weights before optimization.
• 0 gradients: Confirms the model is in evaluation mode (no backpropagation).
• 15.8 GFLOPs: Computational cost per forward pass—useful for estimating inference speed on different hardware.

Per-Class Validation Results

Class     Images  Instances          P          R      mAP50   mAP50-95
  all         89       1006      0.966      0.935       0.98      0.654
  hat         89       1006      0.966      0.935       0.98      0.654

In this single-class scenario ("hat"), per-class metrics match the overall results. All output and weights are saved under runs/train/exp8/.