ultralytics/yolov5

Clone the repository and install dependencies in a Python>=3.8.0 environment. Ensure you have PyTorch>=1.8 installed.

# Clone the YOLOv5 repository
git clone https://github.com/ultralytics/yolov5

# Navigate to the cloned directory
cd yolov5

# Install required packages
pip install -r requirements.txt

Use YOLOv5 via PyTorch Hub for inference. Models are automatically downloaded from the latest YOLOv5 release.

import torch

# Load a YOLOv5 model (options: yolov5n, yolov5s, yolov5m, yolov5l, yolov5x)
model = torch.hub.load("ultralytics/yolov5", "yolov5s")  # Default: yolov5s

# Define the input image source (URL, local file, PIL image, OpenCV frame, numpy array, or list)
img = "https://ultralytics.com/images/zidane.jpg"  # Example image

# Perform inference (handles batching, resizing, normalization automatically)
results = model(img)

# Process the results (options: .print(), .show(), .save(), .crop(), .pandas())
results.print()  # Print results to console
results.show()  # Display results in a window
results.save()  # Save results to runs/detect/exp

The detect.py script runs inference on various sources. It automatically downloads models from the latest YOLOv5 release and saves the results to the runs/detect directory.

# Run inference using a webcam
python detect.py --weights yolov5s.pt --source 0

# Run inference on a local image file
python detect.py --weights yolov5s.pt --source img.jpg

# Run inference on a local video file
python detect.py --weights yolov5s.pt --source vid.mp4

# Run inference on a screen capture
python detect.py --weights yolov5s.pt --source screen

# Run inference on a directory of images
python detect.py --weights yolov5s.pt --source path/to/images/

# Run inference on a text file listing image paths
python detect.py --weights yolov5s.pt --source list.txt

# Run inference on a text file listing stream URLs
python detect.py --weights yolov5s.pt --source list.streams

# Run inference using a glob pattern for images
python detect.py --weights yolov5s.pt --source 'path/to/*.jpg'

# Run inference on a YouTube video URL
python detect.py --weights yolov5s.pt --source 'https://youtu.be/LNwODJXcvt4'

# Run inference on an RTSP, RTMP, or HTTP stream
python detect.py --weights yolov5s.pt --source 'rtsp://example.com/media.mp4'

The commands below demonstrate how to reproduce YOLOv5 COCO dataset results. Both models and datasets are downloaded automatically from the latest YOLOv5 release. Training times for YOLOv5n/s/m/l/x are approximately 1/2/4/6/8 days on a single NVIDIA V100 GPU. Using Multi-GPU training can significantly reduce training time. Use the largest --batch-size your hardware allows, or use --batch-size -1 for YOLOv5 AutoBatch. The batch sizes shown below are for V100-16GB GPUs.

# Train YOLOv5n on COCO for 300 epochs
python train.py --data coco.yaml --epochs 300 --weights '' --cfg yolov5n.yaml --batch-size 128

# Train YOLOv5s on COCO for 300 epochs
python train.py --data coco.yaml --epochs 300 --weights '' --cfg yolov5s.yaml --batch-size 64

# Train YOLOv5m on COCO for 300 epochs
python train.py --data coco.yaml --epochs 300 --weights '' --cfg yolov5m.yaml --batch-size 40

# Train YOLOv5l on COCO for 300 epochs
python train.py --data coco.yaml --epochs 300 --weights '' --cfg yolov5l.yaml --batch-size 24

# Train YOLOv5x on COCO for 300 epochs
python train.py --data coco.yaml --epochs 300 --weights '' --cfg yolov5x.yaml --batch-size 16

• Train Custom Data 🚀 RECOMMENDED: Learn how to train YOLOv5 on your own datasets.
• Tips for Best Training Results ☘️: Improve your model's performance with expert tips.
• Multi-GPU Training: Speed up training using multiple GPUs.
• PyTorch Hub Integration: Easily load models using PyTorch Hub.
• Model Export (TFLite, ONNX, CoreML, TensorRT) 🚀: Convert your models to various deployment formats like ONNX or TensorRT.
• NVIDIA Jetson Deployment: Deploy YOLOv5 on NVIDIA Jetson devices.
• Test-Time Augmentation (TTA): Enhance prediction accuracy with TTA.
• Model Ensembling: Combine multiple models for better performance.
• Model Pruning/Sparsity: Optimize models for size and speed.
• Hyperparameter Evolution: Automatically find the best training hyperparameters.
• Transfer Learning with Frozen Layers: Adapt pretrained models to new tasks efficiently using transfer learning.
• Architecture Summary: Understand the YOLOv5 model architecture.
• Ultralytics Platform Training 🚀 RECOMMENDED: Train and deploy YOLO models using Ultralytics Platform.
• ClearML Logging: Integrate with ClearML for experiment tracking.
• Neural Magic DeepSparse Integration: Accelerate inference with DeepSparse.
• Comet Logging: Log experiments using Comet ML.

• COCO AP val denotes the mean Average Precision (mAP) at Intersection over Union (IoU) thresholds from 0.5 to 0.95, measured on the 5,000-image COCO val2017 dataset across various inference sizes (256 to 1536 pixels).
• GPU Speed measures the average inference time per image on the COCO val2017 dataset using an AWS p3.2xlarge V100 instance with a batch size of 32.
• EfficientDet data is sourced from the google/automl repository at batch size 8.
• Reproduce these results using the command: python val.py --task study --data coco.yaml --iou 0.7 --weights yolov5n6.pt yolov5s6.pt yolov5m6.pt yolov5l6.pt yolov5x6.pt

• The 640-pixel rows report the legacy anchor-based YOLOv5 checkpoints, while their links open the corresponding anchor-free YOLOv5u successor models on Ultralytics Platform. P6 links continue to download the legacy 1280-pixel checkpoints directly.
• All checkpoints were trained for 300 epochs using default settings. Nano (n) and Small (s) models use hyp.scratch-low.yaml hyperparameters, while Medium (m), Large (l), and Extra-Large (x) models use hyp.scratch-high.yaml.
• mAP^val values represent single-model, single-scale performance on the COCO val2017 dataset.
  Reproduce using: python val.py --data coco.yaml --img 640 --conf 0.001 --iou 0.65
• Speed metrics are averaged over COCO val images using an AWS p3.2xlarge V100 instance. Non-Maximum Suppression (NMS) time (~1 ms/image) is not included.
  Reproduce using: python val.py --data coco.yaml --img 640 --task speed --batch 1
• TTA (Test Time Augmentation) includes reflection and scale augmentations for improved accuracy.
  Reproduce using: python val.py --data coco.yaml --img 1536 --iou 0.7 --augment

YOLOv5 segmentation models were trained on the COCO dataset for 300 epochs at an image size of 640 pixels using A100 GPUs. Models were exported to ONNX FP32 for CPU speed tests and TensorRT FP16 for GPU speed tests. All speed tests were conducted on Google Colab Pro notebooks for reproducibility.

• All checkpoints were trained for 300 epochs using the SGD optimizer with lr0=0.01 and weight_decay=5e-5 at an image size of 640 pixels, using default settings.
  Training runs are logged at https://wandb.ai/glenn-jocher/YOLOv5_v70_official.
• Accuracy values represent single-model, single-scale performance on the COCO dataset.
  Reproduce using: python segment/val.py --data coco.yaml --weights yolov5s-seg.pt
• Speed metrics are averaged over 100 inference images using a Colab Pro A100 High-RAM instance. Values indicate inference speed only (NMS adds approximately 1ms per image).
  Reproduce using: python segment/val.py --data coco.yaml --weights yolov5s-seg.pt --batch 1
• Export to ONNX (FP32) and TensorRT (FP16) was performed using export.py.
  Reproduce using: python export.py --weights yolov5s-seg.pt --include engine --device 0 --half

Segmentation Usage Examples  

Train

YOLOv5 segmentation training supports automatic download of the COCO128-seg dataset via the --data coco128-seg.yaml argument. For the full COCO-segments dataset, download it manually using bash data/scripts/get_coco.sh --train --val --segments and then train with python segment/train.py --data coco.yaml.

# Train on a single GPU
python segment/train.py --data coco128-seg.yaml --weights yolov5s-seg.pt --img 640

# Train using Multi-GPU Distributed Data Parallel (DDP)
python -m torch.distributed.run --nproc_per_node 4 --master_port 1 segment/train.py --data coco128-seg.yaml --weights yolov5s-seg.pt --img 640 --device 0,1,2,3

Val

Validate the mask mean Average Precision (mAP) of YOLOv5s-seg on the COCO dataset:

# Download COCO validation segments split (780MB, 5000 images)
bash data/scripts/get_coco.sh --val --segments

# Validate the model
python segment/val.py --weights yolov5s-seg.pt --data coco.yaml --img 640

Predict

Use the pretrained YOLOv5m-seg.pt model to perform segmentation on bus.jpg:

# Run prediction
python segment/predict.py --weights yolov5m-seg.pt --source data/images/bus.jpg

# Load model from PyTorch Hub (Note: Inference support might vary)
model = torch.hub.load("ultralytics/yolov5", "custom", "yolov5m-seg.pt")

Export

Export the YOLOv5s-seg model to ONNX and TensorRT formats:

# Export model
python export.py --weights yolov5s-seg.pt --include onnx engine --img 640 --device 0

YOLOv5-cls classification models were trained on ImageNet for 90 epochs using a 4xA100 instance. ResNet and EfficientNet models were trained alongside under identical settings for comparison. Models were exported to ONNX FP32 (CPU speed tests) and TensorRT FP16 (GPU speed tests). All speed tests were run on Google Colab Pro for reproducibility.

Classification Usage Examples  

Train

YOLOv5 classification training supports automatic download for datasets like MNIST, Fashion-MNIST, CIFAR10, CIFAR100, Imagenette, Imagewoof, and ImageNet using the --data argument. For example, start training on MNIST with --data mnist.

# Train on a single GPU using CIFAR-100 dataset
python classify/train.py --model yolov5s-cls.pt --data cifar100 --epochs 5 --img 224 --batch 128

# Train using Multi-GPU DDP on ImageNet dataset
python -m torch.distributed.run --nproc_per_node 4 --master_port 1 classify/train.py --model yolov5s-cls.pt --data imagenet --epochs 5 --img 224 --device 0,1,2,3

Val

Validate the accuracy of the YOLOv5m-cls model on the ImageNet-1k validation dataset:

# Download ImageNet validation split (6.3GB, 50,000 images)
bash data/scripts/get_imagenet.sh --val

# Validate the model
python classify/val.py --weights yolov5m-cls.pt --data ../datasets/imagenet --img 224

Predict

Use the pretrained YOLOv5s-cls.pt model to classify the image bus.jpg:

# Run prediction
python classify/predict.py --weights yolov5s-cls.pt --source data/images/bus.jpg

# Load model from PyTorch Hub
model = torch.hub.load("ultralytics/yolov5", "custom", "yolov5s-cls.pt")

Export

Export trained YOLOv5s-cls, ResNet50, and EfficientNet_b0 models to ONNX and TensorRT formats:

# Export models
python export.py --weights yolov5s-cls.pt resnet50.pt efficientnet_b0.pt --include onnx engine --img 224