Get Started

Image Classification Datasets Overview

Dataset Structure for YOLO Classification Tasks#

For Ultralytics YOLO classification tasks, the dataset must be organized in a specific split-directory structure under the root directory. This structure includes a train directory and a val directory (valid and validation are also accepted), with an optional test directory. If no validation directory exists, the test directory is used for validation instead. If the root has no train directory at all but contains one folder per class, Ultralytics automatically splits the images 80/20 into a new <root>_split directory with train and val subdirectories.

Each of these directories should contain one subdirectory for each class in the dataset. The subdirectories are named after the corresponding class and contain all the images for that class. Ensure that each image file is named uniquely and stored in a common format such as JPEG or PNG.

Folder Structure Example#

Consider the CIFAR-10 dataset as an example. The folder structure should look like this:

cifar10/
|
|-- train/
|   |-- airplane/
|   |   |-- 10008_airplane.png
|   |   |-- 10009_airplane.png
|   |   |-- ...
|   |
|   |-- automobile/
|   |   |-- 1000_automobile.png
|   |   |-- 1001_automobile.png
|   |   |-- ...
|   |
|   |-- bird/
|   |   |-- 10014_bird.png
|   |   |-- 10015_bird.png
|   |   |-- ...
|   |
|   |-- ...
|
|-- test/ (optional)
|   |-- airplane/
|   |   |-- 10_airplane.png
|   |   |-- 11_airplane.png
|   |   |-- ...
|   |
|   |-- automobile/
|   |   |-- 100_automobile.png
|   |   |-- 101_automobile.png
|   |   |-- ...
|   |
|   |-- bird/
|   |   |-- 1000_bird.png
|   |   |-- 1001_bird.png
|   |   |-- ...
|   |
|   |-- ...
|
|-- val/
|   |-- airplane/
|   |   |-- 105_airplane.png
|   |   |-- 106_airplane.png
|   |   |-- ...
|   |
|   |-- automobile/
|   |   |-- 102_automobile.png
|   |   |-- 103_automobile.png
|   |   |-- ...
|   |
|   |-- bird/
|   |   |-- 1045_bird.png
|   |   |-- 1046_bird.png
|   |   |-- ...
|   |
|   |-- ...

This structured approach ensures that the model can effectively learn from well-organized classes during the training phase and accurately evaluate performance during testing and validation phases.

Usage#

Example
from ultralytics import YOLO

# Load a model
model = YOLO("yolo26n-cls.pt")  # load a pretrained model (recommended for training)

# Train the model
results = model.train(data="path/to/dataset", epochs=100, imgsz=640)
Tip

Most built-in dataset names (for example cifar10, imagenette, or mnist160) will automatically download and cache the data the first time you reference them. Point data to a folder path only when you have curated a custom dataset.

Supported Datasets#

Ultralytics supports the following datasets with automatic download. Most of these datasets are also hosted on Ultralytics Platform, where you can browse the images and annotations, view dataset statistics, and clone them for cloud training.

  • Caltech 101: A dataset containing images of 101 object categories for image classification tasks.
  • Caltech 256: An extended version of Caltech 101 with 256 object categories and more challenging images.
  • CIFAR-10: A dataset of 60K 32x32 color images in 10 classes, with 6K images per class.
  • CIFAR-100: An extended version of CIFAR-10 with 100 object categories and 600 images per class.
  • Fashion-MNIST: A dataset consisting of 70,000 grayscale images of 10 fashion categories for image classification tasks.
  • ImageNet: The ImageNet-1k (ILSVRC-2012) image classification benchmark with 1,000 classes, 1.28 million training images, and 50,000 validation images.
  • ImageNet-10: A smaller subset of ImageNet with 10 categories for faster experimentation and testing.
  • Imagenette: A smaller subset of ImageNet that contains 10 easily distinguishable classes for quicker training and testing.
  • Imagewoof: A more challenging subset of ImageNet containing 10 dog breed categories for image classification tasks.
  • MNIST: A dataset of 70,000 grayscale images of handwritten digits for image classification tasks.
  • MNIST160: First 8 images of each digit (0-9) from both the MNIST train and test splits. Dataset contains 160 images total.

Adding your own dataset#

If you have your own dataset and would like to use it for training classification models with Ultralytics YOLO, ensure that it follows the format specified above under "Dataset Structure" and then point your data argument to the dataset directory when initializing your training script.

FAQ#

  • To structure your dataset for Ultralytics YOLO classification tasks, you should follow a specific split-directory format. Organize your dataset into separate directories for train and val, and optionally test. Each of these directories should contain subdirectories named after each class, with the corresponding images inside. This facilitates smooth training and evaluation processes. For an example, consider the CIFAR-10 dataset format:

    cifar10/
    |-- train/
    |   |-- airplane/
    |   |-- automobile/
    |   |-- bird/
    |   ...
    |-- val/
    |   |-- airplane/
    |   |-- automobile/
    |   |-- bird/
    |   ...
    |-- test/ (optional)
    |   |-- airplane/
    |   |-- automobile/
    |   |-- bird/
    |   ...

    For more details, visit the Dataset Structure for YOLO Classification Tasks section.

  • Ultralytics YOLO supports automatic downloading of several datasets for image classification, including Caltech 101, Caltech 256, CIFAR-10, CIFAR-100, Fashion-MNIST, ImageNet, ImageNet-10, Imagenette, Imagewoof, and MNIST. These datasets are structured in a way that makes them easy to use with YOLO. Each dataset's page provides further details about its structure and applications.

  • To use your own dataset with Ultralytics YOLO, ensure it follows the specified directory format required for the classification task, with separate train and val directories (and optionally test), and subdirectories for each class containing the respective images. Once your dataset is structured correctly, point the data argument to your dataset's root directory when initializing the training script. Here's an example in Python:

    from ultralytics import YOLO
    
    # Load a model
    model = YOLO("yolo26n-cls.pt")  # load a pretrained model (recommended for training)
    
    # Train the model
    results = model.train(data="path/to/your/dataset", epochs=100, imgsz=640)

    More details can be found in the Adding your own dataset section.

  • Ultralytics YOLO offers several benefits for image classification, including:

    • Pretrained Models: Load pretrained models like yolo26n-cls.pt to jump-start your training process.
    • Ease of Use: Simple API and CLI commands for training and evaluation.
    • High Performance: Strong accuracy and speed, suitable for real-time applications.
    • Support for Multiple Datasets: Built-in support for popular datasets like CIFAR-10, ImageNet, and more.
    • Community and Support: Access to extensive documentation and an active community for troubleshooting and improvements.

    For additional insights and real-world applications, you can explore Ultralytics YOLO.

  • You can train a model with Ultralytics YOLO from either Python or the CLI. Here's an example:

    Example
    from ultralytics import YOLO
    
    # Load a model
    model = YOLO("yolo26n-cls.pt")  # load a pretrained model
    
    # Train the model
    results = model.train(data="path/to/dataset", epochs=100, imgsz=640)

    These examples demonstrate the straightforward process of training a YOLO model using either approach. For more information, visit the Usage section and the Train page for classification tasks.

Comments45

Ask a question, share your results, or help someone else.

  1. iongionGitHub

    My annotations have regions that I know variate a lot and cannot be used for classification. Is there a way to feed these regions so that it avoids learning from them during training ? If I mask/fill the areas with black color on the image for training itself, wouldn't that affect classification ? Sometimes some images might have black shapes as features themselves.

  2. lauravchrobakGitHub

    Its confusing to me that the val set is optional for classification but the test set is optional for detection. Should there be some consistency there?

    1. Glenn JocherUltralytics(edited)

      @lauravchrobak great observation! For classification, train plus either a val or a test folder is required: if one of val/test is missing, Ultralytics uses the other for validation. Detection datasets use a YAML file where train and val are required and test is optional. Both layouts are described in the classification and detection dataset docs.

  3. 1ZZZyyyhGitHub

    that's why? I've made my dataset as the sample.

    File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/engine/trainer.py", line 567, in get_dataset raise RuntimeError(emojis(f"Dataset '{clean_url(self.args.data)}' error ❌ {e}")) from e RuntimeError: Dataset '/home/visitor/yolov8-tb11' error ❌ [Errno 21] Is a directory: '/home/visitor/yolov8-tb11'

    1. Glenn JocherUltralytics(edited)

      This error appears when a directory is passed as data to a non-classification task. For classification, data should be the dataset root directory (with train/ and val/ or test/ folders, one subfolder per class), and you should load a classification model such as yolo26n-cls.pt (or set task=classify). For detection, segmentation, pose or OBB, data must be a data.yaml file. See Dataset Structure for YOLO Classification Tasks.

    2. 1ZZZyyyhGitHub

      I want to train my model , it reports error. yolo task=classify mode=train model=yolov8m-cls.pt data=/home/visitor/yolov8-tb11 epochs=100 imgsz=640 ... Traceback (most recent call last): File "/home/visitor/anaconda3/bin/yolo", line 8, in sys.exit(entrypoint()) ^^^^^^^^^^^^ File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/cfg/init.py", line 991, in entrypoint getattr(model, mode)(**overrides) # default args from model ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/engine/model.py", line 806, in train self.trainer.train() File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/engine/trainer.py", line 207, in train self._do_train(world_size) File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/engine/trainer.py", line 323, in _do_train self._setup_train(world_size) File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/engine/trainer.py", line 264, in _setup_train self.amp = torch.tensor(check_amp(self.model), device=self.device) ^^^^^^^^^^^^^^^^^^^^^ File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/utils/checks.py", line 707, in check_amp assert amp_allclose(YOLO("yolo11n.pt"), im) ^^^^^^^^^^^^^^^^^^ File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/models/yolo/model.py", line 23, in init super().init(model=model, task=task, verbose=verbose) File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/engine/model.py", line 146, in init self._load(model, task=task) File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/engine/model.py", line 289, in _load self.model, self.ckpt = attempt_load_one_weight(weights) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/nn/tasks.py", line 908, in attempt_load_one_weight ckpt, weight = torch_safe_load(weight) # load ckpt ^^^^^^^^^^^^^^^^^^^^^^^ File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/nn/tasks.py", line 835, in torch_safe_load ckpt = torch.load(file, map_location="cpu") ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/visitor/anaconda3/lib/python3.12/site-packages/ultralytics/utils/patches.py", line 86, in torch_load return _torch_load(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/visitor/anaconda3/lib/python3.12/site-packages/torch/serialization.py", line 1326, in load with _open_zipfile_reader(opened_file) as opened_zipfile: ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/visitor/anaconda3/lib/python3.12/site-packages/torch/serialization.py", line 671, in init super().init(torch._C.PyTorchFileReader(name_or_buffer)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ RuntimeError: PytorchStreamReader failed reading zip archive: failed finding central directory

    3. Glenn JocherUltralytics

      @1ZZZyyyh the error indicates an issue with your dataset path or structure. Ensure the data argument points to a valid dataset configuration file (.yaml) or a properly structured classification dataset with train/, test/, and optionally val/ directories, as outlined in the classification dataset structure guide. Let me know if you need further assistance!

  4. rishabh-jain14GitHub

    I have this yaml file for caltech-256 classification: path: D:/Coursework/Fall 24/CSC591-007 DLBA/Project/Project/data/tvt_Caltech_256 train: train/ val: val/ test: test/ nc: 257 names: ['001.ak47', '002.american-flag', '003.backpack', '004.baseball-bat', '005.baseball-glove', '006.basketball-hoop', '007.bat', '008.bathtub', '009.bear', '010.beer-mug', '011.billiards', '012.binoculars', '013.birdbath', '014.blimp', '015.bonsai-101', '016.boom-box', '017.bowling-ball', '018.bowling-pin', '019.boxing-glove', '020.brain-101', '021.breadmaker', '022.buddha-101', '023.bulldozer', '024.butterfly', '025.cactus', '026.cake', '027.calculator', '028.camel', '029.cannon', '030.canoe', '031.car-tire', '032.cartman', '033.cd', '034.centipede', '035.cereal-box', '036.chandelier-101', '037.chess-board', '038.chimp', '039.chopsticks', '040.cockroach', '041.coffee-mug', '042.coffin', '043.coin', '044.comet', '045.computer-keyboard', '046.computer-monitor', '047.computer-mouse', '048.conch', '049.cormorant', '050.covered-wagon', '051.cowboy-hat', '052.crab-101', '053.desk-globe', '054.diamond-ring', '055.dice', '056.dog', '057.dolphin-101', '058.doorknob', '059.drinking-straw', '060.duck', '061.dumb-bell', '062.eiffel-tower', '063.electric-guitar-101', '064.elephant-101', '065.elk', '066.ewer-101', '067.eyeglasses', '068.fern', '069.fighter-jet', '070.fire-extinguisher', '071.fire-hydrant', '072.fire-truck', '073.fireworks', '074.flashlight', '075.floppy-disk', '076.football-helmet', '077.french-horn', '078.fried-egg', '079.frisbee', '080.frog', '081.frying-pan', '082.galaxy', '083.gas-pump', '084.giraffe', '085.goat', '086.golden-gate-bridge', '087.goldfish', '088.golf-ball', '089.goose', '090.gorilla', '091.grand-piano-101', '092.grapes', '093.grasshopper', '094.guitar-pick', '095.hamburger', '096.hammock', '097.harmonica', '098.harp', '099.harpsichord', '100.hawksbill-101', '101.head-phones', '102.helicopter-101', '103.hibiscus', '104.homer-simpson', '105.horse', '106.horseshoe-crab', '107.hot-air-balloon', '108.hot-dog', '109.hot-tub', '110.hourglass', '111.house-fly', '112.human-skeleton', '113.hummingbird', '114.ibis-101', '115.ice-cream-cone', '116.iguana', '117.ipod', '118.iris', '119.jesus-christ', '120.joy-stick', '121.kangaroo-101', '122.kayak', '123.ketch-101', '124.killer-whale', '125.knife', '126.ladder', '127.laptop-101', '128.lathe', '129.leopards-101', '130.license-plate', '131.lightbulb', '132.light-house', '133.lightning', '134.llama-101', '135.mailbox', '136.mandolin', '137.mars', '138.mattress', '139.megaphone', '140.menorah-101', '141.microscope', '142.microwave', '143.minaret', '144.minotaur', '145.motorbikes-101', '146.mountain-bike', '147.mushroom', '148.mussels', '149.necktie', '150.octopus', '151.ostrich', '152.owl', '153.palm-pilot', '154.palm-tree', '155.paperclip', '156.paper-shredder', '157.pci-card', '158.penguin', '159.people', '160.pez-dispenser', '161.photocopier', '162.picnic-table', '163.playing-card', '164.porcupine', '165.pram', '166.praying-mantis', '167.pyramid', '168.raccoon', '169.radio-telescope', '170.rainbow', '171.refrigerator', '172.revolver-101', '173.rifle', '174.rotary-phone', '175.roulette-wheel', '176.saddle', '177.saturn', '178.school-bus', '179.scorpion-101', '180.screwdriver', '181.segway', '182.self-propelled-lawn-mower', '183.sextant', '184.sheet-music', '185.skateboard', '186.skunk', '187.skyscraper', '188.smokestack', '189.snail', '190.snake', '191.sneaker', '192.snowmobile', '193.soccer-ball', '194.socks', '195.soda-can', '196.spaghetti', '197.speed-boat', '198.spider', '199.spoon', '200.stained-glass', '201.starfish-101', '202.steering-wheel', '203.stirrups', '204.sunflower-101', '205.superman', '206.sushi', '207.swan', '208.swiss-army-knife', '209.sword', '210.syringe', '211.tambourine', '212.teapot', '213.teddy-bear', '214.teepee', '215.telephone-box', '216.tennis-ball', '217.tennis-court', '218.tennis-racket', '219.theodolite', '220.toaster', '221.tomato', '222.tombstone', '223.top-hat', '224.touring-bike', '225.tower-pisa', '226.traffic-light', '227.treadmill', '228.triceratops', '229.tricycle', '230.trilobite-101', '231.tripod', '232.t-shirt', '233.tuning-fork', '234.tweezer', '235.umbrella-101', '236.unicorn', '237.vcr', '238.video-projector', '239.washing-machine', '240.watch-101', '241.waterfall', '242.watermelon', '243.welding-mask', '244.wheelbarrow', '245.windmill', '246.wine-bottle', '247.xylophone', '248.yarmulke', '249.yo-yo', '250.zebra', '251.airplanes-101', '252.car-side-101', '253.faces-easy-101', '254.greyhound', '255.tennis-shoes', '256.toad', '257.clutter']

    However, when I train, it shows: train: Scanning D:\Coursework\Fall 24\CSC591-007 DLBA\Project\Project\data\tvt_Caltech_256\train\001... 0 images, 24485 backgrounds, 0 corrupt: 100%|██████████| 24485/24485 [00:43<00:00, 565.92it/s] train: WARNING No labels found in D:\Coursework\Fall 24\CSC591-007 DLBA\Project\Project\data\tvt_Caltech_256\train\001.cache. See https://docs.ultralytics.com/datasets for dataset formatting guidance.

    train: New cache created: D:\Coursework\Fall 24\CSC591-007 DLBA\Project\Project\data\tvt_Caltech_256\train\001.cache WARNING No labels found in D:\Coursework\Fall 24\CSC591-007 DLBA\Project\Project\data\tvt_Caltech_256\train\001.cache, training may not work correctly. See https://docs.ultralytics.com/datasets for dataset formatting guidance. val: Scanning D:\Coursework\Fall 24\CSC591-007 DLBA\Project\Project\data\tvt_Caltech_256\val\001... 0 images, 3061 backgrounds, 0 corrupt: 100%|██████████| 3061/3061 [00:02<00:00, 1231.29it/s] val: WARNING No labels found in D:\Coursework\Fall 24\CSC591-007 DLBA\Project\Project\data\tvt_Caltech_256\val\001.cache. See https://docs.ultralytics.com/datasets for dataset formatting guidance.

    val: New cache created: D:\Coursework\Fall 24\CSC591-007 DLBA\Project\Project\data\tvt_Caltech_256\val\001.cache WARNING No labels found in D:\Coursework\Fall 24\CSC591-007 DLBA\Project\Project\data\tvt_Caltech_256\val\001.cache, training may not work correctly. See https://docs.ultralytics.com/datasets for dataset formatting guidance. Plotting labels to runs\detect\train\labels.jpg... zero-size array to reduction operation maximum which has no identity optimizer: 'optimizer=auto' found, ignoring 'lr0=0.01' and 'momentum=0.937' and determining best 'optimizer', 'lr0' and 'momentum' automatically... optimizer: AdamW(lr=3.8e-05, momentum=0.9) with parameter groups 81 weight(decay=0.0), 88 weight(decay=0.0005), 87 bias(decay=0.0) TensorBoard: model graph visualization added Image sizes 128 train, 128 val

    What should I do to resolve the issue?

    1. Glenn JocherUltralytics

      @rishabh-jain14 it looks like the issue stems from the dataset structure not aligning with the expected format for classification tasks. Each class should have its own directory containing the respective images. Please ensure your dataset follows this structure. For detailed guidance, refer to the Image Classification Datasets Overview.

  5. NajeebAhmad00GitHub

    I am using a yolov8-cls.pt model and want to add a new class to it. I need to train the model on this new class without affecting the existing classes. How can I do this?

    1. Glenn JocherUltralytics(edited)

      @NajeebAhmad00 Ultralytics classification models cannot add a class without retraining: the classifier head is rebuilt for the classes in your dataset. If you fine-tune only on images of the new class, the model will forget the existing ones. To keep the existing classes, include their images (or a representative sample) along with the new class in your dataset, organized as described in the Ultralytics documentation, then fine-tune:

      from ultralytics import YOLO
      
      # Load the pretrained model
      model = YOLO("yolo26n-cls.pt")
      
      # Train on the dataset containing the old and new classes
      results = model.train(data="path/to/your/dataset", epochs=100, imgsz=224)
  6. leolle520GitHub

    Can I use this model to train an one class image classifier as an abnormal image detector? Thank you.

    1. Glenn JocherUltralytics

      Yes, you can use the YOLO model to train a one-class image classifier for abnormal image detection. Simply organize your dataset with one class for normal images and another for abnormal images, then proceed with training. For more details on dataset structure, visit our documentation.

  7. MrLi-creatorGitHub

    This is my own small_data.yaml file: path: C:/Users/AIX1WX/Documents/10-POC/train_yolov8/yolov8-all-in-one-main_classification/datasets train: train/ val: val/ test: test/

    nc: 7 names: ['ABS','Airbag','Background','EPB','EPBGREEN','EPBRed','QMtt']

    class_weights: [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 3.0]

    And my train coda here: from ultralytics import YOLO import os os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE" if name == 'main': model = YOLO("yolov8n-cls.pt") # load a pretrained model (recommended for training) results = model.train(data="small_data.yaml", epochs=30, imgsz=50, device=None, batch=16)

    Why will report the error: Dataset not found , missing path C:\Users\AIX1WX\Documents\10-POC\train_yolov8\yolov8-all-in-one-main_detection\datasets\small_data.yaml, attempting download... Downloading https://github.com/ultralytics/yolov5/releases/download/v1.0/small_data.yaml.zip to 'C:\Users\AIX1WX\Documents\10-POC\train_yolov8\yolov8-all-in-one-main_detection\datasets\small_data.yaml.zip'...

    1. Glenn JocherUltralytics(edited)

      @MrLi-creator the error occurs because classification does not use a YAML file. Pass the dataset root directory as data (for example data="C:/Users/AIX1WX/Documents/10-POC/train_yolov8/yolov8-all-in-one-main_classification/datasets"), containing train/, val/ and optionally test/ folders, each with one subfolder per class. Keys such as nc, names and class_weights in your YAML are ignored. For more details on dataset structure, refer to the Ultralytics documentation.

    2. MrLi-creatorGitHub

      Hi glenn-jocher, I find the case: I first train yolov8n.pt in yolov8-all-in-one-main_detection file. And second time I train yolov8n-cls.pt in another file yolov8-all-in-one-main_classification. Then everytime train yolov8n-cls.pt, the dataset path always points to yolov8-all-in-one-main_detection. So is there any way to change the default path of the dataset

    3. Glenn JocherUltralytics(edited)

      @MrLi-creator it seems the issue is that small_data.yaml is not a valid classification dataset. Pass the dataset root directory (with train/, val/ and optionally test/ folders, each containing one subfolder per class) as data instead of a YAML file. For further details on dataset structure, you can refer to the Ultralytics documentation.

    4. Glenn JocherUltralytics

      Thank you for pointing that out. For classification tasks with YOLO, the dataset should be organized in a specific directory structure rather than using a data.yaml file. Ensure your dataset follows the format with separate train, val, and test directories, each containing subdirectories for each class. This structure is essential for proper training and evaluation. For more details, you can refer to the Ultralytics documentation on dataset structure.

    1. Paula DerrengerUltralytics(edited)

      @MrLi-creator Ultralytics does not have a class_weights training argument, so there is no setting to change class weights when training yolo26n-cls.pt. To handle class imbalance in classification, balance your dataset (for example by oversampling the minority classes) or modify the loss in the source code. See the classification dataset guide for the expected dataset structure.

    2. MrLi-creatorGitHub

      Hi pderrenger, Firstly, thank you for your response. I am trying to add classw_weights to the data.yaml file. However, it was found that the classification task cannot use the data.yaml file, as you suggested in the link(https://github.com/ultralytics/ultralytics/issues/13789). So are there any other ways to modify classw_weights in classification, preferably with an example

    3. Paula DerrengerUltralytics(edited)

      @MrLi-creator Ultralytics does not support a class_weights parameter for classification training, and a data.yaml file isn't used for classification tasks (you pass the dataset directory as data). To handle class imbalance, balance your dataset or modify the loss function in the source. See the Ultralytics documentation on classification datasets.

    4. MrLi-creatorGitHub

      @pderrenger Hi pderrenger. I have set class_weights parameter in the training script. But it report error.

      results = model.train(data='./datasets', epochs=30, imgsz=50, device=None, batch=16,classes =[1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 3.0] )
      
      SyntaxError: 'class_weights' is not a valid YOLO argument. Similar arguments are i.e. ['classes'].
      
    5. Paula DerrengerUltralytics

      Hi @MrLi-creator,

      The class_weights parameter is not directly supported in the train method for YOLO classification tasks. Instead, you can handle class imbalance by adjusting the dataset or using techniques like data augmentation. If you encounter any issues, please ensure you are using the latest version of the Ultralytics package. For more detailed guidance, you can refer to the Ultralytics documentation on classification datasets.

  8. casperking666GitHub

    for classification task, do I still need the data.yaml file? If not, how should I carry out validation and testing using the folder structure shown above?

    1. Glenn JocherUltralytics(edited)

      @casperking666 hi there!

      For classification tasks with Ultralytics YOLO, you do not need a data.yaml file. Instead, you should organize your dataset in the specific folder structure as outlined in the documentation. This structure includes separate directories for training (train), testing (test), and optionally validation (val), with each class having its own subdirectory containing the relevant images.

      Here’s a quick example of how your dataset should be structured:

      dataset/
      |
      |-- train/
      |   |-- class1/
      |   |   |-- image1.jpg
      |   |   |-- image2.jpg
      |   |   |-- ...
      |   |
      |   |-- class2/
      |   |   |-- image1.jpg
      |   |   |-- image2.jpg
      |   |   |-- ...
      |   |
      |   |-- ...
      |
      |-- test/
      |   |-- class1/
      |   |   |-- image1.jpg
      |   |   |-- image2.jpg
      |   |   |-- ...
      |   |
      |   |-- class2/
      |   |   |-- image1.jpg
      |   |   |-- image2.jpg
      |   |   |-- ...
      |   |
      |   |-- ...
      |
      |-- val/ (optional)
      |   |-- class1/
      |   |   |-- image1.jpg
      |   |   |-- image2.jpg
      |   |   |-- ...
      |   |
      |   |-- class2/
      |   |   |-- image1.jpg
      |   |   |-- image2.jpg
      |   |   |-- ...
      |   |
      |   |-- ...

      To carry out validation and testing, simply ensure that your dataset is organized as shown above. The train, test, and val directories should each contain subdirectories named after the classes, with the images for each class inside these subdirectories.

      Here’s how you can train and validate your model using the Ultralytics YOLO API:

      Python Example
      from ultralytics import YOLO
      
      # Load a pretrained classification model
      model = YOLO("yolo26n-cls.pt")
      
      # Train the model
      results = model.train(data="path/to/dataset", epochs=100, imgsz=640)
      
      # Validate the model
      metrics = model.val(data="path/to/dataset")
      CLI Example
      # Train the model
      yolo classify train data=path/to/dataset model=yolo26n-cls.pt epochs=100 imgsz=640
      
      # Validate the model
      yolo classify val data=path/to/dataset model=yolo26n-cls.pt

      For more detailed information, you can refer to the Ultralytics documentation on classification datasets.

      If you encounter any issues, please ensure you are using the latest version of the packages. If the problem persists, providing a minimum reproducible example can greatly help in diagnosing the issue. You can find more details on creating a reproducible example here.

      Happy training! 🚀

    2. casperking666GitHub

      thanks! I used comet ml during training, and it produced plots metrics/accuracy_top1 VS step and metrics/accuracy_top5 VS step. What sets did they use to generate these plots? I couldn't find any label indicate what set that is. I would assume it's the training set, as the accuracy is getting to close to 100%.

      second question, it's not clear to me what set this line of code is using metrics = model.val(data="path/to/dataset")? Is it the validation set or the testing set? If it's the validation set, how would I test the test set as they are all organized under the same directory as you previously showed?

    3. Glenn JocherUltralytics(edited)

      Hi @casperking666,

      Thank you for your questions! Let's address them one by one:

      1. Comet ML Metrics: The plots for metrics/accuracy_top1 VS step and metrics/accuracy_top5 VS step generated by Comet ML typically use the validation set during training. This is because the validation set provides a more accurate measure of the model's performance on unseen data, which helps in monitoring overfitting. If the accuracy is approaching 100%, it might indicate that the model is overfitting to the training data. To confirm, you can check the Comet ML experiment logs or the training script to see which dataset was used for these metrics.

      2. Validation and Testing Sets: The metrics = model.val(data="path/to/dataset") line of code uses the validation set if it exists. If no validation set is provided, it defaults to using the test set. To explicitly validate or test your model, ensure your dataset is organized as follows:

        dataset/
        ├── train/
        │   ├── class1/
        │   ├── class2/
        │   └── ...
        ├── test/
        │   ├── class1/
        │   ├── class2/
        │   └── ...
        └── val/ (optional)
            ├── class1/
            ├── class2/
            └── ...

        If you want to validate using the test set, you can specify the split argument in the val method:

        metrics = model.val(data="path/to/dataset", split='test')

        This ensures that the model uses the test set for validation.

      If you encounter any issues, please ensure you are using the latest version of the Ultralytics packages. Providing a minimum reproducible example can greatly help in diagnosing the issue. You can find more details on creating a reproducible example here.

      Feel free to reach out if you have any more questions. Happy training! 🚀

    4. casperking666GitHub

      I see, thank you! I have tried to train from scratch using the yaml file like the following:

      model = YOLO("yolov8n-cls.yaml")

      but I got the following error

      Traceback (most recent call last):
        File "/user/work/yf20630/cow-dataset-project/train.py", line 27, in <module>
          model.save('./models/yolov8s-cow-crops-2080-1e.pt')
        File "/user/work/yf20630/miniconda3/envs/cow_project/lib/python3.10/site-packages/ultralytics/engine/model.py", line 335, in save
          torch.save({**self.ckpt, **updates}, filename, use_dill=use_dill)
      TypeError: 'NoneType' object is not a mapping

      on this line

      model.save('./models/yolov8s-cow-crops-2080-1e.pt')

      the saving works perfectly fine when using the pretrained model, but wouldn't work here, any thoughts on why?

    5. Glenn JocherUltralytics(edited)

      Hi @casperking666,

      Thank you for reaching out and providing the details of your issue!

      It looks like you're encountering an error when trying to save a model that was initialized from a YAML configuration file. The error message TypeError: 'NoneType' object is not a mapping suggests that there might be an issue with the model's checkpoint data.

      Here are a few steps to help resolve this:

      1. Ensure Proper Initialization: When initializing a model from a YAML file, make sure that the YAML file is correctly configured and that the model is properly built before attempting to save it. Here’s an example of how to initialize and train the model:

        from ultralytics import YOLO
        
        # Initialize a model from a YAML file
        model = YOLO("yolo26n-cls.yaml")
        
        # Train the model
        results = model.train(data="path/to/dataset", epochs=100, imgsz=640)
        
        # Save the trained model
        model.save('./models/yolov8s-cow-crops-2080-1e.pt')
      2. Verify Training Completion: Ensure that the training process completes successfully before saving the model. If the training is interrupted or fails, the model's checkpoint data might not be properly populated.

      3. Check for Updates: Verify that you are using the latest version of the Ultralytics package. Sometimes, issues are resolved in newer releases. You can update the package using:

        pip install --upgrade ultralytics
      4. Provide a Reproducible Example: If the issue persists, it would be helpful to provide a minimum reproducible example. This will allow us to better understand the context and provide more accurate assistance. You can find guidelines for creating a reproducible example here.

      By following these steps, you should be able to identify and resolve the issue. If you continue to experience problems, please share more details or a reproducible example, and we’ll be happy to assist further.

      Best of luck with your training! 🚀

      Feel free to reach out if you have any more questions.

    6. MrLi-creatorGitHub

      This is my own small_data.yaml file: path: C:/Users/AIX1WX/Documents/10-POC/train_yolov8/yolov8-all-in-one-main_classification/datasets train: train/ val: val/ test: test/

      nc: 7 names: ['ABS','Airbag','Background','EPB','EPBGREEN','EPBRed','QMtt']

      class_weights: [1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 3.0]

      And my train coda here: from ultralytics import YOLO import os os.environ["KMP_DUPLICATE_LIB_OK"]="TRUE" if name == 'main': model = YOLO("yolov8n-cls.pt") # load a pretrained model (recommended for training) results = model.train(data="small_data.yaml", epochs=30, imgsz=50, device=None, batch=16)

      Why will report the error: Dataset not found , missing path C:\Users\AIX1WX\Documents\10-POC\train_yolov8\yolov8-all-in-one-main_detection\datasets\small_data.yaml, attempting download... Downloading https://github.com/ultralytics/yolov5/releases/download/v1.0/small_data.yaml.zip to 'C:\Users\AIX1WX\Documents\10-POC\train_yolov8\yolov8-all-in-one-main_detection\datasets\small_data.yaml.zip'...

    7. Glenn JocherUltralytics

      @MrLi-creator it looks like the error is due to the dataset path not being correctly recognized. Ensure that the path in your small_data.yaml file is correct and accessible. The path should be absolute or relative to your current working directory. Also, make sure the dataset structure follows the required format with train, val, and test directories. If the issue persists, please verify it with the latest version of the Ultralytics package. If you need further assistance, feel free to ask!

  9. mopa91GitHub

    Hello, I have trained a classification model with my custom dataset. Now, I want to validate the model on another dataset. How can I use this: validation_results = model.val(), to validate the model's predictions and get the confusion matrix for the predicted data?

    1. Paula DerrengerUltralytics

      Hello,

      Great to hear that you've trained a classification model with your custom dataset! To validate your model on another dataset and obtain a confusion matrix, you can use the model.val() method. Here's a concise example:

      from ultralytics import YOLO
      
      # Load your trained model
      model = YOLO("path/to/your/model.pt")
      
      # Validate the model on a new dataset
      validation_results = model.val(data="path/to/validation/dataset")
      
      # Access the confusion matrix
      confusion_matrix = validation_results.confusion_matrix
      print(confusion_matrix)

      Ensure your validation dataset follows the required directory structure. For more details on dataset organization and validation, you can refer to the Image Classification Datasets Overview.

      Happy validating! 🚀

    2. momint2020GitHub

      Hello,

      Thank you very much for your reply. It was a great help. Could you please tell me how I can get the F1 score curve for my classification model on the custom dataset?

    3. Paula DerrengerUltralytics

      @momint2020 hello,

      I'm glad to hear the previous information was helpful! To generate an F1 score curve for your classification model on a custom dataset, you can use the val() method to perform validation and then analyze the predictions to calculate F1 scores at various thresholds.

      Here’s a basic outline of how you might approach this:

      1. Validate the Model: First, run validation to get the predictions.

        from ultralytics import YOLO
        
        # Load your trained model
        model = YOLO("path/to/your/model.pt")
        
        # Validate the model on your custom dataset
        validation_results = model.val(data="path/to/your/validation/dataset")
      2. Calculate F1 Scores: You'll need to write a custom function or script to calculate the F1 score for each class at different thresholds, using the predictions obtained. This involves calculating precision and recall for each threshold and then using these to calculate the F1 score.

      3. Plot the Curve: Use a plotting library like matplotlib to plot the F1 scores against the thresholds.

      Unfortunately, direct support for plotting F1 curves isn't built into the Ultralytics framework, so this will require some custom implementation on your part.

      If you need further guidance on the specifics of calculating and plotting the F1 score, please let me know!