[[["易于理解","easyToUnderstand","thumb-up"],["解决了我的问题","solvedMyProblem","thumb-up"],["其他","otherUp","thumb-up"]],[["没有我需要的信息","missingTheInformationINeed","thumb-down"],["太复杂/步骤太多","tooComplicatedTooManySteps","thumb-down"],["内容需要更新","outOfDate","thumb-down"],["翻译问题","translationIssue","thumb-down"],["示例/代码问题","samplesCodeIssue","thumb-down"],["其他","otherDown","thumb-down"]],["最后更新时间 (UTC)：2025-07-24。"],[],[],null,["MediaPipe Model Maker is a tool for customizing existing machine learning (ML)\nmodels to work with your data and applications. You can use this tool as a\nfaster alternative to building and training a new ML model. Model Maker uses an\nML training technique called\n[transfer learning](https://en.wikipedia.org/wiki/Transfer_learning) which\nretrains existing models with new data. This technique re-uses a significant\nportion of the existing model logic, which means training takes less time than\ntraining a new model, and can be done with less data.\n\nModel Maker works on various types of models including, object detection,\ngesture recognition, or classifiers for images, text, or audio data. The tool\nretrains models by removing the last few layers of the model that classify data\ninto specific categories, and rebuilds those layers using new data you provide.\nModel Maker also supports some option to fine tune model layers to improve\naccuracy and performance.\n\n**Figure 1. Model Maker removes the final layers of an existing model and\nrebuilds them with new data.**\n\nRetraining a model using Model Maker generally makes the model smaller,\nparticularly if you retrain the new model to recognize fewer things. This\nmeans you can use Model Maker to create more focused models that work better for\nyour application. The tool can also help you apply ML techniques like\nquantization so your model uses less resources and runs more efficiently.\n\nTraining data requirements\n\nYou can use Model Maker to retrain models with significantly less data than\ntraining a new model. When retraining a model with new data, you should aim to\nhave approximately 100 data samples for each trained class. For example, if you\nare retraining an image classification model to recognize cats, dogs, and\nparrots, you should have around 100 images of cats, 100 images of dogs, and 100\nimages of parrots. Depending on your application, you may be able to retrain a\nuseful model with even less data per category, although a larger dataset\ngenerally improves the accuracy of your model. When creating your training\ndataset, remember that your training data gets split up during the retraining\nprocess, typically 80% for training, 10% for testing, and the remainder for\nvalidation.\n\nLimitations of customization\n\nSince the retraining process removes the previous classification layers, the\nresulting model can only recognize items, or classes, provided in the new data.\nIf the old model was trained to recognize 30 different item classes, and you use\nModel Maker to retrain with data for 10 different items, the resulting model is\nonly able to recognize those 10 new items.\n\nRetraining a model with Model Maker cannot change what the original ML model\nwas built to solve, even if those jobs are similar. For example, you can't use\nthe tool to make an image classification model perform object detection, even\nthough those tasks share some similarity.\n\nGet started\n\nYou can start using MediaPipe Model Maker by running one of the solution\nCustomization tutorials for MediaPipe Solutions, such as\n[Image Classification](/edge/mediapipe/solutions/customization/image_classifier)"]]
