🕷️ Crawler Inspector

URL Lookup

Direct Parameter Lookup

Raw Queries and Responses

1. Shard Calculation

Query:

Response:

Calculated Shard: 163 (from laksa132)

2. Crawled Status Check

Query:

curl -X POST \
  'http://laksa163.int.ahrefs:8124/' \
  -H 'Content-Type: text/plain' \
  -H 'X-ClickHouse-Database: crawler3' \
  -H 'Authorization: Basic YXBpOg==' \
  -d 'SELECT getAhrefsURLFromUnparsed(src_unparsed) AS found_url, ifNull(toUnixTimestamp(download_stamp), 0) AS crawl_time, ifNull(toUnixTimestamp(props_url_first_seen), 0) AS first_indexed_time, download_http_code AS http_code, src_unparsed AS src_unparsed, src_root_hash AS src_root_hash, history_drop_reason AS history_drop_reason, meta_title AS meta_title, meta_descriptions AS meta_descriptions, attrs_boilerpipe_text AS attrs_boilerpipe_text, attrs_markdown AS attrs_markdown, attrs_readable_markdown AS attrs_readable_markdown, meta_canonical AS meta_canonical FROM crawler3.page_info_local FINAL PREWHERE (src_root_hash, src_unparsed) IN ((getAhrefsRootHashFromUnparsed(getAhrefsUnparsedNoserviceFromURL(\'https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html\')), getAhrefsUnparsedNoserviceFromURL(\'https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html\'))) FORMAT JSONEachRow'

Response:

{"found_url":"https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html","crawl_time":1774375737,"first_indexed_time":1708269487,"http_code":200,"src_unparsed":"ai,flower!\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html s443","src_root_hash":"4426631000825432763","history_drop_reason":null,"meta_title":"快速入门 TensorFlow - Flower Framework","meta_descriptions":["Learn how to train a Convolutional Neural Network on CIFAR-10 using federated learning with Flower and TensorFlow in this step-by-step tutorial."],"attrs_boilerpipe_text":"In this tutorial we will learn how to train a Convolutional Neural Network on CIFAR-10\nusing the Flower framework and TensorFlow. First of all, it is recommended to create a\nvirtual environment and run everything within a\nvirtualenv\n.\nLet's use\nflwr\nnew\nto create a complete Flower+TensorFlow project. It will generate\nall the files needed to run a federation of 10 nodes using\nFedAvg\n. By default,\nthe generated app uses a local simulation profile that\nflwr\nrun\nsubmits to a managed\nlocal SuperLink, which then executes the run with the Flower Simulation Runtime. The\ndataset will be partitioned using Flower Dataset's\nIidPartitioner\n.\nNow that we have a rough idea of what this example is about, let's get started. First,\ninstall Flower in your new environment:\n# In a new Python environment\n$\npip\ninstall\nflwr\n[\nsimulation\n]\nThen, run the command below:\n$\nflwr\nnew\n@flwrlabs\/quickstart-tensorflow\nAfter running it you'll notice a new directory named\nquickstart-tensorflow\nhas been\ncreated. It should have the following structure:\nquickstart-tensorflow\n├──\ntfexample\n│\n├──\n__init__.py\n│\n├──\nclient_app.py\n# Defines your ClientApp\n│\n├──\nserver_app.py\n# Defines your ServerApp\n│\n└──\ntask.py\n# Defines your model, training and data loading\n├──\npyproject.toml\n# Project metadata like dependencies and configs\n└──\nREADME.md\nIf you haven't yet installed the project and its dependencies, you can do so by:\n# From the directory where your pyproject.toml is\n$\npip\ninstall\n-e\n.\nTo run the project, do:\n# Run with default arguments and stream logs\n$\nflwr\nrun\n.\n--stream\nPlain\nflwr\nrun\n.\nsubmits the run, prints the run ID, and returns without streaming\nlogs. For the full local workflow, see\nRun Flower Locally with a Managed SuperLink\n.\nWith default arguments you will see streamed output like this:\nSuccessfully\nbuilt\nflwrlabs.quickstart-tensorflow.1-0-0.014c8eb3.fab\nStarting\nlocal\nSuperLink\non\n127\n.0.0.1:39093...\nSuccessfully\nstarted\nrun\n1859953118041441032\nINFO\n:\nStarting\nFedAvg\nstrategy:\nINFO\n:\n├──\nNumber\nof\nrounds:\n3\nINFO\n:\n[\nROUND\n1\n\/3\n]\nINFO\n:\nconfigure_train:\nSampled\n5\nnodes\n(\nout\nof\n10\n)\nINFO\n:\naggregate_train:\nReceived\n5\nresults\nand\n0\nfailures\nINFO\n:\n└──>\nAggregated\nMetricRecord:\n{\n'train_loss'\n:\n2\n.0013,\n'train_acc'\n:\n0\n.2624\n}\nINFO\n:\nconfigure_evaluate:\nSampled\n10\nnodes\n(\nout\nof\n10\n)\nINFO\n:\naggregate_evaluate:\nReceived\n10\nresults\nand\n0\nfailures\nINFO\n:\n└──>\nAggregated\nMetricRecord:\n{\n'eval_acc'\n:\n0\n.1216,\n'eval_loss'\n:\n2\n.2686\n}\nINFO\n:\n[\nROUND\n2\n\/3\n]\nINFO\n:\n...\nINFO\n:\n[\nROUND\n3\n\/3\n]\nINFO\n:\n...\nINFO\n:\nStrategy\nexecution\nfinished\nin\n16\n.60s\nINFO\n:\nFinal\nresults:\nINFO\n:\nServerApp-side\nEvaluate\nMetrics:\nINFO\n:\n{}\nSaving\nfinal\nmodel\nto\ndisk\nas\nfinal_model.keras...\nYou can also override the parameters defined in the\n[tool.flwr.app.config]\nsection\nin\npyproject.toml\nlike this:\n# Override some arguments\n$\nflwr\nrun\n.\n--run-config\n\"num-server-rounds=5 batch-size=16\"\nThe Data\n¶\nThis tutorial uses\nFlower Datasets\nto easily\ndownload and partition the\nCIFAR-10\ndataset. In this example you'll make use of the\nIidPartitioner\nto generate\nnum_partitions\npartitions. You can choose\nother partitioners\navailable in\nFlower Datasets. Each\nClientApp\nwill call this function to create the\nNumPy\narrays that correspond to their data partition.\npartitioner\n=\nIidPartitioner\n(\nnum_partitions\n=\nnum_partitions\n)\nfds\n=\nFederatedDataset\n(\ndataset\n=\n\"uoft-cs\/cifar10\"\n,\npartitioners\n=\n{\n\"train\"\n:\npartitioner\n},\n)\npartition\n=\nfds\n.\nload_partition\n(\npartition_id\n,\n\"train\"\n)\npartition\n.\nset_format\n(\n\"numpy\"\n)\n# Divide data on each node: 80% train, 20% test\npartition\n=\npartition\n.\ntrain_test_split\n(\ntest_size\n=\n0.2\n)\nx_train\n,\ny_train\n=\npartition\n[\n\"train\"\n][\n\"img\"\n]\n\/\n255.0\n,\npartition\n[\n\"train\"\n][\n\"label\"\n]\nx_test\n,\ny_test\n=\npartition\n[\n\"test\"\n][\n\"img\"\n]\n\/\n255.0\n,\npartition\n[\n\"test\"\n][\n\"label\"\n]\nThe Model\n¶\nNext, we need a model. We defined a simple Convolutional Neural Network (CNN), but feel\nfree to replace it with a more sophisticated model if you'd like:\ndef\nload_model\n(\nlearning_rate\n:\nfloat\n=\n0.001\n):\n# Define a simple CNN for CIFAR-10 and set Adam optimizer\nmodel\n=\nkeras\n.\nSequential\n(\n[\nkeras\n.\nInput\n(\nshape\n=\n(\n32\n,\n32\n,\n3\n)),\nlayers\n.\nConv2D\n(\n32\n,\nkernel_size\n=\n(\n3\n,\n3\n),\nactivation\n=\n\"relu\"\n),\nlayers\n.\nMaxPooling2D\n(\npool_size\n=\n(\n2\n,\n2\n)),\nlayers\n.\nConv2D\n(\n64\n,\nkernel_size\n=\n(\n3\n,\n3\n),\nactivation\n=\n\"relu\"\n),\nlayers\n.\nMaxPooling2D\n(\npool_size\n=\n(\n2\n,\n2\n)),\nlayers\n.\nFlatten\n(),\nlayers\n.\nDropout\n(\n0.5\n),\nlayers\n.\nDense\n(\n10\n,\nactivation\n=\n\"softmax\"\n),\n]\n)\noptimizer\n=\nkeras\n.\noptimizers\n.\nAdam\n(\nlearning_rate\n)\nmodel\n.\ncompile\n(\noptimizer\n=\noptimizer\n,\nloss\n=\n\"sparse_categorical_crossentropy\"\n,\nmetrics\n=\n[\n\"accuracy\"\n],\n)\nreturn\nmodel\nThe ClientApp\n¶\nThe main changes we have to make to use\nTensorflow\nwith\nFlower\nhave to do with\nconverting the\nArrayRecord\nreceived in the\nMessage\ninto numpy ndarrays\nfor use with the built-in\nset_weights()\nfunction. After training, the\nget_weights()\nfunction can be used to extract then pack the updated numpy ndarrays\ninto a\nMessage\nfrom the ClientApp. We can make use of built-in methods in the\nArrayRecord\nto make these conversions:\n@app\n.\ntrain\n()\ndef\ntrain\n(\nmsg\n:\nMessage\n,\ncontext\n:\nContext\n):\n# Load the model\nmodel\n=\nload_model\n(\ncontext\n.\nrun_config\n[\n\"learning-rate\"\n])\n# Extract the ArrayRecord from Message and convert to numpy ndarrays\nmodel\n.\nset_weights\n(\nmsg\n.\ncontent\n[\n\"arrays\"\n]\n.\nto_numpy_ndarrays\n())\n# Train the model\n...\n# Pack the model weights into an ArrayRecord\nmodel_record\n=\nArrayRecord\n(\nmodel\n.\nget_weights\n())\nThe rest of the functionality is directly inspired by the centralized case. The\nClientApp\ncomes with three core methods (\ntrain\n,\nevaluate\n, and\nquery\n)\nthat we can implement for different purposes. For example:\ntrain\nto train the\nreceived model using the local data;\nevaluate\nto assess its performance of the\nreceived model on a validation set; and\nquery\nto retrieve information about the node\nexecuting the\nClientApp\n. In this tutorial we will only make use of\ntrain\nand\nevaluate\n.\nLet's see how the\ntrain\nmethod can be implemented. It receives as input arguments a\nMessage\nfrom the\nServerApp\n. By default it carries:\nan\nArrayRecord\nwith the arrays of the model to federate. By default they can be\nretrieved with key\n\"arrays\"\nwhen accessing the message content.\na\nConfigRecord\nwith the configuration sent from the\nServerApp\n. By default it\ncan be retrieved with key\n\"config\"\nwhen accessing the message content.\nThe\ntrain\nmethod also receives the\nContext\n, giving access to configs for your\nrun and node. The run config hyperparameters are defined in the\npyproject.toml\nof\nyour Flower App. The node config can only be set when running Flower with the Deployment\nRuntime and is not directly configurable during simulations.\n# Flower ClientApp\napp\n=\nClientApp\n()\n@app\n.\ntrain\n()\ndef\ntrain\n(\nmsg\n:\nMessage\n,\ncontext\n:\nContext\n):\n\"\"\"Train the model on local data.\"\"\"\n# Reset local Tensorflow state\nkeras\n.\nbackend\n.\nclear_session\n()\n# Load the data\npartition_id\n=\ncontext\n.\nnode_config\n[\n\"partition-id\"\n]\nnum_partitions\n=\ncontext\n.\nnode_config\n[\n\"num-partitions\"\n]\nx_train\n,\ny_train\n,\n_\n,\n_\n=\nload_data\n(\npartition_id\n,\nnum_partitions\n)\n# Load the model\nmodel\n=\nload_model\n(\ncontext\n.\nrun_config\n[\n\"learning-rate\"\n])\nmodel\n.\nset_weights\n(\nmsg\n.\ncontent\n[\n\"arrays\"\n]\n.\nto_numpy_ndarrays\n())\nepochs\n=\ncontext\n.\nrun_config\n[\n\"local-epochs\"\n]\nbatch_size\n=\ncontext\n.\nrun_config\n[\n\"batch-size\"\n]\nverbose\n=\ncontext\n.\nrun_config\n.\nget\n(\n\"verbose\"\n)\n# Train the model\nhistory\n=\nmodel\n.\nfit\n(\nx_train\n,\ny_train\n,\nepochs\n=\nepochs\n,\nbatch_size\n=\nbatch_size\n,\nverbose\n=\nverbose\n,\n)\n# Get training metrics\ntrain_loss\n=\nhistory\n.\nhistory\n[\n\"loss\"\n][\n-\n1\n]\nif\n\"loss\"\nin\nhistory\n.\nhistory\nelse\nNone\ntrain_acc\n=\n(\nhistory\n.\nhistory\n[\n\"accuracy\"\n][\n-\n1\n]\nif\n\"accuracy\"\nin\nhistory\n.\nhistory\nelse\nNone\n)\n# Pack and send the model weights and metrics as a message\nmodel_record\n=\nArrayRecord\n(\nmodel\n.\nget_weights\n())\nmetrics\n=\n{\n\"num-examples\"\n:\nlen\n(\nx_train\n)}\nif\ntrain_loss\nis\nnot\nNone\n:\nmetrics\n[\n\"train_loss\"\n]\n=\ntrain_loss\nif\ntrain_acc\nis\nnot\nNone\n:\nmetrics\n[\n\"train_acc\"\n]\n=\ntrain_acc\ncontent\n=\nRecordDict\n({\n\"arrays\"\n:\nmodel_record\n,\n\"metrics\"\n:\nMetricRecord\n(\nmetrics\n)})\nreturn\nMessage\n(\ncontent\n=\ncontent\n,\nreply_to\n=\nmsg\n)\nThe\n@app.evaluate()\nmethod would be near identical with two exceptions: (1) the\nmodel is not locally trained, instead it is used to evaluate its performance on the\nlocally held-out validation set; (2) including the model in the reply Message is no\nlonger needed because it is not locally modified.\nThe ServerApp\n¶\nTo construct a\nServerApp\nwe define its\n@app.main()\nmethod. This method\nreceive as input arguments:\na\nGrid\nobject that will be used to interface with the nodes running the\nClientApp\nto involve them in a round of train\/evaluate\/query or other.\na\nContext\nobject that provides access to the run configuration.\nIn this example we use the\nFedAvg\nand configure it with a specific value of\nfraction_train\nwhich is read from the run config. You can find the default value\ndefined in the\npyproject.toml\n. Then, the execution of the strategy is launched when\ninvoking its\nstart\nmethod. To it we pass:\nthe\nGrid\nobject.\nan\nArrayRecord\ncarrying a randomly initialized model that will serve as the global\nmodel to federated.\nthe\nnum_rounds\nparameter specifying how many rounds of\nFedAvg\nto perform.\n# Create the ServerApp\napp\n=\nServerApp\n()\n@app\n.\nmain\n()\ndef\nmain\n(\ngrid\n:\nGrid\n,\ncontext\n:\nContext\n)\n->\nNone\n:\n\"\"\"Main entry point for the ServerApp.\"\"\"\n# Load config\nnum_rounds\n=\ncontext\n.\nrun_config\n[\n\"num-server-rounds\"\n]\nfraction_train\n=\ncontext\n.\nrun_config\n[\n\"fraction-train\"\n]\n# Load initial model\nmodel\n=\nload_model\n()\narrays\n=\nArrayRecord\n(\nmodel\n.\nget_weights\n())\n# Define and start FedAvg strategy\nstrategy\n=\nFedAvg\n(\nfraction_train\n=\nfraction_train\n,\n)\nresult\n=\nstrategy\n.\nstart\n(\ngrid\n=\ngrid\n,\ninitial_arrays\n=\narrays\n,\nnum_rounds\n=\nnum_rounds\n,\n)\n# Save the final model\nndarrays\n=\nresult\n.\narrays\n.\nto_numpy_ndarrays\n()\nfinal_model_name\n=\n\"final_model.keras\"\nprint\n(\nf\n\"Saving final model to disk as\n{\nfinal_model_name\n}\n...\"\n)\nmodel\n.\nset_weights\n(\nndarrays\n)\nmodel\n.\nsave\n(\nfinal_model_name\n)\nNote the\nstart\nmethod of the strategy returns a result object. This object contains\nall the relevant information about the FL process, including the final model weights as\nan\nArrayRecord\n, and federated training and evaluation metrics as\nMetricRecords\n.\nYou can easily log the metrics using Python's\npprint\nand save the final model weights using\nTensorflow's\nsave()\nfunction.\nCongratulations! You've successfully built and run your first federated learning system.\nTip\nCheck the\n运行模拟\ndocumentation to learn more about how to\nconfigure and run Flower simulations.","attrs_markdown":"Hide navigation sidebar\n\nHide table of contents sidebar\n\n[Skip to content](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#furo-main-content)\n\nToggle site navigation sidebar\n\n[Flower Framework](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/index.html)\n\nToggle Light \/ Dark \/ Auto color theme\n\nToggle table of contents sidebar\n\n[![Light Logo](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/_static\/flower-logo-light.png) ![Dark Logo](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/_static\/flower-logo-dark.png) main](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/index.html)\n\n教程\n\n- [什么是联邦学习？](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-series-what-is-federated-learning.html)\n- [开始使用Flower](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-series-get-started-with-flower-pytorch.html)\n- [使用联邦学习策略](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-series-use-a-federated-learning-strategy-pytorch.html)\n- [Customize a Flower Strategy](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-series-build-a-strategy-from-scratch-pytorch.html)\n- [Communicate custom Messages](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-series-customize-the-client-pytorch.html)\n\n- [快速入门教程](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart.html)\n  Toggle navigation of 快速入门教程\n  - [PyTorch快速入门](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-pytorch.html)\n  - [快速入门 TensorFlow](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html)\n  - [Quickstart MLX](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-mlx.html)\n  - [🤗 Transformers快速入门](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-huggingface.html)\n  - [快速入门 JAX](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-jax.html)\n  - [快速入门Pandas](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-pandas.html)\n  - [快速入门 fastai](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-fastai.html)\n  - [快速入门 PyTorch Lightning](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-pytorch-lightning.html)\n  - [scikit-learn快速入门](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-scikitlearn.html)\n  - [XGBoost快速入门](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-xgboost.html)\n  - [快速入门 Android](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-android.html)\n  - [快速入门 iOS](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-ios.html)\n\nHow-to Guides\n\n- [Build](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/build.html)\n  Toggle navigation of Build\n  - [安装Flower](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-install-flower.html)\n  - [Configure `pyproject.toml`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-configure-pyproject-toml.html)\n  - [Configure a `ClientApp`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-configure-clients.html)\n  - [Design stateful ClientApps](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-design-stateful-clients.html)\n  - [使用策略](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-use-strategies.html)\n  - [实施策略](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-implement-strategies.html)\n  - [整合评估结果](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-aggregate-evaluation-results.html)\n  - [Save and load model checkpoints](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-save-and-load-model-checkpoints.html)\n  - [Use Built-in Mods](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-use-built-in-mods.html)\n  - [Use Differential Privacy](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-use-differential-privacy.html)\n  - [Implement FedBN](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-implement-fedbn.html)\n  - [Use CLI JSON output](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-use-cli-json-output.html)\n  - [OpenFL Migration Guide](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-migrate-from-openfl.html)\n  - [升级至 Flower 1.0](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-upgrade-to-flower-1.0.html)\n  - [Upgrade to Flower 1.13](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-upgrade-to-flower-1.13.html)\n  - [Upgrade to Message API](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-upgrade-to-message-api.html)\n- [Simulate](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/simulate.html)\n  Toggle navigation of Simulate\n  - [Run Flower Locally with a Managed SuperLink](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-run-flower-locally.html)\n  - [运行模拟](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-run-simulations.html)\n- [Deploy](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/deploy.html)\n  Toggle navigation of Deploy\n  - [Run Flower with the Deployment Runtime](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-run-flower-with-deployment-engine.html)\n  - [Enable TLS connections](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-enable-tls-connections.html)\n  - [Authenticate SuperNodes](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-authenticate-supernodes.html)\n  - [Configure logging](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-configure-logging.html)\n  - [Run Flower on GCP](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-run-flower-on-gcp.html)\n  - [Run Flower on Azure](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-run-flower-on-azure.html)\n  - [Run Flower on Red Hat OpenShift](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-run-flower-on-red-hat-openshift.html)\n  - [Run Flower on Multiple OpenShift Clusters](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-run-flower-on-multiple-openshift-clusters.html)\n  - [Authenticate Accounts via OpenID Connect](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-authenticate-accounts.html)\n  - [Configure Audit Logging](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-configure-audit-logging.html)\n  - [Manage Flower Federations](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-manage-flower-federations.html)\n  - [Run Flower using Docker](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/index.html)\n    Toggle navigation of Run Flower using Docker\n    - [Quickstart with Docker](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/tutorial-quickstart-docker.html)\n    - [Enable TLS for Secure Connections](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/enable-tls.html)\n    - [Persist the State of the SuperLink](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/persist-superlink-state.html)\n    - [Set Environment Variables](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/set-environment-variables.html)\n    - [Run with Root User Privileges](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/run-as-root-user.html)\n    - [Run ServerApp or ClientApp as a Subprocess](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/run-as-subprocess.html)\n    - [Pin a Docker Image to a Specific Version](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/pin-version.html)\n    - [Use a Different Flower Version](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/use-a-different-version.html)\n    - [Quickstart with Docker Compose](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/tutorial-quickstart-docker-compose.html)\n    - [Run Flower Quickstart Examples with Docker Compose](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/run-quickstart-examples-docker-compose.html)\n    - [Deploy Flower on Multiple Machines with Docker Compose](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/docker\/tutorial-deploy-on-multiple-machines.html)\n  - [Run Flower using Helm](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/helm\/index.html)\n    Toggle navigation of Run Flower using Helm\n    - [Deploy SuperLink](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/helm\/how-to-deploy-superlink-using-helm.html)\n    - [Deploy SuperNode](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/helm\/how-to-deploy-supernode-using-helm.html)\n\n说明\n\n- [联邦学习评估](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/explanation-federated-evaluation.html)\n- [差分隐私](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/explanation-differential-privacy.html)\n- [安全聚合协议](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/explanation-ref-secure-aggregation-protocols.html)\n- [Flower的架构](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/explanation-flower-architecture.html)\n- [Flower Strategy Abstraction](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/explanation-flower-strategy-abstraction.html)\n\n参考资料\n\n- [Reference](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/reference.html)\n  Toggle navigation of Reference\n  - [flwr](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.html)\n    Toggle navigation of flwr\n    - [app](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.html)\n      Toggle navigation of app\n      - [Array](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.Array.html)\n      - [ArrayRecord](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.ArrayRecord.html)\n      - [ConfigRecord](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.ConfigRecord.html)\n      - [Context](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.Context.html)\n      - [Error](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.Error.html)\n      - [Message](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.Message.html)\n      - [MessageType](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.MessageType.html)\n      - [描述数据](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.Metadata.html)\n      - [MetricRecord](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.MetricRecord.html)\n      - [RecordDict](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.RecordDict.html)\n      - [UserConfig](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.UserConfig.html)\n    - [clientapp](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.clientapp.html)\n      Toggle navigation of clientapp\n      - [ClientApp](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.clientapp.ClientApp.html)\n      - [mod](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.clientapp.mod.html)\n        Toggle navigation of mod\n        - [adaptiveclipping\\_mod](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.clientapp.mod.adaptiveclipping_mod.html)\n        - [arrays\\_size\\_mod](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.clientapp.mod.arrays_size_mod.html)\n        - [fixedclipping\\_mod](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.clientapp.mod.fixedclipping_mod.html)\n        - [message\\_size\\_mod](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.clientapp.mod.message_size_mod.html)\n        - [LocalDpMod](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.clientapp.mod.LocalDpMod.html)\n    - [serverapp](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.html)\n      Toggle navigation of serverapp\n      - [Grid](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.Grid.html)\n      - [ServerApp](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.ServerApp.html)\n      - [strategy](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.html)\n        Toggle navigation of strategy\n        - [Bulyan](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.Bulyan.html)\n        - [DifferentialPrivacyClientSideAdaptiveClipping](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.DifferentialPrivacyClientSideAdaptiveClipping.html)\n        - [DifferentialPrivacyClientSideFixedClipping](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.DifferentialPrivacyClientSideFixedClipping.html)\n        - [DifferentialPrivacyServerSideAdaptiveClipping](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.DifferentialPrivacyServerSideAdaptiveClipping.html)\n        - [DifferentialPrivacyServerSideFixedClipping](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.DifferentialPrivacyServerSideFixedClipping.html)\n        - [FedAdagrad](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedAdagrad.html)\n        - [FedAdam](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedAdam.html)\n        - [FedAvg](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedAvg.html)\n        - [FedAvgM](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedAvgM.html)\n        - [FedMedian](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedMedian.html)\n        - [FedProx](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedProx.html)\n        - [FedTrimmedAvg](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedTrimmedAvg.html)\n        - [FedXgbBagging](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedXgbBagging.html)\n        - [FedXgbCyclic](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedXgbCyclic.html)\n        - [FedYogi](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedYogi.html)\n        - [Krum](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.Krum.html)\n        - [MultiKrum](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.MultiKrum.html)\n        - [QFedAvg](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.QFedAvg.html)\n        - [Result](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.Result.html)\n        - [Strategy](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.Strategy.html)\n  - [Flower CLI 参考](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api-cli.html)\n  - [Flower Configuration](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-flower-configuration.html)\n  - [项目实例](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-example-projects.html)\n  - [遥测功能](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-telemetry.html)\n  - [更新日志](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-changelog.html)\n  - [Flower Runtime Comparison](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-flower-runtime-comparison.html)\n  - [Flower Network Communication](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-flower-network-communication.html)\n  - [Exit Codes](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes-dir.html)\n    Toggle navigation of Exit Codes\n    - [\\[0\\] SUCCESS](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/0.html)\n    - [\\[1\\] GRACEFUL\\_EXIT\\_SIGINT](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/1.html)\n    - [\\[100\\] SUPERLINK\\_THREAD\\_CRASH](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/100.html)\n    - [\\[101\\] SUPERLINK\\_LICENSE\\_INVALID](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/101.html)\n    - [\\[102\\] SUPERLINK\\_LICENSE\\_MISSING](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/102.html)\n    - [\\[103\\] SUPERLINK\\_LICENSE\\_URL\\_INVALID](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/103.html)\n    - [\\[104\\] SUPERLINK\\_INVALID\\_ARGS](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/104.html)\n    - [\\[105\\] SUPERLINK\\_DATABASE\\_SCHEMA\\_MISMATCH](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/105.html)\n    - [\\[2\\] GRACEFUL\\_EXIT\\_SIGQUIT](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/2.html)\n    - [\\[200\\] SERVERAPP\\_STRATEGY\\_PRECONDITION\\_UNMET](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/200.html)\n    - [\\[201\\] SERVERAPP\\_EXCEPTION](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/201.html)\n    - [\\[202\\] SERVERAPP\\_STRATEGY\\_AGGREGATION\\_ERROR](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/202.html)\n    - [\\[203\\] SERVERAPP\\_RUN\\_START\\_REJECTED](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/203.html)\n    - [\\[3\\] GRACEFUL\\_EXIT\\_SIGTERM](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/3.html)\n    - [\\[300\\] SUPERNODE\\_REST\\_ADDRESS\\_INVALID](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/300.html)\n    - [\\[302\\] SUPERNODE\\_NODE\\_AUTH\\_KEY\\_INVALID](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/302.html)\n    - [\\[303\\] SUPERNODE\\_STARTED\\_WITHOUT\\_TLS\\_BUT\\_NODE\\_AUTH\\_ENABLED](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/303.html)\n    - [\\[304\\] SUPERNODE\\_INVALID\\_TRUSTED\\_ENTITIES](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/304.html)\n    - [\\[400\\] SUPEREXEC\\_INVALID\\_PLUGIN\\_CONFIG](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/400.html)\n    - [\\[500\\] FLWRCLI\\_NODE\\_AUTH\\_PUBLIC\\_KEY\\_INVALID](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/500.html)\n    - [\\[600\\] COMMON\\_ADDRESS\\_INVALID](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/600.html)\n    - [\\[601\\] COMMON\\_MISSING\\_EXTRA\\_REST](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/601.html)\n    - [\\[602\\] COMMON\\_TLS\\_NOT\\_SUPPORTED](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/602.html)\n    - [\\[700\\] SIMULATION\\_EXCEPTION](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/700.html)\n    - [\\[701\\] SIMULATION\\_MISSING\\_EXTRA](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-exit-codes\/701.html)\n  - [常见问题](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-faq.html)\n\n贡献者文档\n\n- [Contribute](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contribute.html)\n  Toggle navigation of Contribute\n  - [在 GitHub 上投稿](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-tutorial-contribute-on-github.html)\n  - [成为贡献者](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-tutorial-get-started-as-a-contributor.html)\n  - [安装开发版本](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-how-to-install-development-versions.html)\n  - [建立虚拟环境](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-how-to-set-up-a-virtual-env.html)\n  - [使用 VSCode Dev Containers 进行开发](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-how-to-develop-in-vscode-dev-containers.html)\n  - [编写文件](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-how-to-write-documentation.html)\n  - [发布 Flower](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-how-to-release-flower.html)\n  - [贡献译文](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-how-to-contribute-translations.html)\n  - [How to Build Docker Flower Images Locally](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-how-to-build-docker-images.html)\n  - [Migrate Flower Database Schema](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-how-to-migrate-database.html)\n  - [Public and private APIs](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-explanation-public-and-private-apis.html)\n  - [首次代码贡献](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-ref-good-first-contributions.html)\n\nVersions\n\n- [main](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html)\n\n[🇬🇧](https:\/\/flower.ai\/docs\/framework\/main\/en\/tutorial-quickstart-tensorflow.html) [🇫🇷](https:\/\/flower.ai\/docs\/framework\/main\/fr\/tutorial-quickstart-tensorflow.html) [🇨🇳](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html) [🇰🇷](https:\/\/flower.ai\/docs\/framework\/main\/ko\/tutorial-quickstart-tensorflow.html)\n\n[Back to top](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html)\n\n[View this page](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/_sources\/tutorial-quickstart-tensorflow.rst.txt \"View this page\")\n\nToggle Light \/ Dark \/ Auto color theme\n\nToggle table of contents sidebar\n\n# 快速入门 TensorFlow[¶](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#id1 \"Link to this heading\")\nIn this tutorial we will learn how to train a Convolutional Neural Network on CIFAR-10 using the Flower framework and TensorFlow. First of all, it is recommended to create a virtual environment and run everything within a [virtualenv](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-how-to-set-up-a-virtual-env.html).\n\nLet's use `flwr new` to create a complete Flower+TensorFlow project. It will generate all the files needed to run a federation of 10 nodes using [`FedAvg`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedAvg.html). By default, the generated app uses a local simulation profile that `flwr run` submits to a managed local SuperLink, which then executes the run with the Flower Simulation Runtime. The dataset will be partitioned using Flower Dataset's [IidPartitioner](https:\/\/flower.ai\/docs\/datasets\/ref-api\/flwr_datasets.partitioner.IidPartitioner.html#flwr_datasets.partitioner.IidPartitioner).\n\nNow that we have a rough idea of what this example is about, let's get started. First, install Flower in your new environment:\n```\n# In a new Python environment\n$ pip install flwr[simulation]\n```\nThen, run the command below:\n```\n$ flwr new @flwrlabs\/quickstart-tensorflow\n```\nAfter running it you'll notice a new directory named `quickstart-tensorflow` has been created. It should have the following structure:\n```\nquickstart-tensorflow\n├── tfexample\n│   ├── __init__.py\n│   ├── client_app.py   # Defines your ClientApp\n│   ├── server_app.py   # Defines your ServerApp\n│   └── task.py         # Defines your model, training and data loading\n├── pyproject.toml      # Project metadata like dependencies and configs\n└── README.md\n```\nIf you haven't yet installed the project and its dependencies, you can do so by:\n```\n# From the directory where your pyproject.toml is\n$ pip install -e .\n```\nTo run the project, do:\n```\n# Run with default arguments and stream logs\n$ flwr run . --stream\n```\nPlain `flwr run .` submits the run, prints the run ID, and returns without streaming logs. For the full local workflow, see [Run Flower Locally with a Managed SuperLink](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-run-flower-locally.html).\n\nWith default arguments you will see streamed output like this:\n```\nSuccessfully built flwrlabs.quickstart-tensorflow.1-0-0.014c8eb3.fab\nStarting local SuperLink on 127.0.0.1:39093...\nSuccessfully started run 1859953118041441032\nINFO :      Starting FedAvg strategy:\nINFO :          ├── Number of rounds: 3\nINFO :      [ROUND 1\/3]\nINFO :      configure_train: Sampled 5 nodes (out of 10)\nINFO :      aggregate_train: Received 5 results and 0 failures\nINFO :          └──> Aggregated MetricRecord: {'train_loss': 2.0013, 'train_acc': 0.2624}\nINFO :      configure_evaluate: Sampled 10 nodes (out of 10)\nINFO :      aggregate_evaluate: Received 10 results and 0 failures\nINFO :          └──> Aggregated MetricRecord: {'eval_acc': 0.1216, 'eval_loss': 2.2686}\nINFO :      [ROUND 2\/3]\nINFO :      ...\nINFO :      [ROUND 3\/3]\nINFO :      ...\nINFO :      Strategy execution finished in 16.60s\nINFO :      Final results:\nINFO :          ServerApp-side Evaluate Metrics:\nINFO :          {}\nSaving final model to disk as final_model.keras...\n```\nYou can also override the parameters defined in the `[tool.flwr.app.config]` section in `pyproject.toml` like this:\n```\n# Override some arguments\n$ flwr run . --run-config \"num-server-rounds=5 batch-size=16\"\n```\n## The Data[¶](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-data \"Link to this heading\")\nThis tutorial uses [Flower Datasets](https:\/\/flower.ai\/docs\/datasets\/) to easily download and partition the CIFAR-10 dataset. In this example you'll make use of the [IidPartitioner](https:\/\/flower.ai\/docs\/datasets\/ref-api\/flwr_datasets.partitioner.IidPartitioner.html#flwr_datasets.partitioner.IidPartitioner) to generate num\\_partitions partitions. You can choose [other partitioners](https:\/\/flower.ai\/docs\/datasets\/ref-api\/flwr_datasets.partitioner.html) available in Flower Datasets. Each `ClientApp` will call this function to create the `NumPy` arrays that correspond to their data partition.\n```\npartitioner = IidPartitioner(num_partitions=num_partitions)\nfds = FederatedDataset(\n    dataset=\"uoft-cs\/cifar10\",\n    partitioners={\"train\": partitioner},\n)\npartition = fds.load_partition(partition_id, \"train\")\npartition.set_format(\"numpy\")\n\n# Divide data on each node: 80% train, 20% test\npartition = partition.train_test_split(test_size=0.2)\nx_train, y_train = partition[\"train\"][\"img\"] \/ 255.0, partition[\"train\"][\"label\"]\nx_test, y_test = partition[\"test\"][\"img\"] \/ 255.0, partition[\"test\"][\"label\"]\n```\n\n## The Model[¶](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-model \"Link to this heading\")\nNext, we need a model. We defined a simple Convolutional Neural Network (CNN), but feel free to replace it with a more sophisticated model if you'd like:\n```\ndef load_model(learning_rate: float = 0.001):\n    # Define a simple CNN for CIFAR-10 and set Adam optimizer\n    model = keras.Sequential(\n        [\n            keras.Input(shape=(32, 32, 3)),\n            layers.Conv2D(32, kernel_size=(3, 3), activation=\"relu\"),\n            layers.MaxPooling2D(pool_size=(2, 2)),\n            layers.Conv2D(64, kernel_size=(3, 3), activation=\"relu\"),\n            layers.MaxPooling2D(pool_size=(2, 2)),\n            layers.Flatten(),\n            layers.Dropout(0.5),\n            layers.Dense(10, activation=\"softmax\"),\n        ]\n    )\n    optimizer = keras.optimizers.Adam(learning_rate)\n    model.compile(\n        optimizer=optimizer,\n        loss=\"sparse_categorical_crossentropy\",\n        metrics=[\"accuracy\"],\n    )\n    return model\n```\n\n## The ClientApp[¶](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-clientapp \"Link to this heading\")\nThe main changes we have to make to use Tensorflow with Flower have to do with converting the [`ArrayRecord`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.ArrayRecord.html) received in the [`Message`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.Message.html) into numpy ndarrays for use with the built-in `set_weights()` function. After training, the `get_weights()` function can be used to extract then pack the updated numpy ndarrays into a `Message` from the ClientApp. We can make use of built-in methods in the `ArrayRecord` to make these conversions:\n```\n@app.train()\ndef train(msg: Message, context: Context):\n\n    # Load the model\n    model = load_model(context.run_config[\"learning-rate\"])\n    # Extract the ArrayRecord from Message and convert to numpy ndarrays\n    model.set_weights(msg.content[\"arrays\"].to_numpy_ndarrays())\n\n    # Train the model\n    ...\n\n    # Pack the model weights into an ArrayRecord\n    model_record = ArrayRecord(model.get_weights())\n```\nThe rest of the functionality is directly inspired by the centralized case. The [`ClientApp`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.clientapp.ClientApp.html) comes with three core methods (`train`, `evaluate`, and `query`) that we can implement for different purposes. For example: `train` to train the received model using the local data; `evaluate` to assess its performance of the received model on a validation set; and `query` to retrieve information about the node executing the `ClientApp`. In this tutorial we will only make use of `train` and `evaluate`.\n\nLet's see how the `train` method can be implemented. It receives as input arguments a [`Message`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.Message.html) from the `ServerApp`. By default it carries:\n\n- an `ArrayRecord` with the arrays of the model to federate. By default they can be retrieved with key `\"arrays\"` when accessing the message content.\n- a `ConfigRecord` with the configuration sent from the `ServerApp`. By default it can be retrieved with key `\"config\"` when accessing the message content.\n\nThe `train` method also receives the `Context`, giving access to configs for your run and node. The run config hyperparameters are defined in the `pyproject.toml` of your Flower App. The node config can only be set when running Flower with the Deployment Runtime and is not directly configurable during simulations.\n```\n# Flower ClientApp\napp = ClientApp()\n\n\n@app.train()\ndef train(msg: Message, context: Context):\n    \"\"\"Train the model on local data.\"\"\"\n\n    # Reset local Tensorflow state\n    keras.backend.clear_session()\n\n    # Load the data\n    partition_id = context.node_config[\"partition-id\"]\n    num_partitions = context.node_config[\"num-partitions\"]\n    x_train, y_train, _, _ = load_data(partition_id, num_partitions)\n\n    # Load the model\n    model = load_model(context.run_config[\"learning-rate\"])\n    model.set_weights(msg.content[\"arrays\"].to_numpy_ndarrays())\n    epochs = context.run_config[\"local-epochs\"]\n    batch_size = context.run_config[\"batch-size\"]\n    verbose = context.run_config.get(\"verbose\")\n\n    # Train the model\n    history = model.fit(\n        x_train,\n        y_train,\n        epochs=epochs,\n        batch_size=batch_size,\n        verbose=verbose,\n    )\n\n    # Get training metrics\n    train_loss = history.history[\"loss\"][-1] if \"loss\" in history.history else None\n    train_acc = (\n        history.history[\"accuracy\"][-1] if \"accuracy\" in history.history else None\n    )\n\n    # Pack and send the model weights and metrics as a message\n    model_record = ArrayRecord(model.get_weights())\n    metrics = {\"num-examples\": len(x_train)}\n    if train_loss is not None:\n        metrics[\"train_loss\"] = train_loss\n    if train_acc is not None:\n        metrics[\"train_acc\"] = train_acc\n    content = RecordDict({\"arrays\": model_record, \"metrics\": MetricRecord(metrics)})\n    return Message(content=content, reply_to=msg)\n```\nThe `@app.evaluate()` method would be near identical with two exceptions: (1) the model is not locally trained, instead it is used to evaluate its performance on the locally held-out validation set; (2) including the model in the reply Message is no longer needed because it is not locally modified.\n\n## The ServerApp[¶](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-serverapp \"Link to this heading\")\nTo construct a [`ServerApp`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.ServerApp.html) we define its `@app.main()` method. This method receive as input arguments:\n\n- a `Grid` object that will be used to interface with the nodes running the `ClientApp` to involve them in a round of train\/evaluate\/query or other.\n- a `Context` object that provides access to the run configuration.\n\nIn this example we use the [`FedAvg`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedAvg.html) and configure it with a specific value of `fraction_train` which is read from the run config. You can find the default value defined in the `pyproject.toml`. Then, the execution of the strategy is launched when invoking its [`start`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.Strategy.html#flwr.serverapp.Strategy.start) method. To it we pass:\n\n- the `Grid` object.\n- an `ArrayRecord` carrying a randomly initialized model that will serve as the global model to federated.\n- the `num_rounds` parameter specifying how many rounds of `FedAvg` to perform.\n```\n# Create the ServerApp\napp = ServerApp()\n\n\n@app.main()\ndef main(grid: Grid, context: Context) -> None:\n    \"\"\"Main entry point for the ServerApp.\"\"\"\n    # Load config\n    num_rounds = context.run_config[\"num-server-rounds\"]\n    fraction_train = context.run_config[\"fraction-train\"]\n\n    # Load initial model\n    model = load_model()\n    arrays = ArrayRecord(model.get_weights())\n\n    # Define and start FedAvg strategy\n    strategy = FedAvg(\n        fraction_train=fraction_train,\n    )\n\n    result = strategy.start(\n        grid=grid,\n        initial_arrays=arrays,\n        num_rounds=num_rounds,\n    )\n\n    # Save the final model\n    ndarrays = result.arrays.to_numpy_ndarrays()\n    final_model_name = \"final_model.keras\"\n    print(f\"Saving final model to disk as {final_model_name}...\")\n    model.set_weights(ndarrays)\n    model.save(final_model_name)\n```\nNote the `start` method of the strategy returns a result object. This object contains all the relevant information about the FL process, including the final model weights as an `ArrayRecord`, and federated training and evaluation metrics as `MetricRecords`. You can easily log the metrics using Python's [pprint](https:\/\/docs.python.org\/3\/library\/pprint.html) and save the final model weights using Tensorflow's `save()` function.\n\nCongratulations! You've successfully built and run your first federated learning system.\n\nTip\n\nCheck the [运行模拟](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-run-simulations.html) documentation to learn more about how to configure and run Flower simulations.\n\nNote\n\nCheck the source code of the extended version of this tutorial in [`examples\/quickstart-tensorflow`](https:\/\/github.com\/flwrlabs\/flower\/blob\/main\/examples\/quickstart-tensorflow) in the Flower GitHub repository.\n\n[Next Quickstart MLX](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-mlx.html)\n\n[Previous PyTorch快速入门](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-pytorch.html)\n\nCopyright © 2026 Flower Labs GmbH\n\nMade with [Sphinx](https:\/\/www.sphinx-doc.org\/) and [@pradyunsg](https:\/\/pradyunsg.me\/)'s [Furo](https:\/\/github.com\/pradyunsg\/furo)\n\nOn this page\n\n- [快速入门 TensorFlow](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html)\n  - [The Data](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-data)\n  - [The Model](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-model)\n  - [The ClientApp](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-clientapp)\n  - [The ServerApp](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-serverapp)","attrs_readable_markdown":"In this tutorial we will learn how to train a Convolutional Neural Network on CIFAR-10 using the Flower framework and TensorFlow. First of all, it is recommended to create a virtual environment and run everything within a [virtualenv](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/contributor-how-to-set-up-a-virtual-env.html).\n\nLet's use `flwr new` to create a complete Flower+TensorFlow project. It will generate all the files needed to run a federation of 10 nodes using [`FedAvg`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedAvg.html). By default, the generated app uses a local simulation profile that `flwr run` submits to a managed local SuperLink, which then executes the run with the Flower Simulation Runtime. The dataset will be partitioned using Flower Dataset's [IidPartitioner](https:\/\/flower.ai\/docs\/datasets\/ref-api\/flwr_datasets.partitioner.IidPartitioner.html#flwr_datasets.partitioner.IidPartitioner).\n\nNow that we have a rough idea of what this example is about, let's get started. First, install Flower in your new environment:\n```\n# In a new Python environment\n$ pip install flwr[simulation]\n```\nThen, run the command below:\n```\n$ flwr new @flwrlabs\/quickstart-tensorflow\n```\nAfter running it you'll notice a new directory named `quickstart-tensorflow` has been created. It should have the following structure:\n```\nquickstart-tensorflow\n├── tfexample\n│   ├── __init__.py\n│   ├── client_app.py   # Defines your ClientApp\n│   ├── server_app.py   # Defines your ServerApp\n│   └── task.py         # Defines your model, training and data loading\n├── pyproject.toml      # Project metadata like dependencies and configs\n└── README.md\n```\nIf you haven't yet installed the project and its dependencies, you can do so by:\n```\n# From the directory where your pyproject.toml is\n$ pip install -e .\n```\nTo run the project, do:\n```\n# Run with default arguments and stream logs\n$ flwr run . --stream\n```\nPlain `flwr run .` submits the run, prints the run ID, and returns without streaming logs. For the full local workflow, see [Run Flower Locally with a Managed SuperLink](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-run-flower-locally.html).\n\nWith default arguments you will see streamed output like this:\n```\nSuccessfully built flwrlabs.quickstart-tensorflow.1-0-0.014c8eb3.fab\nStarting local SuperLink on 127.0.0.1:39093...\nSuccessfully started run 1859953118041441032\nINFO :      Starting FedAvg strategy:\nINFO :          ├── Number of rounds: 3\nINFO :      [ROUND 1\/3]\nINFO :      configure_train: Sampled 5 nodes (out of 10)\nINFO :      aggregate_train: Received 5 results and 0 failures\nINFO :          └──> Aggregated MetricRecord: {'train_loss': 2.0013, 'train_acc': 0.2624}\nINFO :      configure_evaluate: Sampled 10 nodes (out of 10)\nINFO :      aggregate_evaluate: Received 10 results and 0 failures\nINFO :          └──> Aggregated MetricRecord: {'eval_acc': 0.1216, 'eval_loss': 2.2686}\nINFO :      [ROUND 2\/3]\nINFO :      ...\nINFO :      [ROUND 3\/3]\nINFO :      ...\nINFO :      Strategy execution finished in 16.60s\nINFO :      Final results:\nINFO :          ServerApp-side Evaluate Metrics:\nINFO :          {}\nSaving final model to disk as final_model.keras...\n```\nYou can also override the parameters defined in the `[tool.flwr.app.config]` section in `pyproject.toml` like this:\n```\n# Override some arguments\n$ flwr run . --run-config \"num-server-rounds=5 batch-size=16\"\n```\n## The Data[¶](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-data \"Link to this heading\")\nThis tutorial uses [Flower Datasets](https:\/\/flower.ai\/docs\/datasets\/) to easily download and partition the CIFAR-10 dataset. In this example you'll make use of the [IidPartitioner](https:\/\/flower.ai\/docs\/datasets\/ref-api\/flwr_datasets.partitioner.IidPartitioner.html#flwr_datasets.partitioner.IidPartitioner) to generate num\\_partitions partitions. You can choose [other partitioners](https:\/\/flower.ai\/docs\/datasets\/ref-api\/flwr_datasets.partitioner.html) available in Flower Datasets. Each `ClientApp` will call this function to create the `NumPy` arrays that correspond to their data partition.\n```\npartitioner = IidPartitioner(num_partitions=num_partitions)\nfds = FederatedDataset(\n    dataset=\"uoft-cs\/cifar10\",\n    partitioners={\"train\": partitioner},\n)\npartition = fds.load_partition(partition_id, \"train\")\npartition.set_format(\"numpy\")\n\n# Divide data on each node: 80% train, 20% test\npartition = partition.train_test_split(test_size=0.2)\nx_train, y_train = partition[\"train\"][\"img\"] \/ 255.0, partition[\"train\"][\"label\"]\nx_test, y_test = partition[\"test\"][\"img\"] \/ 255.0, partition[\"test\"][\"label\"]\n```\n\n## The Model[¶](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-model \"Link to this heading\")\nNext, we need a model. We defined a simple Convolutional Neural Network (CNN), but feel free to replace it with a more sophisticated model if you'd like:\n```\ndef load_model(learning_rate: float = 0.001):\n    # Define a simple CNN for CIFAR-10 and set Adam optimizer\n    model = keras.Sequential(\n        [\n            keras.Input(shape=(32, 32, 3)),\n            layers.Conv2D(32, kernel_size=(3, 3), activation=\"relu\"),\n            layers.MaxPooling2D(pool_size=(2, 2)),\n            layers.Conv2D(64, kernel_size=(3, 3), activation=\"relu\"),\n            layers.MaxPooling2D(pool_size=(2, 2)),\n            layers.Flatten(),\n            layers.Dropout(0.5),\n            layers.Dense(10, activation=\"softmax\"),\n        ]\n    )\n    optimizer = keras.optimizers.Adam(learning_rate)\n    model.compile(\n        optimizer=optimizer,\n        loss=\"sparse_categorical_crossentropy\",\n        metrics=[\"accuracy\"],\n    )\n    return model\n```\n\n## The ClientApp[¶](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-clientapp \"Link to this heading\")\nThe main changes we have to make to use Tensorflow with Flower have to do with converting the [`ArrayRecord`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.ArrayRecord.html) received in the [`Message`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.Message.html) into numpy ndarrays for use with the built-in `set_weights()` function. After training, the `get_weights()` function can be used to extract then pack the updated numpy ndarrays into a `Message` from the ClientApp. We can make use of built-in methods in the `ArrayRecord` to make these conversions:\n```\n@app.train()\ndef train(msg: Message, context: Context):\n\n    # Load the model\n    model = load_model(context.run_config[\"learning-rate\"])\n    # Extract the ArrayRecord from Message and convert to numpy ndarrays\n    model.set_weights(msg.content[\"arrays\"].to_numpy_ndarrays())\n\n    # Train the model\n    ...\n\n    # Pack the model weights into an ArrayRecord\n    model_record = ArrayRecord(model.get_weights())\n```\nThe rest of the functionality is directly inspired by the centralized case. The [`ClientApp`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.clientapp.ClientApp.html) comes with three core methods (`train`, `evaluate`, and `query`) that we can implement for different purposes. For example: `train` to train the received model using the local data; `evaluate` to assess its performance of the received model on a validation set; and `query` to retrieve information about the node executing the `ClientApp`. In this tutorial we will only make use of `train` and `evaluate`.\n\nLet's see how the `train` method can be implemented. It receives as input arguments a [`Message`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.app.Message.html) from the `ServerApp`. By default it carries:\n\n- an `ArrayRecord` with the arrays of the model to federate. By default they can be retrieved with key `\"arrays\"` when accessing the message content.\n- a `ConfigRecord` with the configuration sent from the `ServerApp`. By default it can be retrieved with key `\"config\"` when accessing the message content.\n\nThe `train` method also receives the `Context`, giving access to configs for your run and node. The run config hyperparameters are defined in the `pyproject.toml` of your Flower App. The node config can only be set when running Flower with the Deployment Runtime and is not directly configurable during simulations.\n```\n# Flower ClientApp\napp = ClientApp()\n\n\n@app.train()\ndef train(msg: Message, context: Context):\n    \"\"\"Train the model on local data.\"\"\"\n\n    # Reset local Tensorflow state\n    keras.backend.clear_session()\n\n    # Load the data\n    partition_id = context.node_config[\"partition-id\"]\n    num_partitions = context.node_config[\"num-partitions\"]\n    x_train, y_train, _, _ = load_data(partition_id, num_partitions)\n\n    # Load the model\n    model = load_model(context.run_config[\"learning-rate\"])\n    model.set_weights(msg.content[\"arrays\"].to_numpy_ndarrays())\n    epochs = context.run_config[\"local-epochs\"]\n    batch_size = context.run_config[\"batch-size\"]\n    verbose = context.run_config.get(\"verbose\")\n\n    # Train the model\n    history = model.fit(\n        x_train,\n        y_train,\n        epochs=epochs,\n        batch_size=batch_size,\n        verbose=verbose,\n    )\n\n    # Get training metrics\n    train_loss = history.history[\"loss\"][-1] if \"loss\" in history.history else None\n    train_acc = (\n        history.history[\"accuracy\"][-1] if \"accuracy\" in history.history else None\n    )\n\n    # Pack and send the model weights and metrics as a message\n    model_record = ArrayRecord(model.get_weights())\n    metrics = {\"num-examples\": len(x_train)}\n    if train_loss is not None:\n        metrics[\"train_loss\"] = train_loss\n    if train_acc is not None:\n        metrics[\"train_acc\"] = train_acc\n    content = RecordDict({\"arrays\": model_record, \"metrics\": MetricRecord(metrics)})\n    return Message(content=content, reply_to=msg)\n```\nThe `@app.evaluate()` method would be near identical with two exceptions: (1) the model is not locally trained, instead it is used to evaluate its performance on the locally held-out validation set; (2) including the model in the reply Message is no longer needed because it is not locally modified.\n\n## The ServerApp[¶](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/tutorial-quickstart-tensorflow.html#the-serverapp \"Link to this heading\")\nTo construct a [`ServerApp`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.ServerApp.html) we define its `@app.main()` method. This method receive as input arguments:\n\n- a `Grid` object that will be used to interface with the nodes running the `ClientApp` to involve them in a round of train\/evaluate\/query or other.\n- a `Context` object that provides access to the run configuration.\n\nIn this example we use the [`FedAvg`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.strategy.FedAvg.html) and configure it with a specific value of `fraction_train` which is read from the run config. You can find the default value defined in the `pyproject.toml`. Then, the execution of the strategy is launched when invoking its [`start`](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/ref-api\/flwr.serverapp.Strategy.html#flwr.serverapp.Strategy.start) method. To it we pass:\n\n- the `Grid` object.\n- an `ArrayRecord` carrying a randomly initialized model that will serve as the global model to federated.\n- the `num_rounds` parameter specifying how many rounds of `FedAvg` to perform.\n```\n# Create the ServerApp\napp = ServerApp()\n\n\n@app.main()\ndef main(grid: Grid, context: Context) -> None:\n    \"\"\"Main entry point for the ServerApp.\"\"\"\n    # Load config\n    num_rounds = context.run_config[\"num-server-rounds\"]\n    fraction_train = context.run_config[\"fraction-train\"]\n\n    # Load initial model\n    model = load_model()\n    arrays = ArrayRecord(model.get_weights())\n\n    # Define and start FedAvg strategy\n    strategy = FedAvg(\n        fraction_train=fraction_train,\n    )\n\n    result = strategy.start(\n        grid=grid,\n        initial_arrays=arrays,\n        num_rounds=num_rounds,\n    )\n\n    # Save the final model\n    ndarrays = result.arrays.to_numpy_ndarrays()\n    final_model_name = \"final_model.keras\"\n    print(f\"Saving final model to disk as {final_model_name}...\")\n    model.set_weights(ndarrays)\n    model.save(final_model_name)\n```\nNote the `start` method of the strategy returns a result object. This object contains all the relevant information about the FL process, including the final model weights as an `ArrayRecord`, and federated training and evaluation metrics as `MetricRecords`. You can easily log the metrics using Python's [pprint](https:\/\/docs.python.org\/3\/library\/pprint.html) and save the final model weights using Tensorflow's `save()` function.\n\nCongratulations! You've successfully built and run your first federated learning system.\n\nTip\n\nCheck the [运行模拟](https:\/\/flower.ai\/docs\/framework\/main\/zh_Hans\/how-to-run-simulations.html) documentation to learn more about how to configure and run Flower simulations.","meta_canonical":null}

3. Robots.txt Check

Query:

Response:

4. Spam/Ban Check

Query:

Response:

5. Seen Status Check

ℹ️ Skipped - page is already crawled

📄

INDEXABLE

✅

CRAWLED

13 days ago

🤖

ROBOTS ALLOWED

Page Info Filters

Filter	Status	Condition	Details
HTTP status	PASS	`download_http_code = 200`	HTTP 200
Age cutoff	PASS	`download_stamp > now() - 6 MONTH`	0.4 months ago
History drop	PASS	`isNull(history_drop_reason)`	No drop reason
Spam/ban	PASS	`fh_dont_index != 1 AND ml_spam_score = 0`	ml_spam_score=0
Canonical	PASS	`meta_canonical IS NULL OR = '' OR = src_unparsed`	Not set

Page Details

Property	Value
URL	https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html
Last Crawled	2026-03-24 18:08:57 (13 days ago)
First Indexed	2024-02-18 15:18:07 (2 years ago)
HTTP Status Code	200
Meta Title	快速入门 TensorFlow - Flower Framework
Meta Description	Learn how to train a Convolutional Neural Network on CIFAR-10 using federated learning with Flower and TensorFlow in this step-by-step tutorial.
Meta Canonical	null
Boilerpipe Text	In this tutorial we will learn how to train a Convolutional Neural Network on CIFAR-10 using the Flower framework and TensorFlow. First of all, it is recommended to create a virtual environment and run everything within a virtualenv . Let's use flwr new to create a complete Flower+TensorFlow project. It will generate all the files needed to run a federation of 10 nodes using FedAvg . By default, the generated app uses a local simulation profile that flwr run submits to a managed local SuperLink, which then executes the run with the Flower Simulation Runtime. The dataset will be partitioned using Flower Dataset's IidPartitioner . Now that we have a rough idea of what this example is about, let's get started. First, install Flower in your new environment: # In a new Python environment $ pip install flwr [ simulation ] Then, run the command below: $ flwr new @flwrlabs/quickstart-tensorflow After running it you'll notice a new directory named quickstart-tensorflow has been created. It should have the following structure: quickstart-tensorflow ├── tfexample │ ├── __init__.py │ ├── client_app.py # Defines your ClientApp │ ├── server_app.py # Defines your ServerApp │ └── task.py # Defines your model, training and data loading ├── pyproject.toml # Project metadata like dependencies and configs └── README.md If you haven't yet installed the project and its dependencies, you can do so by: # From the directory where your pyproject.toml is $ pip install -e . To run the project, do: # Run with default arguments and stream logs $ flwr run . --stream Plain flwr run . submits the run, prints the run ID, and returns without streaming logs. For the full local workflow, see Run Flower Locally with a Managed SuperLink . With default arguments you will see streamed output like this: Successfully built flwrlabs.quickstart-tensorflow.1-0-0.014c8eb3.fab Starting local SuperLink on 127 .0.0.1:39093... Successfully started run 1859953118041441032 INFO : Starting FedAvg strategy: INFO : ├── Number of rounds: 3 INFO : [ ROUND 1 /3 ] INFO : configure_train: Sampled 5 nodes ( out of 10 ) INFO : aggregate_train: Received 5 results and 0 failures INFO : └──> Aggregated MetricRecord: { 'train_loss' : 2 .0013, 'train_acc' : 0 .2624 } INFO : configure_evaluate: Sampled 10 nodes ( out of 10 ) INFO : aggregate_evaluate: Received 10 results and 0 failures INFO : └──> Aggregated MetricRecord: { 'eval_acc' : 0 .1216, 'eval_loss' : 2 .2686 } INFO : [ ROUND 2 /3 ] INFO : ... INFO : [ ROUND 3 /3 ] INFO : ... INFO : Strategy execution finished in 16 .60s INFO : Final results: INFO : ServerApp-side Evaluate Metrics: INFO : {} Saving final model to disk as final_model.keras... You can also override the parameters defined in the [tool.flwr.app.config] section in pyproject.toml like this: # Override some arguments $ flwr run . --run-config "num-server-rounds=5 batch-size=16" The Data ¶ This tutorial uses Flower Datasets to easily download and partition the CIFAR-10 dataset. In this example you'll make use of the IidPartitioner to generate num_partitions partitions. You can choose other partitioners available in Flower Datasets. Each ClientApp will call this function to create the NumPy arrays that correspond to their data partition. partitioner = IidPartitioner ( num_partitions = num_partitions ) fds = FederatedDataset ( dataset = "uoft-cs/cifar10" , partitioners = { "train" : partitioner }, ) partition = fds . load_partition ( partition_id , "train" ) partition . set_format ( "numpy" ) # Divide data on each node: 80% train, 20% test partition = partition . train_test_split ( test_size = 0.2 ) x_train , y_train = partition [ "train" ][ "img" ] / 255.0 , partition [ "train" ][ "label" ] x_test , y_test = partition [ "test" ][ "img" ] / 255.0 , partition [ "test" ][ "label" ] The Model ¶ Next, we need a model. We defined a simple Convolutional Neural Network (CNN), but feel free to replace it with a more sophisticated model if you'd like: def load_model ( learning_rate : float = 0.001 ): # Define a simple CNN for CIFAR-10 and set Adam optimizer model = keras . Sequential ( [ keras . Input ( shape = ( 32 , 32 , 3 )), layers . Conv2D ( 32 , kernel_size = ( 3 , 3 ), activation = "relu" ), layers . MaxPooling2D ( pool_size = ( 2 , 2 )), layers . Conv2D ( 64 , kernel_size = ( 3 , 3 ), activation = "relu" ), layers . MaxPooling2D ( pool_size = ( 2 , 2 )), layers . Flatten (), layers . Dropout ( 0.5 ), layers . Dense ( 10 , activation = "softmax" ), ] ) optimizer = keras . optimizers . Adam ( learning_rate ) model . compile ( optimizer = optimizer , loss = "sparse_categorical_crossentropy" , metrics = [ "accuracy" ], ) return model The ClientApp ¶ The main changes we have to make to use Tensorflow with Flower have to do with converting the ArrayRecord received in the Message into numpy ndarrays for use with the built-in set_weights() function. After training, the get_weights() function can be used to extract then pack the updated numpy ndarrays into a Message from the ClientApp. We can make use of built-in methods in the ArrayRecord to make these conversions: @app . train () def train ( msg : Message , context : Context ): # Load the model model = load_model ( context . run_config [ "learning-rate" ]) # Extract the ArrayRecord from Message and convert to numpy ndarrays model . set_weights ( msg . content [ "arrays" ] . to_numpy_ndarrays ()) # Train the model ... # Pack the model weights into an ArrayRecord model_record = ArrayRecord ( model . get_weights ()) The rest of the functionality is directly inspired by the centralized case. The ClientApp comes with three core methods ( train , evaluate , and query ) that we can implement for different purposes. For example: train to train the received model using the local data; evaluate to assess its performance of the received model on a validation set; and query to retrieve information about the node executing the ClientApp . In this tutorial we will only make use of train and evaluate . Let's see how the train method can be implemented. It receives as input arguments a Message from the ServerApp . By default it carries: an ArrayRecord with the arrays of the model to federate. By default they can be retrieved with key "arrays" when accessing the message content. a ConfigRecord with the configuration sent from the ServerApp . By default it can be retrieved with key "config" when accessing the message content. The train method also receives the Context , giving access to configs for your run and node. The run config hyperparameters are defined in the pyproject.toml of your Flower App. The node config can only be set when running Flower with the Deployment Runtime and is not directly configurable during simulations. # Flower ClientApp app = ClientApp () @app . train () def train ( msg : Message , context : Context ): """Train the model on local data.""" # Reset local Tensorflow state keras . backend . clear_session () # Load the data partition_id = context . node_config [ "partition-id" ] num_partitions = context . node_config [ "num-partitions" ] x_train , y_train , _ , _ = load_data ( partition_id , num_partitions ) # Load the model model = load_model ( context . run_config [ "learning-rate" ]) model . set_weights ( msg . content [ "arrays" ] . to_numpy_ndarrays ()) epochs = context . run_config [ "local-epochs" ] batch_size = context . run_config [ "batch-size" ] verbose = context . run_config . get ( "verbose" ) # Train the model history = model . fit ( x_train , y_train , epochs = epochs , batch_size = batch_size , verbose = verbose , ) # Get training metrics train_loss = history . history [ "loss" ][ - 1 ] if "loss" in history . history else None train_acc = ( history . history [ "accuracy" ][ - 1 ] if "accuracy" in history . history else None ) # Pack and send the model weights and metrics as a message model_record = ArrayRecord ( model . get_weights ()) metrics = { "num-examples" : len ( x_train )} if train_loss is not None : metrics [ "train_loss" ] = train_loss if train_acc is not None : metrics [ "train_acc" ] = train_acc content = RecordDict ({ "arrays" : model_record , "metrics" : MetricRecord ( metrics )}) return Message ( content = content , reply_to = msg ) The @app.evaluate() method would be near identical with two exceptions: (1) the model is not locally trained, instead it is used to evaluate its performance on the locally held-out validation set; (2) including the model in the reply Message is no longer needed because it is not locally modified. The ServerApp ¶ To construct a ServerApp we define its @app.main() method. This method receive as input arguments: a Grid object that will be used to interface with the nodes running the ClientApp to involve them in a round of train/evaluate/query or other. a Context object that provides access to the run configuration. In this example we use the FedAvg and configure it with a specific value of fraction_train which is read from the run config. You can find the default value defined in the pyproject.toml . Then, the execution of the strategy is launched when invoking its start method. To it we pass: the Grid object. an ArrayRecord carrying a randomly initialized model that will serve as the global model to federated. the num_rounds parameter specifying how many rounds of FedAvg to perform. # Create the ServerApp app = ServerApp () @app . main () def main ( grid : Grid , context : Context ) -> None : """Main entry point for the ServerApp.""" # Load config num_rounds = context . run_config [ "num-server-rounds" ] fraction_train = context . run_config [ "fraction-train" ] # Load initial model model = load_model () arrays = ArrayRecord ( model . get_weights ()) # Define and start FedAvg strategy strategy = FedAvg ( fraction_train = fraction_train , ) result = strategy . start ( grid = grid , initial_arrays = arrays , num_rounds = num_rounds , ) # Save the final model ndarrays = result . arrays . to_numpy_ndarrays () final_model_name = "final_model.keras" print ( f "Saving final model to disk as { final_model_name } ..." ) model . set_weights ( ndarrays ) model . save ( final_model_name ) Note the start method of the strategy returns a result object. This object contains all the relevant information about the FL process, including the final model weights as an ArrayRecord , and federated training and evaluation metrics as MetricRecords . You can easily log the metrics using Python's pprint and save the final model weights using Tensorflow's save() function. Congratulations! You've successfully built and run your first federated learning system. Tip Check the 运行模拟 documentation to learn more about how to configure and run Flower simulations.
Markdown	Hide navigation sidebar Hide table of contents sidebar [Skip to content](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#furo-main-content) Toggle site navigation sidebar [Flower Framework](https://flower.ai/docs/framework/main/zh_Hans/index.html) Toggle Light / Dark / Auto color theme Toggle table of contents sidebar [![Light Logo](https://flower.ai/docs/framework/main/zh_Hans/_static/flower-logo-light.png) ![Dark Logo](https://flower.ai/docs/framework/main/zh_Hans/_static/flower-logo-dark.png) main](https://flower.ai/docs/framework/main/zh_Hans/index.html) 教程 - [什么是联邦学习？](https://flower.ai/docs/framework/main/zh_Hans/tutorial-series-what-is-federated-learning.html) - [开始使用Flower](https://flower.ai/docs/framework/main/zh_Hans/tutorial-series-get-started-with-flower-pytorch.html) - [使用联邦学习策略](https://flower.ai/docs/framework/main/zh_Hans/tutorial-series-use-a-federated-learning-strategy-pytorch.html) - [Customize a Flower Strategy](https://flower.ai/docs/framework/main/zh_Hans/tutorial-series-build-a-strategy-from-scratch-pytorch.html) - [Communicate custom Messages](https://flower.ai/docs/framework/main/zh_Hans/tutorial-series-customize-the-client-pytorch.html) - [快速入门教程](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart.html) Toggle navigation of 快速入门教程 - [PyTorch快速入门](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-pytorch.html) - [快速入门 TensorFlow](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html) - [Quickstart MLX](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-mlx.html) - [🤗 Transformers快速入门](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-huggingface.html) - [快速入门 JAX](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-jax.html) - [快速入门Pandas](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-pandas.html) - [快速入门 fastai](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-fastai.html) - [快速入门 PyTorch Lightning](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-pytorch-lightning.html) - [scikit-learn快速入门](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-scikitlearn.html) - [XGBoost快速入门](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-xgboost.html) - [快速入门 Android](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-android.html) - [快速入门 iOS](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-ios.html) How-to Guides - [Build](https://flower.ai/docs/framework/main/zh_Hans/build.html) Toggle navigation of Build - [安装Flower](https://flower.ai/docs/framework/main/zh_Hans/how-to-install-flower.html) - [Configure `pyproject.toml`](https://flower.ai/docs/framework/main/zh_Hans/how-to-configure-pyproject-toml.html) - [Configure a `ClientApp`](https://flower.ai/docs/framework/main/zh_Hans/how-to-configure-clients.html) - [Design stateful ClientApps](https://flower.ai/docs/framework/main/zh_Hans/how-to-design-stateful-clients.html) - [使用策略](https://flower.ai/docs/framework/main/zh_Hans/how-to-use-strategies.html) - [实施策略](https://flower.ai/docs/framework/main/zh_Hans/how-to-implement-strategies.html) - [整合评估结果](https://flower.ai/docs/framework/main/zh_Hans/how-to-aggregate-evaluation-results.html) - [Save and load model checkpoints](https://flower.ai/docs/framework/main/zh_Hans/how-to-save-and-load-model-checkpoints.html) - [Use Built-in Mods](https://flower.ai/docs/framework/main/zh_Hans/how-to-use-built-in-mods.html) - [Use Differential Privacy](https://flower.ai/docs/framework/main/zh_Hans/how-to-use-differential-privacy.html) - [Implement FedBN](https://flower.ai/docs/framework/main/zh_Hans/how-to-implement-fedbn.html) - [Use CLI JSON output](https://flower.ai/docs/framework/main/zh_Hans/how-to-use-cli-json-output.html) - [OpenFL Migration Guide](https://flower.ai/docs/framework/main/zh_Hans/how-to-migrate-from-openfl.html) - [升级至 Flower 1.0](https://flower.ai/docs/framework/main/zh_Hans/how-to-upgrade-to-flower-1.0.html) - [Upgrade to Flower 1.13](https://flower.ai/docs/framework/main/zh_Hans/how-to-upgrade-to-flower-1.13.html) - [Upgrade to Message API](https://flower.ai/docs/framework/main/zh_Hans/how-to-upgrade-to-message-api.html) - [Simulate](https://flower.ai/docs/framework/main/zh_Hans/simulate.html) Toggle navigation of Simulate - [Run Flower Locally with a Managed SuperLink](https://flower.ai/docs/framework/main/zh_Hans/how-to-run-flower-locally.html) - [运行模拟](https://flower.ai/docs/framework/main/zh_Hans/how-to-run-simulations.html) - [Deploy](https://flower.ai/docs/framework/main/zh_Hans/deploy.html) Toggle navigation of Deploy - [Run Flower with the Deployment Runtime](https://flower.ai/docs/framework/main/zh_Hans/how-to-run-flower-with-deployment-engine.html) - [Enable TLS connections](https://flower.ai/docs/framework/main/zh_Hans/how-to-enable-tls-connections.html) - [Authenticate SuperNodes](https://flower.ai/docs/framework/main/zh_Hans/how-to-authenticate-supernodes.html) - [Configure logging](https://flower.ai/docs/framework/main/zh_Hans/how-to-configure-logging.html) - [Run Flower on GCP](https://flower.ai/docs/framework/main/zh_Hans/how-to-run-flower-on-gcp.html) - [Run Flower on Azure](https://flower.ai/docs/framework/main/zh_Hans/how-to-run-flower-on-azure.html) - [Run Flower on Red Hat OpenShift](https://flower.ai/docs/framework/main/zh_Hans/how-to-run-flower-on-red-hat-openshift.html) - [Run Flower on Multiple OpenShift Clusters](https://flower.ai/docs/framework/main/zh_Hans/how-to-run-flower-on-multiple-openshift-clusters.html) - [Authenticate Accounts via OpenID Connect](https://flower.ai/docs/framework/main/zh_Hans/how-to-authenticate-accounts.html) - [Configure Audit Logging](https://flower.ai/docs/framework/main/zh_Hans/how-to-configure-audit-logging.html) - [Manage Flower Federations](https://flower.ai/docs/framework/main/zh_Hans/how-to-manage-flower-federations.html) - [Run Flower using Docker](https://flower.ai/docs/framework/main/zh_Hans/docker/index.html) Toggle navigation of Run Flower using Docker - [Quickstart with Docker](https://flower.ai/docs/framework/main/zh_Hans/docker/tutorial-quickstart-docker.html) - [Enable TLS for Secure Connections](https://flower.ai/docs/framework/main/zh_Hans/docker/enable-tls.html) - [Persist the State of the SuperLink](https://flower.ai/docs/framework/main/zh_Hans/docker/persist-superlink-state.html) - [Set Environment Variables](https://flower.ai/docs/framework/main/zh_Hans/docker/set-environment-variables.html) - [Run with Root User Privileges](https://flower.ai/docs/framework/main/zh_Hans/docker/run-as-root-user.html) - [Run ServerApp or ClientApp as a Subprocess](https://flower.ai/docs/framework/main/zh_Hans/docker/run-as-subprocess.html) - [Pin a Docker Image to a Specific Version](https://flower.ai/docs/framework/main/zh_Hans/docker/pin-version.html) - [Use a Different Flower Version](https://flower.ai/docs/framework/main/zh_Hans/docker/use-a-different-version.html) - [Quickstart with Docker Compose](https://flower.ai/docs/framework/main/zh_Hans/docker/tutorial-quickstart-docker-compose.html) - [Run Flower Quickstart Examples with Docker Compose](https://flower.ai/docs/framework/main/zh_Hans/docker/run-quickstart-examples-docker-compose.html) - [Deploy Flower on Multiple Machines with Docker Compose](https://flower.ai/docs/framework/main/zh_Hans/docker/tutorial-deploy-on-multiple-machines.html) - [Run Flower using Helm](https://flower.ai/docs/framework/main/zh_Hans/helm/index.html) Toggle navigation of Run Flower using Helm - [Deploy SuperLink](https://flower.ai/docs/framework/main/zh_Hans/helm/how-to-deploy-superlink-using-helm.html) - [Deploy SuperNode](https://flower.ai/docs/framework/main/zh_Hans/helm/how-to-deploy-supernode-using-helm.html) 说明 - [联邦学习评估](https://flower.ai/docs/framework/main/zh_Hans/explanation-federated-evaluation.html) - [差分隐私](https://flower.ai/docs/framework/main/zh_Hans/explanation-differential-privacy.html) - [安全聚合协议](https://flower.ai/docs/framework/main/zh_Hans/explanation-ref-secure-aggregation-protocols.html) - [Flower的架构](https://flower.ai/docs/framework/main/zh_Hans/explanation-flower-architecture.html) - [Flower Strategy Abstraction](https://flower.ai/docs/framework/main/zh_Hans/explanation-flower-strategy-abstraction.html) 参考资料 - [Reference](https://flower.ai/docs/framework/main/zh_Hans/reference.html) Toggle navigation of Reference - [flwr](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.html) Toggle navigation of flwr - [app](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.html) Toggle navigation of app - [Array](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.Array.html) - [ArrayRecord](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.ArrayRecord.html) - [ConfigRecord](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.ConfigRecord.html) - [Context](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.Context.html) - [Error](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.Error.html) - [Message](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.Message.html) - [MessageType](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.MessageType.html) - [描述数据](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.Metadata.html) - [MetricRecord](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.MetricRecord.html) - [RecordDict](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.RecordDict.html) - [UserConfig](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.UserConfig.html) - [clientapp](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.clientapp.html) Toggle navigation of clientapp - [ClientApp](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.clientapp.ClientApp.html) - [mod](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.clientapp.mod.html) Toggle navigation of mod - [adaptiveclipping\_mod](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.clientapp.mod.adaptiveclipping_mod.html) - [arrays\_size\_mod](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.clientapp.mod.arrays_size_mod.html) - [fixedclipping\_mod](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.clientapp.mod.fixedclipping_mod.html) - [message\_size\_mod](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.clientapp.mod.message_size_mod.html) - [LocalDpMod](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.clientapp.mod.LocalDpMod.html) - [serverapp](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.html) Toggle navigation of serverapp - [Grid](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.Grid.html) - [ServerApp](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.ServerApp.html) - [strategy](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.html) Toggle navigation of strategy - [Bulyan](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.Bulyan.html) - [DifferentialPrivacyClientSideAdaptiveClipping](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.DifferentialPrivacyClientSideAdaptiveClipping.html) - [DifferentialPrivacyClientSideFixedClipping](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.DifferentialPrivacyClientSideFixedClipping.html) - [DifferentialPrivacyServerSideAdaptiveClipping](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.DifferentialPrivacyServerSideAdaptiveClipping.html) - [DifferentialPrivacyServerSideFixedClipping](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.DifferentialPrivacyServerSideFixedClipping.html) - [FedAdagrad](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedAdagrad.html) - [FedAdam](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedAdam.html) - [FedAvg](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedAvg.html) - [FedAvgM](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedAvgM.html) - [FedMedian](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedMedian.html) - [FedProx](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedProx.html) - [FedTrimmedAvg](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedTrimmedAvg.html) - [FedXgbBagging](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedXgbBagging.html) - [FedXgbCyclic](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedXgbCyclic.html) - [FedYogi](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedYogi.html) - [Krum](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.Krum.html) - [MultiKrum](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.MultiKrum.html) - [QFedAvg](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.QFedAvg.html) - [Result](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.Result.html) - [Strategy](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.Strategy.html) - [Flower CLI 参考](https://flower.ai/docs/framework/main/zh_Hans/ref-api-cli.html) - [Flower Configuration](https://flower.ai/docs/framework/main/zh_Hans/ref-flower-configuration.html) - [项目实例](https://flower.ai/docs/framework/main/zh_Hans/ref-example-projects.html) - [遥测功能](https://flower.ai/docs/framework/main/zh_Hans/ref-telemetry.html) - [更新日志](https://flower.ai/docs/framework/main/zh_Hans/ref-changelog.html) - [Flower Runtime Comparison](https://flower.ai/docs/framework/main/zh_Hans/ref-flower-runtime-comparison.html) - [Flower Network Communication](https://flower.ai/docs/framework/main/zh_Hans/ref-flower-network-communication.html) - [Exit Codes](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes-dir.html) Toggle navigation of Exit Codes - [\[0\] SUCCESS](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/0.html) - [\[1\] GRACEFUL\_EXIT\_SIGINT](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/1.html) - [\[100\] SUPERLINK\_THREAD\_CRASH](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/100.html) - [\[101\] SUPERLINK\_LICENSE\_INVALID](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/101.html) - [\[102\] SUPERLINK\_LICENSE\_MISSING](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/102.html) - [\[103\] SUPERLINK\_LICENSE\_URL\_INVALID](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/103.html) - [\[104\] SUPERLINK\_INVALID\_ARGS](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/104.html) - [\[105\] SUPERLINK\_DATABASE\_SCHEMA\_MISMATCH](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/105.html) - [\[2\] GRACEFUL\_EXIT\_SIGQUIT](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/2.html) - [\[200\] SERVERAPP\_STRATEGY\_PRECONDITION\_UNMET](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/200.html) - [\[201\] SERVERAPP\_EXCEPTION](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/201.html) - [\[202\] SERVERAPP\_STRATEGY\_AGGREGATION\_ERROR](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/202.html) - [\[203\] SERVERAPP\_RUN\_START\_REJECTED](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/203.html) - [\[3\] GRACEFUL\_EXIT\_SIGTERM](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/3.html) - [\[300\] SUPERNODE\_REST\_ADDRESS\_INVALID](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/300.html) - [\[302\] SUPERNODE\_NODE\_AUTH\_KEY\_INVALID](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/302.html) - [\[303\] SUPERNODE\_STARTED\_WITHOUT\_TLS\_BUT\_NODE\_AUTH\_ENABLED](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/303.html) - [\[304\] SUPERNODE\_INVALID\_TRUSTED\_ENTITIES](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/304.html) - [\[400\] SUPEREXEC\_INVALID\_PLUGIN\_CONFIG](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/400.html) - [\[500\] FLWRCLI\_NODE\_AUTH\_PUBLIC\_KEY\_INVALID](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/500.html) - [\[600\] COMMON\_ADDRESS\_INVALID](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/600.html) - [\[601\] COMMON\_MISSING\_EXTRA\_REST](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/601.html) - [\[602\] COMMON\_TLS\_NOT\_SUPPORTED](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/602.html) - [\[700\] SIMULATION\_EXCEPTION](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/700.html) - [\[701\] SIMULATION\_MISSING\_EXTRA](https://flower.ai/docs/framework/main/zh_Hans/ref-exit-codes/701.html) - [常见问题](https://flower.ai/docs/framework/main/zh_Hans/ref-faq.html) 贡献者文档 - [Contribute](https://flower.ai/docs/framework/main/zh_Hans/contribute.html) Toggle navigation of Contribute - [在 GitHub 上投稿](https://flower.ai/docs/framework/main/zh_Hans/contributor-tutorial-contribute-on-github.html) - [成为贡献者](https://flower.ai/docs/framework/main/zh_Hans/contributor-tutorial-get-started-as-a-contributor.html) - [安装开发版本](https://flower.ai/docs/framework/main/zh_Hans/contributor-how-to-install-development-versions.html) - [建立虚拟环境](https://flower.ai/docs/framework/main/zh_Hans/contributor-how-to-set-up-a-virtual-env.html) - [使用 VSCode Dev Containers 进行开发](https://flower.ai/docs/framework/main/zh_Hans/contributor-how-to-develop-in-vscode-dev-containers.html) - [编写文件](https://flower.ai/docs/framework/main/zh_Hans/contributor-how-to-write-documentation.html) - [发布 Flower](https://flower.ai/docs/framework/main/zh_Hans/contributor-how-to-release-flower.html) - [贡献译文](https://flower.ai/docs/framework/main/zh_Hans/contributor-how-to-contribute-translations.html) - [How to Build Docker Flower Images Locally](https://flower.ai/docs/framework/main/zh_Hans/contributor-how-to-build-docker-images.html) - [Migrate Flower Database Schema](https://flower.ai/docs/framework/main/zh_Hans/contributor-how-to-migrate-database.html) - [Public and private APIs](https://flower.ai/docs/framework/main/zh_Hans/contributor-explanation-public-and-private-apis.html) - [首次代码贡献](https://flower.ai/docs/framework/main/zh_Hans/contributor-ref-good-first-contributions.html) Versions - [main](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html) [🇬🇧](https://flower.ai/docs/framework/main/en/tutorial-quickstart-tensorflow.html) [🇫🇷](https://flower.ai/docs/framework/main/fr/tutorial-quickstart-tensorflow.html) [🇨🇳](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html) [🇰🇷](https://flower.ai/docs/framework/main/ko/tutorial-quickstart-tensorflow.html) [Back to top](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html) [View this page](https://flower.ai/docs/framework/main/zh_Hans/_sources/tutorial-quickstart-tensorflow.rst.txt "View this page") Toggle Light / Dark / Auto color theme Toggle table of contents sidebar # 快速入门 TensorFlow[¶](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#id1 "Link to this heading") In this tutorial we will learn how to train a Convolutional Neural Network on CIFAR-10 using the Flower framework and TensorFlow. First of all, it is recommended to create a virtual environment and run everything within a [virtualenv](https://flower.ai/docs/framework/main/zh_Hans/contributor-how-to-set-up-a-virtual-env.html). Let's use `flwr new` to create a complete Flower+TensorFlow project. It will generate all the files needed to run a federation of 10 nodes using [`FedAvg`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedAvg.html). By default, the generated app uses a local simulation profile that `flwr run` submits to a managed local SuperLink, which then executes the run with the Flower Simulation Runtime. The dataset will be partitioned using Flower Dataset's [IidPartitioner](https://flower.ai/docs/datasets/ref-api/flwr_datasets.partitioner.IidPartitioner.html#flwr_datasets.partitioner.IidPartitioner). Now that we have a rough idea of what this example is about, let's get started. First, install Flower in your new environment: ``` # In a new Python environment $ pip install flwr[simulation] ``` Then, run the command below: ``` $ flwr new @flwrlabs/quickstart-tensorflow ``` After running it you'll notice a new directory named `quickstart-tensorflow` has been created. It should have the following structure: ``` quickstart-tensorflow ├── tfexample │ ├── __init__.py │ ├── client_app.py # Defines your ClientApp │ ├── server_app.py # Defines your ServerApp │ └── task.py # Defines your model, training and data loading ├── pyproject.toml # Project metadata like dependencies and configs └── README.md ``` If you haven't yet installed the project and its dependencies, you can do so by: ``` # From the directory where your pyproject.toml is $ pip install -e . ``` To run the project, do: ``` # Run with default arguments and stream logs $ flwr run . --stream ``` Plain `flwr run .` submits the run, prints the run ID, and returns without streaming logs. For the full local workflow, see [Run Flower Locally with a Managed SuperLink](https://flower.ai/docs/framework/main/zh_Hans/how-to-run-flower-locally.html). With default arguments you will see streamed output like this: ``` Successfully built flwrlabs.quickstart-tensorflow.1-0-0.014c8eb3.fab Starting local SuperLink on 127.0.0.1:39093... Successfully started run 1859953118041441032 INFO : Starting FedAvg strategy: INFO : ├── Number of rounds: 3 INFO : [ROUND 1/3] INFO : configure_train: Sampled 5 nodes (out of 10) INFO : aggregate_train: Received 5 results and 0 failures INFO : └──> Aggregated MetricRecord: {'train_loss': 2.0013, 'train_acc': 0.2624} INFO : configure_evaluate: Sampled 10 nodes (out of 10) INFO : aggregate_evaluate: Received 10 results and 0 failures INFO : └──> Aggregated MetricRecord: {'eval_acc': 0.1216, 'eval_loss': 2.2686} INFO : [ROUND 2/3] INFO : ... INFO : [ROUND 3/3] INFO : ... INFO : Strategy execution finished in 16.60s INFO : Final results: INFO : ServerApp-side Evaluate Metrics: INFO : {} Saving final model to disk as final_model.keras... ``` You can also override the parameters defined in the `[tool.flwr.app.config]` section in `pyproject.toml` like this: ``` # Override some arguments $ flwr run . --run-config "num-server-rounds=5 batch-size=16" ``` ## The Data[¶](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-data "Link to this heading") This tutorial uses [Flower Datasets](https://flower.ai/docs/datasets/) to easily download and partition the CIFAR-10 dataset. In this example you'll make use of the [IidPartitioner](https://flower.ai/docs/datasets/ref-api/flwr_datasets.partitioner.IidPartitioner.html#flwr_datasets.partitioner.IidPartitioner) to generate num\_partitions partitions. You can choose [other partitioners](https://flower.ai/docs/datasets/ref-api/flwr_datasets.partitioner.html) available in Flower Datasets. Each `ClientApp` will call this function to create the `NumPy` arrays that correspond to their data partition. ``` partitioner = IidPartitioner(num_partitions=num_partitions) fds = FederatedDataset( dataset="uoft-cs/cifar10", partitioners={"train": partitioner}, ) partition = fds.load_partition(partition_id, "train") partition.set_format("numpy") # Divide data on each node: 80% train, 20% test partition = partition.train_test_split(test_size=0.2) x_train, y_train = partition["train"]["img"] / 255.0, partition["train"]["label"] x_test, y_test = partition["test"]["img"] / 255.0, partition["test"]["label"] ``` ## The Model[¶](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-model "Link to this heading") Next, we need a model. We defined a simple Convolutional Neural Network (CNN), but feel free to replace it with a more sophisticated model if you'd like: ``` def load_model(learning_rate: float = 0.001): # Define a simple CNN for CIFAR-10 and set Adam optimizer model = keras.Sequential( [ keras.Input(shape=(32, 32, 3)), layers.Conv2D(32, kernel_size=(3, 3), activation="relu"), layers.MaxPooling2D(pool_size=(2, 2)), layers.Conv2D(64, kernel_size=(3, 3), activation="relu"), layers.MaxPooling2D(pool_size=(2, 2)), layers.Flatten(), layers.Dropout(0.5), layers.Dense(10, activation="softmax"), ] ) optimizer = keras.optimizers.Adam(learning_rate) model.compile( optimizer=optimizer, loss="sparse_categorical_crossentropy", metrics=["accuracy"], ) return model ``` ## The ClientApp[¶](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-clientapp "Link to this heading") The main changes we have to make to use Tensorflow with Flower have to do with converting the [`ArrayRecord`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.ArrayRecord.html) received in the [`Message`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.Message.html) into numpy ndarrays for use with the built-in `set_weights()` function. After training, the `get_weights()` function can be used to extract then pack the updated numpy ndarrays into a `Message` from the ClientApp. We can make use of built-in methods in the `ArrayRecord` to make these conversions: ``` @app.train() def train(msg: Message, context: Context): # Load the model model = load_model(context.run_config["learning-rate"]) # Extract the ArrayRecord from Message and convert to numpy ndarrays model.set_weights(msg.content["arrays"].to_numpy_ndarrays()) # Train the model ... # Pack the model weights into an ArrayRecord model_record = ArrayRecord(model.get_weights()) ``` The rest of the functionality is directly inspired by the centralized case. The [`ClientApp`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.clientapp.ClientApp.html) comes with three core methods (`train`, `evaluate`, and `query`) that we can implement for different purposes. For example: `train` to train the received model using the local data; `evaluate` to assess its performance of the received model on a validation set; and `query` to retrieve information about the node executing the `ClientApp`. In this tutorial we will only make use of `train` and `evaluate`. Let's see how the `train` method can be implemented. It receives as input arguments a [`Message`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.Message.html) from the `ServerApp`. By default it carries: - an `ArrayRecord` with the arrays of the model to federate. By default they can be retrieved with key `"arrays"` when accessing the message content. - a `ConfigRecord` with the configuration sent from the `ServerApp`. By default it can be retrieved with key `"config"` when accessing the message content. The `train` method also receives the `Context`, giving access to configs for your run and node. The run config hyperparameters are defined in the `pyproject.toml` of your Flower App. The node config can only be set when running Flower with the Deployment Runtime and is not directly configurable during simulations. ``` # Flower ClientApp app = ClientApp() @app.train() def train(msg: Message, context: Context): """Train the model on local data.""" # Reset local Tensorflow state keras.backend.clear_session() # Load the data partition_id = context.node_config["partition-id"] num_partitions = context.node_config["num-partitions"] x_train, y_train, _, _ = load_data(partition_id, num_partitions) # Load the model model = load_model(context.run_config["learning-rate"]) model.set_weights(msg.content["arrays"].to_numpy_ndarrays()) epochs = context.run_config["local-epochs"] batch_size = context.run_config["batch-size"] verbose = context.run_config.get("verbose") # Train the model history = model.fit( x_train, y_train, epochs=epochs, batch_size=batch_size, verbose=verbose, ) # Get training metrics train_loss = history.history["loss"][-1] if "loss" in history.history else None train_acc = ( history.history["accuracy"][-1] if "accuracy" in history.history else None ) # Pack and send the model weights and metrics as a message model_record = ArrayRecord(model.get_weights()) metrics = {"num-examples": len(x_train)} if train_loss is not None: metrics["train_loss"] = train_loss if train_acc is not None: metrics["train_acc"] = train_acc content = RecordDict({"arrays": model_record, "metrics": MetricRecord(metrics)}) return Message(content=content, reply_to=msg) ``` The `@app.evaluate()` method would be near identical with two exceptions: (1) the model is not locally trained, instead it is used to evaluate its performance on the locally held-out validation set; (2) including the model in the reply Message is no longer needed because it is not locally modified. ## The ServerApp[¶](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-serverapp "Link to this heading") To construct a [`ServerApp`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.ServerApp.html) we define its `@app.main()` method. This method receive as input arguments: - a `Grid` object that will be used to interface with the nodes running the `ClientApp` to involve them in a round of train/evaluate/query or other. - a `Context` object that provides access to the run configuration. In this example we use the [`FedAvg`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedAvg.html) and configure it with a specific value of `fraction_train` which is read from the run config. You can find the default value defined in the `pyproject.toml`. Then, the execution of the strategy is launched when invoking its [`start`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.Strategy.html#flwr.serverapp.Strategy.start) method. To it we pass: - the `Grid` object. - an `ArrayRecord` carrying a randomly initialized model that will serve as the global model to federated. - the `num_rounds` parameter specifying how many rounds of `FedAvg` to perform. ``` # Create the ServerApp app = ServerApp() @app.main() def main(grid: Grid, context: Context) -> None: """Main entry point for the ServerApp.""" # Load config num_rounds = context.run_config["num-server-rounds"] fraction_train = context.run_config["fraction-train"] # Load initial model model = load_model() arrays = ArrayRecord(model.get_weights()) # Define and start FedAvg strategy strategy = FedAvg( fraction_train=fraction_train, ) result = strategy.start( grid=grid, initial_arrays=arrays, num_rounds=num_rounds, ) # Save the final model ndarrays = result.arrays.to_numpy_ndarrays() final_model_name = "final_model.keras" print(f"Saving final model to disk as {final_model_name}...") model.set_weights(ndarrays) model.save(final_model_name) ``` Note the `start` method of the strategy returns a result object. This object contains all the relevant information about the FL process, including the final model weights as an `ArrayRecord`, and federated training and evaluation metrics as `MetricRecords`. You can easily log the metrics using Python's [pprint](https://docs.python.org/3/library/pprint.html) and save the final model weights using Tensorflow's `save()` function. Congratulations! You've successfully built and run your first federated learning system. Tip Check the [运行模拟](https://flower.ai/docs/framework/main/zh_Hans/how-to-run-simulations.html) documentation to learn more about how to configure and run Flower simulations. Note Check the source code of the extended version of this tutorial in [`examples/quickstart-tensorflow`](https://github.com/flwrlabs/flower/blob/main/examples/quickstart-tensorflow) in the Flower GitHub repository. [Next Quickstart MLX](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-mlx.html) [Previous PyTorch快速入门](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-pytorch.html) Copyright © 2026 Flower Labs GmbH Made with [Sphinx](https://www.sphinx-doc.org/) and [@pradyunsg](https://pradyunsg.me/)'s [Furo](https://github.com/pradyunsg/furo) On this page - [快速入门 TensorFlow](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html) - [The Data](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-data) - [The Model](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-model) - [The ClientApp](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-clientapp) - [The ServerApp](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-serverapp)
Readable Markdown	In this tutorial we will learn how to train a Convolutional Neural Network on CIFAR-10 using the Flower framework and TensorFlow. First of all, it is recommended to create a virtual environment and run everything within a [virtualenv](https://flower.ai/docs/framework/main/zh_Hans/contributor-how-to-set-up-a-virtual-env.html). Let's use `flwr new` to create a complete Flower+TensorFlow project. It will generate all the files needed to run a federation of 10 nodes using [`FedAvg`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedAvg.html). By default, the generated app uses a local simulation profile that `flwr run` submits to a managed local SuperLink, which then executes the run with the Flower Simulation Runtime. The dataset will be partitioned using Flower Dataset's [IidPartitioner](https://flower.ai/docs/datasets/ref-api/flwr_datasets.partitioner.IidPartitioner.html#flwr_datasets.partitioner.IidPartitioner). Now that we have a rough idea of what this example is about, let's get started. First, install Flower in your new environment: ``` # In a new Python environment $ pip install flwr[simulation] ``` Then, run the command below: ``` $ flwr new @flwrlabs/quickstart-tensorflow ``` After running it you'll notice a new directory named `quickstart-tensorflow` has been created. It should have the following structure: ``` quickstart-tensorflow ├── tfexample │ ├── __init__.py │ ├── client_app.py # Defines your ClientApp │ ├── server_app.py # Defines your ServerApp │ └── task.py # Defines your model, training and data loading ├── pyproject.toml # Project metadata like dependencies and configs └── README.md ``` If you haven't yet installed the project and its dependencies, you can do so by: ``` # From the directory where your pyproject.toml is $ pip install -e . ``` To run the project, do: ``` # Run with default arguments and stream logs $ flwr run . --stream ``` Plain `flwr run .` submits the run, prints the run ID, and returns without streaming logs. For the full local workflow, see [Run Flower Locally with a Managed SuperLink](https://flower.ai/docs/framework/main/zh_Hans/how-to-run-flower-locally.html). With default arguments you will see streamed output like this: ``` Successfully built flwrlabs.quickstart-tensorflow.1-0-0.014c8eb3.fab Starting local SuperLink on 127.0.0.1:39093... Successfully started run 1859953118041441032 INFO : Starting FedAvg strategy: INFO : ├── Number of rounds: 3 INFO : [ROUND 1/3] INFO : configure_train: Sampled 5 nodes (out of 10) INFO : aggregate_train: Received 5 results and 0 failures INFO : └──> Aggregated MetricRecord: {'train_loss': 2.0013, 'train_acc': 0.2624} INFO : configure_evaluate: Sampled 10 nodes (out of 10) INFO : aggregate_evaluate: Received 10 results and 0 failures INFO : └──> Aggregated MetricRecord: {'eval_acc': 0.1216, 'eval_loss': 2.2686} INFO : [ROUND 2/3] INFO : ... INFO : [ROUND 3/3] INFO : ... INFO : Strategy execution finished in 16.60s INFO : Final results: INFO : ServerApp-side Evaluate Metrics: INFO : {} Saving final model to disk as final_model.keras... ``` You can also override the parameters defined in the `[tool.flwr.app.config]` section in `pyproject.toml` like this: ``` # Override some arguments $ flwr run . --run-config "num-server-rounds=5 batch-size=16" ``` ## The Data[¶](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-data "Link to this heading") This tutorial uses [Flower Datasets](https://flower.ai/docs/datasets/) to easily download and partition the CIFAR-10 dataset. In this example you'll make use of the [IidPartitioner](https://flower.ai/docs/datasets/ref-api/flwr_datasets.partitioner.IidPartitioner.html#flwr_datasets.partitioner.IidPartitioner) to generate num\_partitions partitions. You can choose [other partitioners](https://flower.ai/docs/datasets/ref-api/flwr_datasets.partitioner.html) available in Flower Datasets. Each `ClientApp` will call this function to create the `NumPy` arrays that correspond to their data partition. ``` partitioner = IidPartitioner(num_partitions=num_partitions) fds = FederatedDataset( dataset="uoft-cs/cifar10", partitioners={"train": partitioner}, ) partition = fds.load_partition(partition_id, "train") partition.set_format("numpy") # Divide data on each node: 80% train, 20% test partition = partition.train_test_split(test_size=0.2) x_train, y_train = partition["train"]["img"] / 255.0, partition["train"]["label"] x_test, y_test = partition["test"]["img"] / 255.0, partition["test"]["label"] ``` ## The Model[¶](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-model "Link to this heading") Next, we need a model. We defined a simple Convolutional Neural Network (CNN), but feel free to replace it with a more sophisticated model if you'd like: ``` def load_model(learning_rate: float = 0.001): # Define a simple CNN for CIFAR-10 and set Adam optimizer model = keras.Sequential( [ keras.Input(shape=(32, 32, 3)), layers.Conv2D(32, kernel_size=(3, 3), activation="relu"), layers.MaxPooling2D(pool_size=(2, 2)), layers.Conv2D(64, kernel_size=(3, 3), activation="relu"), layers.MaxPooling2D(pool_size=(2, 2)), layers.Flatten(), layers.Dropout(0.5), layers.Dense(10, activation="softmax"), ] ) optimizer = keras.optimizers.Adam(learning_rate) model.compile( optimizer=optimizer, loss="sparse_categorical_crossentropy", metrics=["accuracy"], ) return model ``` ## The ClientApp[¶](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-clientapp "Link to this heading") The main changes we have to make to use Tensorflow with Flower have to do with converting the [`ArrayRecord`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.ArrayRecord.html) received in the [`Message`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.Message.html) into numpy ndarrays for use with the built-in `set_weights()` function. After training, the `get_weights()` function can be used to extract then pack the updated numpy ndarrays into a `Message` from the ClientApp. We can make use of built-in methods in the `ArrayRecord` to make these conversions: ``` @app.train() def train(msg: Message, context: Context): # Load the model model = load_model(context.run_config["learning-rate"]) # Extract the ArrayRecord from Message and convert to numpy ndarrays model.set_weights(msg.content["arrays"].to_numpy_ndarrays()) # Train the model ... # Pack the model weights into an ArrayRecord model_record = ArrayRecord(model.get_weights()) ``` The rest of the functionality is directly inspired by the centralized case. The [`ClientApp`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.clientapp.ClientApp.html) comes with three core methods (`train`, `evaluate`, and `query`) that we can implement for different purposes. For example: `train` to train the received model using the local data; `evaluate` to assess its performance of the received model on a validation set; and `query` to retrieve information about the node executing the `ClientApp`. In this tutorial we will only make use of `train` and `evaluate`. Let's see how the `train` method can be implemented. It receives as input arguments a [`Message`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.app.Message.html) from the `ServerApp`. By default it carries: - an `ArrayRecord` with the arrays of the model to federate. By default they can be retrieved with key `"arrays"` when accessing the message content. - a `ConfigRecord` with the configuration sent from the `ServerApp`. By default it can be retrieved with key `"config"` when accessing the message content. The `train` method also receives the `Context`, giving access to configs for your run and node. The run config hyperparameters are defined in the `pyproject.toml` of your Flower App. The node config can only be set when running Flower with the Deployment Runtime and is not directly configurable during simulations. ``` # Flower ClientApp app = ClientApp() @app.train() def train(msg: Message, context: Context): """Train the model on local data.""" # Reset local Tensorflow state keras.backend.clear_session() # Load the data partition_id = context.node_config["partition-id"] num_partitions = context.node_config["num-partitions"] x_train, y_train, _, _ = load_data(partition_id, num_partitions) # Load the model model = load_model(context.run_config["learning-rate"]) model.set_weights(msg.content["arrays"].to_numpy_ndarrays()) epochs = context.run_config["local-epochs"] batch_size = context.run_config["batch-size"] verbose = context.run_config.get("verbose") # Train the model history = model.fit( x_train, y_train, epochs=epochs, batch_size=batch_size, verbose=verbose, ) # Get training metrics train_loss = history.history["loss"][-1] if "loss" in history.history else None train_acc = ( history.history["accuracy"][-1] if "accuracy" in history.history else None ) # Pack and send the model weights and metrics as a message model_record = ArrayRecord(model.get_weights()) metrics = {"num-examples": len(x_train)} if train_loss is not None: metrics["train_loss"] = train_loss if train_acc is not None: metrics["train_acc"] = train_acc content = RecordDict({"arrays": model_record, "metrics": MetricRecord(metrics)}) return Message(content=content, reply_to=msg) ``` The `@app.evaluate()` method would be near identical with two exceptions: (1) the model is not locally trained, instead it is used to evaluate its performance on the locally held-out validation set; (2) including the model in the reply Message is no longer needed because it is not locally modified. ## The ServerApp[¶](https://flower.ai/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html#the-serverapp "Link to this heading") To construct a [`ServerApp`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.ServerApp.html) we define its `@app.main()` method. This method receive as input arguments: - a `Grid` object that will be used to interface with the nodes running the `ClientApp` to involve them in a round of train/evaluate/query or other. - a `Context` object that provides access to the run configuration. In this example we use the [`FedAvg`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.strategy.FedAvg.html) and configure it with a specific value of `fraction_train` which is read from the run config. You can find the default value defined in the `pyproject.toml`. Then, the execution of the strategy is launched when invoking its [`start`](https://flower.ai/docs/framework/main/zh_Hans/ref-api/flwr.serverapp.Strategy.html#flwr.serverapp.Strategy.start) method. To it we pass: - the `Grid` object. - an `ArrayRecord` carrying a randomly initialized model that will serve as the global model to federated. - the `num_rounds` parameter specifying how many rounds of `FedAvg` to perform. ``` # Create the ServerApp app = ServerApp() @app.main() def main(grid: Grid, context: Context) -> None: """Main entry point for the ServerApp.""" # Load config num_rounds = context.run_config["num-server-rounds"] fraction_train = context.run_config["fraction-train"] # Load initial model model = load_model() arrays = ArrayRecord(model.get_weights()) # Define and start FedAvg strategy strategy = FedAvg( fraction_train=fraction_train, ) result = strategy.start( grid=grid, initial_arrays=arrays, num_rounds=num_rounds, ) # Save the final model ndarrays = result.arrays.to_numpy_ndarrays() final_model_name = "final_model.keras" print(f"Saving final model to disk as {final_model_name}...") model.set_weights(ndarrays) model.save(final_model_name) ``` Note the `start` method of the strategy returns a result object. This object contains all the relevant information about the FL process, including the final model weights as an `ArrayRecord`, and federated training and evaluation metrics as `MetricRecords`. You can easily log the metrics using Python's [pprint](https://docs.python.org/3/library/pprint.html) and save the final model weights using Tensorflow's `save()` function. Congratulations! You've successfully built and run your first federated learning system. Tip Check the [运行模拟](https://flower.ai/docs/framework/main/zh_Hans/how-to-run-simulations.html) documentation to learn more about how to configure and run Flower simulations.
Shard	163 (laksa)
Root Hash	4426631000825432763
Unparsed URL	ai,flower!/docs/framework/main/zh_Hans/tutorial-quickstart-tensorflow.html s443