训练catvsdog 的tflite库

由于我的机器非常老旧所以采用的是依靠cpu训练的模式搭建的。‌硬件特性‌该机型通常搭载Intel酷睿标压处理器、核显无独立高性能GPU训练时优先使用CPU模式避免显存不足报错。‌软件环境‌安装Python 3.7~3.8版本搭配TensorFlow 2.x稳定版避免高版本框架带来的兼容性问题。‌依赖库安装‌执行命令安装所需工具包pip install tensorflow opencv-python numpy matplotlib最后系统选取的是版本如下pip config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple/从国内镜像站中下载的没有设置出现了网速不好的情况断开的异常。pip install tensorflow opencv-python numpy matplotlibLooking in indexes: https://pypi.tuna.tsinghua.edu.cn/simple/Collecting tensorflowDownloading https://pypi.tuna.tsinghua.edu.cn/packages/8f/a2/6d7e6a738e302530586d484895de2cf3fc158ad9c73b4504a670b2956dd9/tensorflow-2.21.0-cp311-cp311-win_amd64.whl (350.8 MB)---------------------------------------- 350.8/350.8 MB 292.4 kB/s eta 0:00:00Collecting opencv-pythonDownloading https://pypi.tuna.tsinghua.edu.cn/packages/21/f0/9fa6e85cb10c8eb36a0222d27e50fe381b86ce49a55446bf39f491727564/opencv_python-5.0.0.93-cp37-abi3-win_amd64.whl (44.0 MB)---------------------------------------- 44.0/44.0 MB 316.4 kB/s eta 0:00:00Requirement already satisfied: numpy in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (2.4.6)Collecting matplotlibDownloading https://pypi.tuna.tsinghua.edu.cn/packages/bc/be/fa26ed085b41298f64a8f9b7592c671bbf1acc8b0df124c1c5de96b859f8/matplotlib-3.11.1-cp311-cp311-win_amd64.whl (9.3 MB)---------------------------------------- 9.3/9.3 MB 322.6 kB/s eta 0:00:00Collecting absl-py1.0.0Downloading https://pypi.tuna.tsinghua.edu.cn/packages/58/0a/a10b45aab35b175aded078a462dc8d0c698f5b13946e7cb0869097b78bb6/absl_py-2.5.0-py3-none-any.whl (137 kB)---------------------------------------- 137.4/137.4 kB 61.7 kB/s eta 0:00:00Collecting astunparse1.6.0Downloading https://pypi.tuna.tsinghua.edu.cn/packages/2b/03/13dde6512ad7b4557eb792fbcf0c653af6076b81e5941d36ec61f7ce6028/astunparse-1.6.3-py2.py3-none-any.whl (12 kB)Requirement already satisfied: flatbuffers25.9.23 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from tensorflow) (25.12.19)Collecting gast!0.5.0,!0.5.1,!0.5.2,0.2.1Downloading https://pypi.tuna.tsinghua.edu.cn/packages/1d/33/f1c6a276de27b7d7339a34749cc33fa87f077f921969c47185d34a887ae2/gast-0.7.0-py3-none-any.whl (22 kB)Collecting google_pasta0.1.1Downloading https://pypi.tuna.tsinghua.edu.cn/packages/a3/de/c648ef6835192e6e2cc03f40b19eeda4382c49b5bafb43d88b931c4c74ac/google_pasta-0.2.0-py3-none-any.whl (57 kB)---------------------------------------- 57.5/57.5 kB 30.2 kB/s eta 0:00:00Collecting libclang13.0.0Downloading https://pypi.tuna.tsinghua.edu.cn/packages/0b/2d/3f480b1e1d31eb3d6de5e3ef641954e5c67430d5ac93b7fa7e07589576c7/libclang-18.1.1-py2.py3-none-win_amd64.whl (26.4 MB)---------------------------------------- 26.4/26.4 MB 299.4 kB/s eta 0:00:00Collecting opt_einsum2.3.2Downloading https://pypi.tuna.tsinghua.edu.cn/packages/23/cd/066e86230ae37ed0be70aae89aabf03ca8d9f39c8aea0dec8029455b5540/opt_einsum-3.4.0-py3-none-any.whl (71 kB)---------------------------------------- 71.9/71.9 kB 33.2 kB/s eta 0:00:00Requirement already satisfied: packaging in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from tensorflow) (26.2)Requirement already satisfied: protobuf8.0.0,6.31.1 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from tensorflow) (7.35.1)Requirement already satisfied: requests3,2.21.0 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from tensorflow) (2.34.2)Requirement already satisfied: setuptools in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from tensorflow) (83.0.0)Requirement already satisfied: six1.12.0 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from tensorflow) (1.17.0)Collecting termcolor1.1.0Downloading https://pypi.tuna.tsinghua.edu.cn/packages/33/d1/8bb87d21e9aeb323cc03034f5eaf2c8f69841e40e4853c2627edf8111ed3/termcolor-3.3.0-py3-none-any.whl (7.7 kB)Requirement already satisfied: typing_extensions3.6.6 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from tensorflow) (4.16.0)Collecting wrapt1.11.0Downloading https://pypi.tuna.tsinghua.edu.cn/packages/4a/62/ecc969b13b141fef89b888c9760821cb01a86ac8fc953911592c8e1e1522/wrapt-2.3.0-cp311-cp311-win_amd64.whl (80 kB)---------------------------------------- 80.9/80.9 kB 37.1 kB/s eta 0:00:00Requirement already satisfied: grpcio2.0,1.24.3 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from tensorflow) (1.83.0)Collecting keras3.12.0Downloading https://pypi.tuna.tsinghua.edu.cn/packages/8d/b3/c9b848bbdba18e765a8051917c0cc82585b64278ce87895f2e521a27438f/keras-3.15.1-py3-none-any.whl (2.4 MB)---------------------------------------- 2.4/2.4 MB 262.2 kB/s eta 0:00:00Collecting h5py3.15.0,3.11.0Downloading https://pypi.tuna.tsinghua.edu.cn/packages/db/0c/6c3f879a0f8e891625817637fad902da6e764e36919ed091dc77529004ac/h5py-3.14.0-cp311-cp311-win_amd64.whl (2.9 MB)---------------------------------------- 2.9/2.9 MB 347.9 kB/s eta 0:00:00Collecting ml_dtypes1.0.0,0.5.1Downloading https://pypi.tuna.tsinghua.edu.cn/packages/b4/24/70bd59276883fdd91600ca20040b41efd4902a923283c4d6edcb1de128d2/ml_dtypes-0.5.4-cp311-cp311-win_amd64.whl (210 kB)---------------------------------------- 210.7/210.7 kB 366.5 kB/s eta 0:00:00Collecting contourpy1.0.1Downloading https://pypi.tuna.tsinghua.edu.cn/packages/98/4b/9bd370b004b5c9d8045c6c33cf65bae018b27aca550a3f657cdc99acdbd8/contourpy-1.3.3-cp311-cp311-win_amd64.whl (225 kB)---------------------------------------- 225.2/225.2 kB 404.9 kB/s eta 0:00:00Collecting cycler0.10Downloading https://pypi.tuna.tsinghua.edu.cn/packages/e7/05/c19819d5e3d95294a6f5947fb9b9629efb316b96de511b418c53d245aae6/cycler-0.12.1-py3-none-any.whl (8.3 kB)Collecting fonttools4.28.2Downloading https://pypi.tuna.tsinghua.edu.cn/packages/08/60/defa5e69641db890a63be281f41345f4c33b157824eaf0b9fad3e08b0dcb/fonttools-4.63.0-cp311-cp311-win_amd64.whl (2.4 MB)---------------------------------------- 2.4/2.4 MB 300.3 kB/s eta 0:00:00Collecting kiwisolver1.3.1Downloading https://pypi.tuna.tsinghua.edu.cn/packages/be/6c/28f17390b62b8f2f520e2915095b3c94d88681ecf0041e75389d9667f202/kiwisolver-1.5.0-cp311-cp311-win_amd64.whl (73 kB)---------------------------------------- 73.5/73.5 kB 30.9 kB/s eta 0:00:00Requirement already satisfied: pillow9 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from matplotlib) (12.3.0)Collecting pyparsing3Downloading https://pypi.tuna.tsinghua.edu.cn/packages/10/bd/c038d7cc38edc1aa5bf91ab8068b63d4308c66c4c8bb3cbba7dfbc049f9c/pyparsing-3.3.2-py3-none-any.whl (122 kB)---------------------------------------- 122.8/122.8 kB 57.2 kB/s eta 0:00:00Requirement already satisfied: python-dateutil2.7 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from matplotlib) (2.9.0.post0)Collecting wheel1.0,0.23.0Downloading https://pypi.tuna.tsinghua.edu.cn/packages/87/1b/9e33c09813d65e248f7f773119148a612516a4bea93e9c6f545f78455b7c/wheel-0.47.0-py3-none-any.whl (32 kB)Requirement already satisfied: rich in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from keras3.12.0-tensorflow) (15.0.0)Collecting namexDownloading https://pypi.tuna.tsinghua.edu.cn/packages/b2/bc/465daf1de06409cdd4532082806770ee0d8d7df434da79c76564d0f69741/namex-0.1.0-py3-none-any.whl (5.9 kB)Collecting optreeDownloading https://pypi.tuna.tsinghua.edu.cn/packages/22/64/3cc7b08cb1c0f1949895f9490217ca8db6ced7f3bf75c65a5bf31c07bf1e/optree-0.19.1-cp311-cp311-win_amd64.whl (337 kB)---------------------------------------- 337.6/337.6 kB 261.9 kB/s eta 0:00:00Requirement already satisfied: charset_normalizer4,2 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from requests3,2.21.0-tensorflow) (3.4.9)Requirement already satisfied: idna4,2.5 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from requests3,2.21.0-tensorflow) (3.18)Requirement already satisfied: urllib33,1.26 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from requests3,2.21.0-tensorflow) (2.7.0)Requirement already satisfied: certifi2023.5.7 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from requests3,2.21.0-tensorflow) (2026.7.22)Requirement already satisfied: markdown-it-py2.2.0 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from rich-keras3.12.0-tensorflow) (4.2.0)Requirement already satisfied: pygments3.0.0,2.13.0 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from rich-keras3.12.0-tensorflow) (2.20.0)Requirement already satisfied: mdurl~0.1 in c:\users\administrator\appdata\local\programs\python\python311\lib\site-packages (from markdown-it-py2.2.0-rich-keras3.12.0-tensorflow) (0.1.2)Installing collected packages: namex, libclang, wrapt, wheel, termcolor, pyparsing, optree, opt_einsum, opencv-python, ml_dtypes, kiwisolver, h5py, google_pasta, gast, fonttools, cycler, contourpy, absl-py, matplotlib, astunparse, keras, tensorflowSuccessfully installed absl-py-2.5.0 astunparse-1.6.3 contourpy-1.3.3 cycler-0.12.1 fonttools-4.63.0 gast-0.7.0 google_pasta-0.2.0 h5py-3.14.0 keras-3.15.1 kiwisolver-1.5.0 libclang-18.1.1 matplotlib-3.11.1 ml_dtypes-0.5.4 namex-0.1.0 opencv-python-5.0.0.93 opt_einsum-3.4.0 optree-0.19.1 pyparsing-3.3.2 tensorflow-2.21.0 termcolor-3.3.0 wheel-0.47.0 wrapt-2.3.0准备好工具后下载数据源‌数据集获取‌下载Kaggle官方的Dogs vs. Cats数据集从中选取猫狗图片各1000~1500张数据清洗‌剔除损坏、非猫狗类的异常图片避免训练过程中断。预处理操作‌将所有图片统一resize为224×224尺寸按8:2比例划分为训练集和验证集完成数值归一化、图像随机翻转/亮度调节等增强操作提升模型泛化能力。在实现图片统一size的时候我们采用了批处理使用 OpenCV推荐稳定性高适合 AI 训练OpenCV 在处理图像解码异常和大规模批量任务时比 Pillow 更健壮且支持断点续传逻辑pip install opencv-python tqdm这个脚本实现了转换到这里就基本上实现了环境的搭建和数据源的基本准备条件。接下来就是开始做数据分配和训练的工作。预处理操作‌将所有图片统一resize为224×224尺寸按8:2比例划分为训练集和验证集完成数值归一化、图像随机翻转/亮度调节等增强操作提升模型泛化能力。由于我的电脑比较老旧所以我在数据源的处理上使用了1500张图的设计原则。那么用于训练就是1200张用于验证就是300张这样的一个比例操作。三、基于迁移学习训练基础模型‌预训练模型选择‌选用轻量型的MobileNet V2作为基础模型冻结大部分预训练层仅微调最后几层分类头大幅减少训练耗时。‌训练配置‌设置批量大小为16初始学习率设为0.0001训练轮次控制在20~30轮开启早停机制防止过拟合。‌训练监控‌训练过程中实时观察损失值和准确率变化在E431的CPU环境下该规模的数据集完整训练耗时约2~3小时。在这里呢就使用了这个脚本进行的训练。训练数据源是有格式要求的。设置相关参数会得到四、转换生成catsvsdogs.tflite文件‌模型导出‌将训练完成的Keras格式模型.h5或SavedModel格式保存到本地。‌格式转换‌使用TensorFlow官方的tf.lite.TFLiteConverter工具加载基础模型可选择动态范围量化策略进一步压缩模型体积、提升后续推理速度。‌导出验证‌转换完成后得到最终的catsvsdogs.tflite文件可通过加载模型输入单张测试图片验证分类输出结果是否正常。把生成的.h5文件进行转换这个脚本完成的这个功能。最终我们得到了当我们训练出来了这个模型后那么它的识别率有多么高呢需要通过图片来进行验证。我在网上找了相关的方案比如colab 在线网站操作比对结果出现各种问题不稳定。然后准备使用tf原生的官网方法应该是没有维护了导致不行。最终使用了streamlit 方案搭建了本地的验证网站。那么接下来就重点介绍这个方式。考虑到 的性能和开发便捷性推荐使用 ‌Python Streamlit‌。它无需编写 HTML/CSS/JS仅需几十行 Python 代码即可生成交互式网页且完美支持加载 TFLite 模型安装相关环境pip install streamlit pillow tensorflow这个地方比较坑多。使用如下脚本实现这个功能错误最多的就是非字符串什么的错误。运行D:\AIsource\archive\PetImagesstreamlit run app.pyXXXXX Uvicorn server started on :::8501You can now view your Streamlit app in your browser.Local URL: http://localhost:8501Network URL: http://192.168.106.191:8501Help agents write better Streamlit apps?Install the official Streamlit skills by running streamlit skills in your terminal.WARNING: All log messages before absl::InitializeLog() is called are written to STDERR最终的效果到这里整个训练验证的工程就完成了。