zh/integrations/tflite/

[giscus[bot]]

zh/integrations/tflite/

了解如何将YOLO11 模型转换为 TFLite，用于边缘设备部署。优化性能，确保在各种平台上无缝执行。

https://docs.ultralytics.com/zh/integrations/tflite/

[xiao5676wen]

导出错误啊！
C:\ProgramData\anaconda3\envs\export1\python.exe C:\Users\Administrator\PycharmProjects\pythonProject\aaa.py
Downloading https://github.com/ultralytics/assets/releases/download/v8.3.0/yolo11n.pt to 'yolo11n.pt'...
100%|██████████| 5.35M/5.35M [00:00<00:00, 7.98MB/s]
Ultralytics 8.3.57 🚀 Python-3.12.0 torch-2.5.1+cpu CPU (Intel Core(TM) i7-6700K 4.00GHz)
YOLO11n summary (fused): 238 layers, 2,616,248 parameters, 0 gradients, 6.5 GFLOPs

PyTorch: starting from 'yolo11n.pt' with input shape (1, 3, 640, 640) BCHW and output shape(s) (1, 84, 8400) (5.4 MB)

TensorFlow SavedModel: starting export with tensorflow 2.18.0...
WARNING:tensorflow:From C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\tf_keras\src\losses.py:2976: The name tf.losses.sparse_softmax_cross_entropy is deprecated. Please use tf.compat.v1.losses.sparse_softmax_cross_entropy instead.

ONNX: starting export with onnx 1.17.0 opset 19...
ONNX: slimming with onnxslim 0.1.31...
ONNX: export success ✅ 1.7s, saved as 'yolo11n.onnx' (10.3 MB)
TensorFlow SavedModel: starting TFLite export with onnx2tf 1.26.3...
WARNING: All log messages before absl::InitializeLog() is called are written to STDERR
I0000 00:00:1736044019.708593 14188 devices.cc:76] Number of eligible GPUs (core count >= 8, compute capability >= 0.0): 0 (Note: TensorFlow was not compiled with CUDA or ROCm support)
I0000 00:00:1736044019.708768 14188 single_machine.cc:361] Starting new session
W0000 00:00:1736044020.665041 14188 tf_tfl_flatbuffer_helpers.cc:365] Ignored output_format.
W0000 00:00:1736044020.665086 14188 tf_tfl_flatbuffer_helpers.cc:368] Ignored drop_control_dependency.
I0000 00:00:1736044021.563802 14188 devices.cc:76] Number of eligible GPUs (core count >= 8, compute capability >= 0.0): 0 (Note: TensorFlow was not compiled with CUDA or ROCm support)
I0000 00:00:1736044021.563934 14188 single_machine.cc:361] Starting new session
W0000 00:00:1736044022.460546 14188 tf_tfl_flatbuffer_helpers.cc:365] Ignored output_format.
W0000 00:00:1736044022.460593 14188 tf_tfl_flatbuffer_helpers.cc:368] Ignored drop_control_dependency.
TensorFlow SavedModel: export failure ❌ 43.0s: No module named 'imp'
Traceback (most recent call last):
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\ultralytics\engine\exporter.py", line 1271, in _add_tflite_metadata
from tensorflow_lite_support.metadata import metadata_schema_py_generated as schema # noqa
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ModuleNotFoundError: No module named 'tensorflow_lite_support'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "C:\Users\Administrator\PycharmProjects\pythonProject\aaa.py", line 7, in
model.export(format="tflite") # creates 'yolo11n_float32.tflite'
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\ultralytics\engine\model.py", line 738, in export
return Exporter(overrides=args, _callbacks=self.callbacks)(model=self.model)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\ultralytics\engine\exporter.py", line 397, in call
f[5], keras_model = self.export_saved_model()
^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\ultralytics\engine\exporter.py", line 170, in outer_func
raise e
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\ultralytics\engine\exporter.py", line 165, in outer_func
f, model = inner_func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\ultralytics\engine\exporter.py", line 1024, in export_saved_model
f.unlink() if "quant_with_int16_act.tflite" in str(f) else self._add_tflite_metadata(file)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\ultralytics\engine\exporter.py", line 1274, in add_tflite_metadata
from tflite_support import metadata # noqa
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\tflite_support\metadata.py", line 28, in
from tflite_support import flatbuffers
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\tflite_support\flatbuffers_init.py", line 15, in
from .builder import Builder
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\tflite_support\flatbuffers\builder.py", line 15, in
from . import number_types as N
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\tflite_support\flatbuffers\number_types.py", line 18, in
from . import packer
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\tflite_support\flatbuffers\packer.py", line 22, in
from . import compat
File "C:\ProgramData\anaconda3\envs\export1\Lib\site-packages\tflite_support\flatbuffers\compat.py", line 19, in
import imp
ModuleNotFoundError: No module named 'imp'

进程已结束，退出代码为 1
imp和tensorflow_lite_support都无法安装啊！

[xiao5676wen]

我用的python3.12

[glenn-jocher]

@xiao5676wen 感谢您的反馈！YOLO11 当前与 Python 3.12 可能存在兼容性问题。建议切换至 Python 3.10 或 3.11 并确保安装最新的 tensorflow 和 tensorflow_lite_support。如果问题持续，请查看 TFLite 导出指南 或提交详细的错误报告以获得进一步支持。

[UltralyticsAssistant]

👋 Hello, thank you for your interest in Ultralytics 🚀! It looks like you're exploring YOLO11 model integration with TFLite for edge device deployment 🤖. We encourage you to check out the 中文文档 for detailed guidance on this topic.

If you encounter any issues 🐛, please provide a minimum reproducible example so our team can better assist you.

If this is a ❓ question about optimization or edge deployment, please include additional details about your device specifications, configurations, and any relevant logs to help us provide personalized guidance. Also, ensure you're following general conversion and export tips for best results.

Join the Ultralytics community where it suits you best! Engage in real-time discussions on our Discord 🎧, explore in-depth topics on Discourse, or exchange ideas on our Subreddit.

Upgrade

Ensure you're using the latest version of the ultralytics package to access all the newest features and fixes. To upgrade, simply run:

pip install -U ultralytics

Environments

YOLO may be run and experimented with in verified environments. Try free GPU-enabled Notebooks for seamless experimentation:

Status

When green, this badge indicates that all Ultralytics CI tests are passing, ensuring compatibility and stability across platforms.

This is an automated response to kickstart the discussion, but an Ultralytics engineer will assist you soon! Let us know if there's anything else you'd like to clarify in the meanwhile. 😊

[xiao5676wen]

我的ultralytics已经是最新版本了，更新不了了，就是导出tflite格式出现我之前的错误

[glenn-jocher]

@xiao5676wen 感谢反馈！导出至 TFLite 出现问题时，建议检查以下几点：1) 确保 TensorFlow 版本兼容，建议使用最新版本；2) 确保 ultralytics 包已正确安装并为最新版本；3) 验证模型支持 TFLite 导出。如果问题仍未解决，请提供完整的错误信息和使用的命令，我们将进一步协助您！

[xiao5676wen]

3. 验证模型支持 TFLite 导出：我怎么才能知道模型支持 TFLite 导出呢？不支持有什么办法解决问题呢？

[glenn-jocher]

感谢您的提问！要验证模型是否支持 TFLite 导出，请确保所用的 YOLO11 模型具备导出功能（如 yolo11n.pt 等已测试支持的模型）。如果导出失败，可能是模型架构不兼容或需要额外的 TensorFlow 配置。可以尝试更新 TensorFlow 版本或切换到支持导出的模型。如有进一步问题，请参考 TFLite 导出指南。

[longzaixiang222]

from ultralytics import YOLO
model = YOLO("best5.pt")
model.export(format="tflite")
模型导出后识别效果不好
使用YOLO("best5_float32.tflite") 这种加载方式，效果就很好

但我使用：
interpreter = tf.lite.Interpreter(model_path="best5_float32.tflite")
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
识别效果不佳，部分类型识别不出来，原图尺寸640*640

请问下是您们内部对图片做了什么处理吗？

[glenn-jocher]

@longzaixiang222 thank you for your question! The difference occurs because when using YOLO("best5_float32.tflite"), we automatically handle preprocessing including image scaling, normalization (pixel values /255), and channel reordering (BGR→RGB). When using the TFLite Interpreter directly, you'll need to manually apply these preprocessing steps.

For implementation details, please refer to our TFLite export documentation which explains the required input normalization and channel order. Ensure your input tensor matches the expected shape (1,640,640,3) with float32 dtype and normalized values.

[longzaixiang222]

thanks

[glenn-jocher]

Great observation! When using the TFLite Interpreter directly, you'll need to manually replicate the preprocessing steps our YOLO class handles automatically. Ensure you're resizing images to 640x640, normalizing pixel values to 0-1 (divide by 255), and converting BGR→RGB before inference. The input tensor should match shape (1,640,640,3) with float32 dtype. For implementation details, refer to our TFLite preprocessing requirements.

[diamondeyess]

我的导出设置是：ormat = "tflite", device = "0", int8 = True, data =.....data.yaml, 为什么输出：fully_quantize: 0, inference_type: 6, input_inference_type: FLOAT32, output_inference_type: FLOAT32 ，没有完全量化吗？

[glenn-jocher]

@diamondeyess 这是正常的TFLite INT8量化行为！输入输出保持FLOAT32是为了保持精度和兼容性，而模型内部计算已经量化为INT8以提升性能和减小模型大小。这种"混合量化"可以在保持推理精度的同时获得INT8的性能优势。如需完全INT8量化（包括输入输出），需要特定的TFLite配置，但通常不推荐，因为可能影响精度。

[diamondeyess]

好的，谢谢！后面输出还有一段：fully_quantize: 0, inference_type: 6, input_inference_type: INT8, output_inference_type: INT8 就是您说的完全量化吗？

[glenn-jocher]

@diamondeyess 是的，第二个输出（input_inference_type: INT8, output_inference_type: INT8）就是完全量化！这意味着模型的输入和输出也被量化为INT8。完全量化可以进一步减小模型大小和提升推理速度，但需要注意输入数据的预处理要匹配INT8格式。在使用时，您可能需要将输入数据从FLOAT32转换为INT8范围，并在推理后将输出从INT8转换回FLOAT32以便后处理。

[longzaixiang222]

您好老师，同一个模型，转换成tflite，同一张图片，每2秒识别一次，和单次识别为什么结果会不样？

[longzaixiang222]

每二秒识别，是通过录屏生成的图片，单次识别，是录屏后保存到本地的图片

[glenn-jocher]

@longzaixiang222 这种差异通常由以下原因造成：1) 录屏图像可能存在压缩损失或格式转换；2) 录屏帧和保存的图片可能在色彩空间、分辨率或编码上有细微差别；3) 图像预处理路径不同。建议检查两种图像的像素值、尺寸和格式是否完全一致，确保使用相同的预处理步骤。您可以尝试保存录屏帧并与本地图片进行像素级对比来定位差异源。

[longzaixiang222]

好的，谢谢了老师

[glenn-jocher]

@longzaixiang222 不客气！如有其他关于YOLO11或TFLite导出的问题，随时欢迎继续提问。祝您项目顺利！