AIoT class three
適用於 Espressif 晶片組的 TensorFlow Lite Micro
依照 ESP-IDF 入門指南的說明設定工具鏈和 ESP-IDF 本身。接下來的步驟假設此安裝成功且 IDF 環境變數已設定。具體來說, 設定 IDF_PATH 環境變數
install.sh
. .export.sh
在 ESP-IDF 專案中執行以下命令來安裝此元件:
cd esp-idf
idf.py add-dependency "esp-tflite-micro"
在console中執行以下命令來安裝此元件:
pip install
若要取得範例,請執行以下命令:
idf.py create-project-from-example "esp-tflite-micro:hello_world"
cd hello_world
idf.py set-target esp32s3
idf.py build
若要取得範例,請執行以下命令:
idf.py create-project-from-example "esp-tflite-micro:micro_speech"
cd micro_speech
idf.py set-target esp32s3
idf.py build
- 分離網路實做及量化實做,專注在神經算子開發
- 有google 提供大量的tflite 模型可用
- 效能比較:在各種晶片組上測量的 ESP-NN 優化的快速總結
| Target | TFLite Micro Example | without ESP-NN | with ESP-NN | CPU Freq |
|---|---|---|---|---|
| ESP32-S3 | Person Detection | 2300ms | 54ms | 240MHz |
| ESP32 | Person Detection | 4084ms | 380ms | 240MHz |
| ESP32-C3 | Person Detection | 3355ms | 426ms | 160MHz |
TensorFlow Lite Micro 目前仅支持有限的 TensorFlow 算子,因此可运行的模型也有所限制。Google 正致力于在参考实现和针对特定结构的优化方面扩展算子支持。Arm 的 CMSIS-NN 开源加速库也为算子的支持和优化提供了另一种可能。 已支持的算子在文件
https://github.com/tensorflow/tensorflow/blob/5e0ed38eb746f3a86463f19bcf7138a959ddb2d4/tensorflow/lite/micro/all_ops_resolver.cc
-
Kernelwise performance on ESP32-S3 chip
- Numbers are ticks taken for kernel to execute
- Chip config: 240MHz, SPI: QPI 80MHz, Data cache: 64KB
Function ANSI C Optimized Opt Ratio Data info Memory elementwise_add 312327 71644 4.36 size = 1615 External elementwise_mul 122046 30950 3.95 size = 1615 External convolution 4642259 461398 10.06 input(10,10), filter(64x1x1x64), pad(0,0), stride(1,1) External convolution 300032 43578 6.9 input(8,8), filter(16x1x1x16), pad(0,0), stride(1,1) External convolution 2106801 643689 3.27 input(10,10), filter(64x3x3x3), pad(0,0), stride(1,1) External depthwise conv 1192832 191931 6.2 input (18, 18), pad(0,0), stride(1,1) filter: 1x3x3x16 External depthwise conv 1679406 366102 4.59 input (12, 12), pad(1,1), stride(1,1) filter: 8x5x5x4 External max pool 485714 76747 6.33 input(16,16), filter (1x3x3x16) Internal avg pool 541462 160580 3.37 input(16,16), filter (1x3x3x16) Internal fully connected 12290 4439 2.77 len: 265, ch = 3 Internal prelu (relu6) 18315 1856 9.87 size, 1615 Internal
-
To configure, please use
idf.py menuconfigand underESP-NNselectNN_OPTIMIZATIONS -
There are two options presented:
- Optimized versions
- ANSI C
-
Default selection is for
Optimized versions. For ESP32-S3, assembly versions are automatically selected, whereas for other chips (viz., ESP32, ESP32-C3), generic optimisations are selected. -
For debugging purposes, you may want to select
ANSI Creference versions.
若要取得範例,請執行以下命令:
idf.py create-project-from-example "esp-tflite-micro:person_detection"
cd person_detection
idf.py set-target esp32s3
idf.py build
同步到最新的 TFLite Micro 上游,根據上游儲存庫策略,tflite-lib 會複製到此儲存庫中的元件目錄中。我們會不時將其更新到最新的上游版本。在任何情況下,如果您希望在本地更新它,您可以執行
scripts/sync_from_tflite_micro.sh
https://dejazzer.com/eece4710/ https://dejazzer.com/eece4710/index.html#3_resources





