7.9.37. Coffee Bean Classification

Classify coffee beans by roast level from images, using a MobileNetV1 model on the MSPM0G5187 with NPU.

Example folder: examples/coffee_bean_classification (GitHub)

7.9.37.1. Overview

Task: image classification (image_classification, target_module: vision), 4 classes
Application: coffee bean roast and quality inspection from color, texture and surface appearance
Device: MSPM0G5187 (config_MSPM0.yaml, the only config in the folder). The folder contains only the config and no readme.
Model: MobileNetV1_28k_NPU

Note

In 1.4.0 this example used MobileNetV1_58k_NPU. In 1.5.0 the config uses MobileNetV1_28k_NPU.

7.9.37.2. The Dataset

The data is the Kaggle Coffee Bean Dataset Resized (224 x 224) with four classes: Dark, Green, Light and Medium. Check the Kaggle page for the license and usage terms. The config downloads the prepared dataset from https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/coffee_bean_classification.zip. The layout is a classes/ folder with one sub-folder per class.

7.9.37.3. Feature Extraction

The config uses the preset CoffeeBean_Default: RGB images (3 channels) resized to 128x128 and normalized with the ImageNet mean (0.485, 0.456, 0.406) and scale (0.229, 0.224, 0.225). The preset CoffeeBean_Augmentation_Default adds random horizontal and vertical flips, random rotation and color jitter if you want augmentation.

7.9.37.4. Running the Example

Install Tiny ML ModelZoo first (Installation, Quickstart), then run from the tinyml-modelzoo folder:

./run_tinyml_modelzoo.sh examples/coffee_bean_classification/config_MSPM0.yaml

See Running Examples for all launchers, for example:

run_tinyml_modelzoo.bat examples\coffee_bean_classification\config_MSPM0.yaml

7.9.37.5. Configuration

common:
  target_module: vision
  task_type: image_classification
  target_device: MSPM0G5187
dataset:
  dataset_name: coffee_bean_classification
  input_data_path: https://software-dl.ti.com/C2000/esd/mcu_ai/datasets/coffee_bean_classification.zip
data_processing_feature_extraction:
  feature_extraction_name: CoffeeBean_Default
training:
  model_name: MobileNetV1_28k_NPU
  batch_size: 64
  training_epochs: 30
  num_gpus: 1
  quantization: 2
  learning_rate: 0.1
testing: {}
compilation: {}

num_gpus: 1 expects a GPU; use 0 for CPU. The image models enabled in 1.5.0 are Lenet5, MobileNetV1_28k_NPU and MobileNetV2_58k_NPU. For the image task type see Image Classification.

7.9.37.6. Results

There is no readme with published results. In the automated 1.5.0 config test run (test set of 160 images, 40 per class) the quantized model reached a validation accuracy of 98.75% (F1 0.988, AUC ROC 0.972) and a test accuracy of 99.38% (AUC ROC 0.966). This is a small test set, so treat it as indicative.

7.9.37.7. Output Location

Results are written to ./data/projects/coffee_bean_classification/run/<date-time>/MobileNetV1_28k_NPU/. See Understanding the Run Output and Project Compilation Guide: Timeseries Classification Examples for deployment with the MSPM0 SDK and a camera interface for image capture.