Minimized version from original source: https://github.com/AdityaKurude/Tensorflow-Object-Detection-API-on-AWS-Greengrass

Plus special notes when using EC2 as GGC.

1. Set up Greengrass Core - GGC v1.6.0, using Amazon EC2.

2. Install dependencies for TensorFlow on GGC:

   $ sudo pip install numpy scipy tensorflow pillow

Use sudo and --no-cache-dir option to install on EC2, python2.7.

Package | Local-macOS| EC2-Linux ------------ | ------------- Numpy | 1.15.2 | 1.15.2 Scipy | 1.1.0 | 1.1.0 Tensorflow | 1.10.0 | 1.11.0 Pillow | 5.3.0 | 5.3.0

3. Download ssd_mobilenet_v1_coco from [the TF model zoo] (https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md) and use frozen_inference_graph.pb to create zip file: tf_models.zip. Upload zip file to S3:

   • S3 bucket name: e.g. qt_greengrass
   • Note: Download the label map for the model (models/mscoco_label_map.pbtxt) from [TF github] (https://github.com/tensorflow/models/tree/master/research/object_detection/data)

4. [Lambda console] Create and Publish Lambda function using source from this repo:

   • Name: greengrassObjectDetection
   • Handler: greengrassObjectDetection.function_handler

5. Add Lambda function to Greengrass group

   • Memory limit: 512MB (observed on MacOS Activity Monitor ~325MB)
   • Timeout: 10 seconds

6. [Greengrass Group] Add Subscription: Lambda -> IoTCloud (Lambda publish message at topic "hello/world")

7. Greengrass console > Add machine learning resource

   • Model from S3: select file from step 3
   • Local path: /greengrass-machine-learning/tf (must follow what is declared in Lambda function)