1. Github: https://github.com/hailusong/models/blob/master/research/object_detection/samples/configs
2. The one we use is faster_rcnn_resnet101_pets.config
   • See this about Google Object API custom training
3. Details about the pipeline.config as well as protobuf
   • See 3 steps to update parameters of Faster R-CNN/SSD models in TensorFlow Object Detection API

See src/dlib

1. god_idclass_gcs_model.ipynb
   • Setup GCS detection model for Google Object Detection API custom training
   • Setup also include pipeline configuration file and train/test TF record files
2. god_idclass_gcs.ipynb
   • Setup GCS (anything other than detection model) for Google Object Detection API custom training
3. god_idclass_colabtrain.ipynb
   • Train Google Object Detection API with custom data and pipeline configuration on CoLab
4. god_idclass_mlabtrain.ipynb
   • Train Google Object Detection API with custom data and pipeline configuration on Google Cloud ML
5. god_idclass_export.ipynb
   • Export the custom train result, a checkpoint model of Google Object Detection API, for inference
6. god_idclass_colabeval.ipynb
   • object detection inference using exported model
   • note that in the legacy Google Object Detection API, you need to run train and eval at the same time in separated processes
   • in the newer Google Object Detection API (>=1.13.1), it is one run (via model_main/py or TPU version) and trigger BOTH train/eval at the same time
   • In tensorboard, all metrics are on Valid dataset, NOT on Train dataset, including IMAGEs
7. god_idclass_flask.ipynb
   • Run inference as a REST service in Flask

1. Load frozen Graph
   • Create Graph object
   • Create Graph Definition object
   • Load frozen graph using Graph Definition object
   • Import Graph Definition object into Graph object
2. Locate all tensors/ops we need:
   • Image input tensor/op: image_tensor:0
   • Output tensor/op: detection_boxes
   • Output tensor/op: detection_masks
   • Output tensor/op: num_detections
3. Compute the graph within a TF session
   • Set the session input dict (a.k.a. feed_dict) to the image to be inferenced
   • From the session output dict (a.k.a. session output) fetch values from those tensors
   • Plot the result