[object_detection] Adding a handy Colab Notebook for training a custom pets detector model using the latest API

Prerequisites

Please answer the following question for yourself before submitting an issue.

• I checked to make sure that this issue has not been filed already.

I prepared a comprehensive Colab Notebook so that the onboarding experience becomes as seamless as possible. The Colab Notebook mere acts as a base VM to:

• Prepare the data
• Copy the data over to a GCS bucket
• Launch training and evaluation jobs (using Cloud TPUs and GPUs respectively)
• Monitor results with TensorBoard within the Colab Notebook itself

I have used the following references to prepare this notebook:

• https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/running_pets.md
• https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/preparing_inputs.md
• Training and serving a realtime mobile object detector in 30 minutes with Cloud TPUs
• https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/tf2_training_and_evaluation.md

In this document it is suggested to use 3.6 as the Python version which is actually no longer supported by AI Platform and it will cause the following error:

ERROR: (gcloud.ai-platform.jobs.submit.training) INVALID_ARGUMENT: Field: python_version Error: The specified Python version '3.6' is not supported.
- '@type': type.googleapis.com/google.rpc.BadRequest
  fieldViolations:
  - description: The specified Python version '3.6' is not supported.
    field: python_version

The fix is to change it to 3.7 and everything will be golden. I hope this Colab Notebook is going to be useful. I plan to author more notebooks showing more involved data preparation processes & popping them up on a GPU for training and that might be even more helpful for the community.

Question about export the TF 2 compatible models

I could not export the custom trained models because it appears to me that it is not documented yet anywhere. Please correct me if I am wrong. I would specifically be interested in exporting the checkpoints as a SavedModel file and then converting it to a TF Lite model.

Update: I was able to export the checkpoints after referring https://github.com/tensorflow/models/issues/8841:

python object_detection/exporter_main_v2.py \
    --input_type encoded_image_string_tensor \
    --trained_checkpoint_dir gs://$YOUR_GCS_BUCKET/$MODEL_DIR \
    --output_directory `pwd` \
    --pipeline_config_path gs://$YOUR_GCS_BUCKET/data/$PIPELINE_CONFIG_PATH

I was able to run inference with the model checkpoints as well. Here's a separate notebook that shows the process. I think it might useful for folks that are just interested in detection models only.

Question: What is the preferred format for running inference? Checkpoints or SavedModel? I find SavedModel to be more convenient with the respect of the TensorFlow ecosystem.