Finding Nemo with Tensorflow on Google Cloud ML Engine

Prerequisits:

• Installed Tensorflow and Object Detection API
• Having set up Google Cloud ML Engine

Steps are as follows:

1. Prepare the dataset
   • Collect images
   • Label the images
   • Generate TFRecords
2. Select a pre-trained model and configure the pipeline
3. Train the model
   • Upload files to the Google Cloud
   • Submit the training and evaluation jobs
4. Use the model

Prepare the dataset

The training and testing image data is fed in form of TFRecords.

Collect and prepare images

In this example, the image set was built using images from Imgenet and Google images. (Didn't feel like diving). If you are using imgenet, you can save the urls from the Downloads tab to a .txt file and use the following to download the images.

$ get -i url_list.txt

Thats it! simple and painless.

After selecting appropriate images, divide them for training, testing and validating the model. In this example, 190 images were used as follows:

• Train - 150
• Test - 30
• Validate - 10

Also, there were more than one clownfish in some images, therefore actal number of exmples was larger.

Label the images

For labeling the object we are interested in, (a clown fish in this case) LabelImg was used.

After labeling is done, we have xml files for each of the image in the training and testing sets.

Generate TFRecords

For creating the TFRecords, pascal_to_tfrecord.py was used. The script is from Object detection in the context of permaculture.

Select a pre-trained model and configure the pipeline

For this example, faster_rcnn_resnet101_coco_11_06_2017 was used, following the Quick Start: Distributed Training on the Oxford-IIIT Pets Dataset on Google Cloud. For configuring the pipeline, one of the example config files provided with the API was used (available here) with changes to the parameters according to the training job. I used the faster_rcnn_resnet101_pets.config. The essential points to change are as follows.

• num_classes: 1
  • We are interested in finding only one object, a clownfish. Therefore, the above variable is 1.
• fine_tune_checkpoint: "gs://my_bucket/data/model.ckpt"
  • This points to the trained model checkpoint file which needs to be uploaded to our storage bucket on the cloud.
• input_path: "gs://your_bucket/data/train.record"(under train_input_reader: for the training set)
• label_map_path: "gs://your_bucket/data/labelmap.pbtxt" (under train_input_reader: for the training set)
• input_path: "gs://your_bucket/data/test.record"(under eval_input_reader: for the testing set)
• label_map_path: "gs://your_bucket/data/labelmap.pbtxt" (under eval_input_reader: for the testing set)

Also, it's reocommended to run a few number of steps first to see how everything works and to see how the loss is reducing. To do this change the num_steps: parameter. I first tried with 500 steps.
In addition to this, there are many other parameters that can be changed to get the best performance such as learning_rate. But, the default values are okay to start and get things going.

Train the model

Upload files to theGoogle Cloud

To do the training on cloud, following structure is suggested for the storage bucket.

your_bucket
	+data
    	    -train.record
            -test.record
            -labelmap.pbtxt
            -model.ckpt.data
            -model.ckpt.meta
            -model.ckpt.index
            -faster_rcnn_resnet101_pets.config
   +train
   +test

Submit training and evaluation jobs

Packaging

# From tensorflow/models/research/
bash object_detection/dataset_tools/create_pycocotools_package.sh /tmp/pycocotools
python setup.py sdist

and,

cd slim && python setup.py sdist

Training

#From tensorflow/models/research
gcloud ml-engine jobs submit training object_detection_`date +%m_%d_%Y_%H_%M_%S` \
--runtime-version 1.8 \
--job-dir=gs://your_bucket/train/ \
--packages dist/object_detection-0.1.tar.gz,slim/dist/slim-0.1.tar.gz,/tmp/pycocotools/pycocotools-2.0.tar.gz \
--module-name object_detection.train \
--region us-central1 \
--config object_detection/samples/cloud/cloud.yml \
-- \
--train_dir=gs://your_bucket/train/ \
--pipeline_config_path=gs://your_bucket/data/faster_rcnn_resnet101_pets.config

Evaluating

#From tensorflow/models/research
gcloud ml-engine jobs submit training object_detection_eval_`date +%m_%d_%Y_%H_%M_%S` \
--job-dir=gs://your_bucket/train \
--packages dist/object_detection-0.1.tar.gz,slim/dist/slim-0.1.tar.gz,/tmp/pycocotools/pycocotools-2.0.tar.gz \
--module-name object_detection.eval \
--region us-central1 \
--scale-tier BASIC_GPU \
--runtime-version=1.8 \
-- \
--checkpoint_dir=gs://your_bucket/train \
--eval_dir=gs://your_bucket/test \
--pipeline_config_path=gs://your_bucket/data/faster_rcnn_resnet101_pets.config

Note: I ran in to some errors, which are apprently related to deploying replicas. As a workaround, I removed the workers. Therefore, the cloud.yml file looked like this. Since this is a simpler training job (one catagory, small no. of imges) this did not affect very much.