Coach CLI

This utility is responsible for managing user interaction with Coach's services.

Specifically you can:

• Sync local training data with remote
• Start and watch training sessions
• Locally evaluate models
• Download models

Installation

pip3 install -U coach-cli

To do anything we must be logged in

coach login
API Key: *****
Storage Key: *****
Storage Key Secret: *****

Usage

Usage: coach [OPTIONS] COMMAND [ARGS]...

  💖 Welcome to the Coach CLI Utility! 💖

  Grab your API keys and view example usage at:
  https://coach.lkuich.com

  Happy training! ⚽

Options:
  --help  Show this message and exit.

Commands:
  cache     Caches a model locally.
  download  Downloads remote training data locally.  
  login     Authenticates with Coach.
  ls        Lists synced projects in Coach.
  new       Uploads your local training directory to Coach.
  predict   Locally runs model prediction on specified image.
  rm        Deletes synced training data.
  status    Retreives the status of models.
  sync      Syncs a local data directory with Coach.
  train     Starts a Coach training session.

Examples

We're going to train a flowers dataset to recognise the following subjects:

• daisy
• dandelion
• roses
• sunflowers
• tulips

Start by downloading and extracting our dataset

wget https://storage.googleapis.com/download.tensorflow.org/example_images/flower_photos.tgz
tar -xvf flower_photos.tgz
mv flower_photos flowers # Our model's going to be called flowers

Note the structure of the dataset, this is important as it defines both the model and label names

flowers
    |-daisy
        |-10090824183_d02c613f10_m.jpg
        |-17040847367_b54d05bf52.jpg
        |-...
    |-dandelion
    |-roses
    |-sunflowers
    |-tulips

Now we're going to upload our dataset to Coach

coach new flowers

Note, if you make changes to this dataset, like delete some samples, you can sync your local directory with Coach by running

coach sync flowers

Now we're going to train. We must specify the name of our synced project, the number of training steps, and the base module to use for transfer learning.
It's typically best to start high with training steps. The default is 1000, and will do fine for this example. We're going to use the default mobilenet_v2_100_224 as our base model, since this gives us a decent tradeoff between final model size, speed, and quality. Make sure to consult the help docs to find our more about supported base modules and find the right fit for your model based on your needs.

coach train flowers
# OR: coach train flowers --steps 1000 --module mobilenet_v2_100_224

This will start a new training session. You can monitor its progress with the status command, by default with no arguments it'll show the status of all models. Since we're just interested in our flowers model, well run with the --model parameter

coach status --model flowers
-----------------------------------------------------
flowers     |    Training   |     2019-06-15 14:30:20
-----------------------------------------------------

Once complete, your status should look something like this

coach status --model flowers
-----------------------------------------------------
flowers     |    Completed   |    2019-06-15 14:32:00
-----------------------------------------------------

Now we can download and run our model locally on some training data

coach cache flowers # Only have to run once
coach predict flowers/roses/13342823005_16d3df58df_n.jpg flowers
{ roses: 0.90, tulips: 0.05, sunflowers: 0.03, daisy: 0.01, dandelion: 0.01 }

Our cached model is stored in:
~/.coach/models/flowers

Conclusions

That's all it takes to train a model end-to-end!

For implementation, check out our client side SDK's:

• Python
• .NET
• Unity