Skip to content

C. Run on Your Use Case

This page guides you through the process of running the app on your data and pipelines, using Docker. Different dataset and text classification models can be supported in Azimuth.

Launch Azimuth with no pipeline, or with multiple pipelines

Azimuth supports specifying no pipelines, to only perform dataset analysis. It also supports supplying mulitple pipelines, to allow for quick comparison. However, only one dataset per config is allowed.

The simplest scenario is if you have a HuggingFace (HF) dataset and model. For the sake of simplicity, we explain the instructions to run the app with this scenario. However, you will quickly need to learn about the Configuration details and Custom Objects to launch more complex use cases.

1. Prepare the Config File

Run our demo first

You haven't run our demo yet? You might want to verify your setup before feeding your own model and dataset. Go back to B. Learn Basics.

Start from an existing config and edit the relevant fields to adapt it to your dataset and models. Examples with an HuggingFace (HF) dataset and model are available in config/examples (CLINC is also shown below).

  1. Put your model checkpoint (results of .save_pretained()) under the folder azimuth_shr.
  2. In config, copy config/examples/clinc_oos/conf.json to a new folder with your project name. Ex: config/my_project/conf.json.

  3. Edit the config:

    1. name: put your project name.
    2. dataset.args: specify the args required to load your dataset with datasets.load_dataset.
    3. Edit columns and rejection_class based on the dataset.
    4. pipelines.models.kwargs.checkpoint_path: put your own checkpoint path to your model. The path should start with /azimuth_shr, since this folder will be mounted on Docker.
    5. Edit the saliency_layer so it is the name of the input layer of the model. It should be set to null if your model is not from PyTorch or without a word-embedding layer.

    Links to full reference

    If you need more details on some of these fields:

{
  "name": "CLINC150", # (1)
  "dataset": {
    "class_name": "datasets.load_dataset", # (2)
    "args": [ # (3)
        "clinc_oos",
        "imbalanced"
    ]
  },
  "columns": { # (4)
    "text_input": "text",
    "label": "intent"
  },
  "rejection_class": "oos", # (5)
  "model_contract": "hf_text_classification", # (6)
  "pipelines": [ # (7)
    {
      "model": {
        "class_name": "loading_resources.load_hf_text_classif_pipeline", # (8)
        "remote": "/azimuth_shr", # (9)
        "kwargs": { # (10)
          "checkpoint_path": "transformersbook/
                              distilbert-base-uncased-distilled-clinc"
        }
      }
    }
  ],
  "saliency_layer": "distilbert.embeddings.word_embeddings", # (11)
}
  1. Name for your project. Shown in the application to identify your config.
  2. If the dataset is a HF dataset, use this class_name.
  3. kwargs to send to the class_name.
  4. Specify the name of the dataset columns, such as the column with the utterance and the label.
  5. Specify the value if a rejection option is present in the classes.
  6. If the pipeline is a HF pipeline, use this model_contract.
  7. Multiples ML pipelines can be listed to be available in the webapp.
  8. If this a HF pipeline, use this class_name.
  9. Change only if class_name is not found in /azimuth_shr.
  10. kwargs to send to the class. Only checkpoint_path if you use the class above.
  11. Name of the layer on which to compute saliency maps.

2. Running the App

  1. In the terminal, go to the azimuth root directory.
  2. Set CFG_PATH=/config/my_project/conf.json with the location of the config.
    • The initial / is required as your local config folder will be mounted on the Docker container at the root.
  3. Execute the following command: 
    make launch
  4. The app will be accessible at http://localhost:8080 after a few minutes of waiting. The start-up tasks will start.

Advanced Settings

Additional Config Fields

The Configuration reference details all additional fields that can be set, such as changing how behavioral tests are executed, the similarity analysis encoder, the batch size and so on.

Environment variables

No matter where you launch the app from, you can always configure some options through environment variables. They are all redundant with the config attributes, so you can set them in either place. They are the following:

  • Specify the threshold of your model by passing TH (ex: TH=0.6 or NaN if there is no threshold) in the command. If multiple pipelines are defined, the threshold will apply to all.
  • Similarly, pass TEMP=Y (ex: TEMP=3) to set the temperature of the model.
  • Disable behavioral tests and similarity by passing respectively BEHAVIORAL_TESTING=null and SIMILARITY=null.
  • Specify the name of the project, passing NAME.
  • You can specify the device on which to run Azimuth, with DEVICE being one of auto, gpu or cpu. If none is provided, auto will be used. Ex: DEVICE=gpu.
  • Specify READ_ONLY_CONFIG=1 to lock the config once Azimuth is launched.

Config file prevails over environment variables

Remember that the values above are defined in the config too. If conflicting values are defined, values from the config file will prevail.