rime-sdk 0.20.0rc0

Package to programmatically access a RIME deployment

Overview

The rime-sdk package contains Python (RIME SDK) and command-line interfaces (RIME CLI).

RIME SDK

The RIME SDK provides an interface to RIME backend services for starting and viewing the progress of RIME stress test jobs. There are four objects available in the rime-sdk package:

• Client
• Job
• Project
• Firewall

To use these objects, import them from the package like so:

from rime_sdk import Client, Job, Project, Firewall

Client

The Client provides an interface to RIME's backend services for creating projects, starting stress test jobs, and querying the backend for current stress test jobs. To initialize the Client, provide the address of your RIME instance.

rime_client = Client("my_vpc.rime.com", "api-key")

start_stress_test()

This allows you to start an AI Stress Test job on the RIME backend.

Arguments:

• test_run_config: dict

  Specifies paths to the model and dataset to be used in the stress test.

• project_id: Optional[str] = None

  Specify the project to file the stress test result under. If omitted, the stress test result will be stored in the default project.

• rime_managed_image: Optional[str] = None

  Specify the name of the RIME Managed Image to run the stress test on. Managed Images are the preferred method of running RIME on an image with custom Pip requirements. See the documentation for create_managed_image() and list_managed_images() for further information.

• custom_image: Optional[CustomImage] = None

  Specify a custom Docker image to run the stress test job on. This image could include custom libraries that your model depends on. If no custom image is provided, the backend will use the default image specified by the cluster configuration.

• ram_request_megabytes: Optional[int] = None

  Specify the megabytes of RAM to request for a stress test job. If none specified, will default to 4000MB. The limit will be set as 2x the request.

• cpu_request_millicores: Optional[int] = None

  Specify the millicores of CPU to request for a stress test job. If none specified, will default to 1500mi. The limit will be set as 2x the request.

Return Value:

A Job object that provides an interface for monitoring the job in the backend.

Example:

# This example will likely not work for you because it requires permissions to a specific S3 bucket.
# This demonstrates how you might specify such a configuration.
config = {
  "run_name": "Titanic",
  "data_info": {
    "label_col": "Survived",
    "ref_path": "s3://rime-datasets/titanic/titanic_example.csv", "eval_path": "s3://rime-datasets/titanic/titanic_example.csv"
  },
  "model_info": {
    "path": "s3://rime-models/titanic_s3_test/titanic_example_model.py"
  }
}
# Run the job using the specified config and the default Docker image in the RIME backend.
# Store the results under project ID "foo"
# Use the RIME Managed Image "tensorflow115".
# This assumes you have already created the Managed Image and waited for it to be ready.
job = rime_client.start_stress_test(test_run_config=config, project_id="foo", rime_managed_image="tensorflow115")

create_managed_image()

This method allow you to create new managed Docker images to run RIME on. These managed Docker images are managed by the RIME backend and will automatically be upgraded when you update your version of RIME. Images take a few minutes to be built. This method returns an object that can be used to track the progress of the image building job. The new custom image is only available for use in a stress test once it has status READY.

Arguments

• name: str

  The name of the new managed image. This acts as the unique identifier of the managed image. The call will fail if an image with the specified name already exists.

• requirements: List[ManagedImage.PipRequirement]

  List of additional pip requirements to be installed on the managed image. A ManagedImage.PipRequirement can be created with the helper method Client.pip_requirement(name: str, version_specifier: Optional[str] = None). The first argument is the name of the library (e.g. "tensorflow" or "xgboost") and the second argument is a valid pip version specifier (e.g. ">=0.1.2" or "==1.0.2").

Return Value

A RIMEImageBuilder object that provides an interface for monitoring the job in the backend.

Example

requirements = [
   # Fix the version of `xgboost` to `1.0.2`.
   rime_client.pip_requirement("xgboost", "==1.0.2"),
   # We do not care about the installed version of `tensorflow`.
   rime_client.pip_requirement("tensorflow")
 ]

# Start a new image building job
builder_job = rime_client.create_managed_image("xgboost102_tensorflow", requirements)
# Wait until the job has finished and print out status information.
# Once this prints out the `READY` status, your image is available for use in stress tests.
builder_job.get_status(verbose=True, wait_until_finish=True)

list_managed_images()

This method allows you to query the backend for managed Docker images. This is where the true power of the managed images feature lies. You can search for images with specific pip libraries installed so that you do not have to create a new managed image every time you need to run a stress test.

Arguments

• pip_library_filters: Optional[List[ListImagesRequest.PipLibraryFilter]] = None

  A list of filters used to query the managed images stored in the backend. Query results will match the intersection of the filters. You may construct ListImagesRequest.PipLibraryFilter objects using the Client.pip_library_filter(name: str, fixed_version: Optional[str] = None) helper method.

• page_token: str = ""

  Page token used for paginating the API results. To get access to the next page of results, use the second value in the tuple returned by the previous call.

• page_size: int = 100

  Number of results to output. Default is 100 managed images to be returned.

Return Value

A Tuple[List[Dict], str] object. The first value in the tuple is a list of dictionary representations of Managed Images. The second value in the tuple is the next page token.

Example

# Filter for an image with catboost1.0.3 and tensorflow installed.
filters = [
  rime_client.pip_library_filter("catboost", "1.0.3"),
  rime_client.pip_library_filter("tensorflow"),
]

# Query for the images.
images, next_page_token = rime_client.list_managed_images(pip_library_filters=filters)

# List comprehension to get all the names of the images.
names = [x["name"] for x in images]

create_project()

Projects allow you to organize stress test runs as you see fit. A natural way to organize stress test runs is to create a project for each specific ML task, such as predicting whether a transaction is fradulent.

Arguments:

• name: str

  The name of the project. You can change this later in the UI.

• description: str

  A short blurb about the project.

Return Value:

A Project that describes the created project. Its project_id attribute can be used in start_stress_test() and list_stress_test_jobs().

Example:

project = rime_client.create_project(name='foo', description='bar')

list_stress_test_jobs()

Query the backend for a list of jobs filtered by status and project ID. This is a good way to recover Job objects. Note that this only returns jobs from the last two days, because the time-to-live of job objects in the backend is set at two days.

Arguments:

• status_filters: Optional[List[str]] = None

  Select jobs by a union of statuses. If this is omitted, jobs will not be filtered by status. Acceptable values are in the following array:

  ['UNKNOWN_JOB_STATUS', 'PENDING', 'RUNNING', 'FAILING', 'SUCCEEDED']
  

• project_id: Optional[str] = None

  Select jobs by project. If this is omitted, jobs from across different projects will be returned.

Return Value:

A list of Job objects. These are not guaranteed to be in any sorted order.

Example:

# Get all running and succeeded jobs for project 'foo'
jobs = rime_client.list_stress_test_jobs(status_filters=['RUNNING', 'SUCCEEDED'], project_id='foo')

get_firewall()

Query the backend for a Firewall which can be used to perform Firewall operations. If the FW you are trying to fetch does not exist, this will error.

Arguments:

• firewall_id: str

  ID of the FW instance to fetch.

Return Value:

A Firewall object.

Example:

# Get FW foo if it exists.
firewall = rime_client.get_firewall("foo")

get_firewall_for_project()

Query the backend for an active Firewall in a specified project which can be used to perform Firewall operations. If there is no active Firewall for the project, this call will error.

Arguments:

• project_id: str

  ID of the project which contains a Firewall.

Return Value:

A Firewall object.

Example:

# Get FW in foo-project if it exists.
firewall = rime_client.get_firewall_for_project("foo-project")

create_firewall()

Create a Firewall for a given project.

Arguments:

• name: str

  FW name.

• bin_size_seconds: int

  Bin size in seconds. Only supports daily or hourly.

• test_run_id: str

  ID of the stress test run that firewall will be based on.

• project_id: str

  ID of the project this FW belongs to.

Return Value:

A Firewall object.

Example:

# Create FW based on foo stress test in bar project.
firewall = rime_client.create_firewall(
  "firewall name", "day", "foo", "bar")

Job

This object provides an interface for monitoring the status of a stress test job in the RIME backend.

get_status()

Query the RIME backend for the job's status. This includes flags for blocking until the job is complete and printing information to stdout. This method can help with monitoring the progress of stress test jobs, because it prints out helpful information such as running time and the progress of the test run.

Arguments:

• verbose: bool = False

  Whether to print additional status information to stdout. If this flag is enabled and the job status is 'SUCCEEDED' or 'FAILING', the logs of the testing engine will be dumped to stdout to help with debuggability. Note that this logs have no strict form and will be subject to significant change in future versions.

• wait_until_finish: bool = False

  Whether to block until the job status is 'SUCCEEDED' or 'FAILING'. If verbose is enabled too, information about the job including running time and progress will be printed to stdout every poll_rate_sec.

• poll_rate_sec: float = 5.0

  How often to ping the RIME backend services for the status of the job. Units are in seconds.

Return Value:

A dictionary representing the status of the Job.

{
  "name": str
  "type": str
  "status": str
  "start_time_secs": int64
  "running_time_secs": double
}

type will be an element in the following array:

['MODEL_STRESS_TEST', 'UNKNOWN_JOB_TYPE']

status will be an element in the following array:

['UNKNOWN_JOB_STATUS', 'PENDING', 'RUNNING', 'FAILING', 'SUCCEEDED']

Example:

# Block until this job is finished and dump monitoring info to stdout.
job_status = job.get_status(verbose=True, wait_until_finish=True)

get_test_cases_result()

Retrieve all the test cases for a completed stress test run in a dataframe. This gives you the ability to perform granular queries on test cases. For example, if you only care about subset performance tests and want to see the results on each feature, you can fetch all the test cases in a dataframe, then query on that dataframe by test type. This only works on stress test jobs that have succeeded.

Note: this does not work on <0.14.0 RIME test runs.

Arguments:

• version: Optional[str] = None

  Semantic version of the results to be returned. This allows users to pin the version of the results, which is helpful if you write any code on top of RIME data. If you upgrade the SDK and do not pin the version in your code, it may break because the output not guaranteed to be stable across versions. The latest output will be returned by default.

Return Value:

A pandas.DataFrame object containing the test case results. Here is a selected list of columns in the output:

1. test_run_id: ID of the parent test run.
2. features: List of features that the test case ran on.
3. test_batch_type: Type of test that was run (e.g. Subset AUC, Must be Int, etc.).
4. status: Status of the test case (e.g. Pass, Fail, Skip, etc.).
5. severity: Metric that denotes the severity of the failure of the test.

Example:

# Wait until the job has finished, since this method only works on SUCCEEDED jobs.
job.get_status(verbose=True, wait_until_finish=True)
# Dump the test cases in dataframe `df`.
# Pin the version to RIME version 0.14.0.
df = job.get_test_cases_result(version="0.14.0")
# Print out the column names and types.
print(df.columns)

get_test_run_result()

Retrieve high level summary information for a complete stress test run in a single-row dataframe. By concatenating these rows together, this allows you to build a table of test run results for sake of comparison. This only works on stress test jobs that have succeeded.

Note: this does not work on <0.14.0 RIME test runs.

Arguments:

• version: Optional[str] = None

  Semantic version of the results to be returned. This allows users to pin the version of the results, which is helpful if you write any code on top of RIME data. If you upgrade the SDK and do not pin the version in your code, it may break because the output not guaranteed to be stable across versions. The latest output will be returned by default.

Return Value:

A pandas.DataFrame object containing the test run result. There are a lot of columns, so it is worth viewing them with the .columns method to see what they are. Generally, these columns have information about the model and datasets as well as summary statistics like the number of failing test cases or number of high severity test cases.

Example:

# Wait until the job has finished, since this method only works on SUCCEEDED jobs.
job.get_status(verbose=True, wait_until_finish=True)
# Dump the test cases in dataframe `df`.
# Pin the version to RIME version 0.14.0.
df = job.get_test_run_result(version="0.14.0")
# Print out the column names and types.
print(df.columns)

Project

This object describes a project in the RIME backend.

Attributes:

• project_id: str

  How to refer to the project in the backend. Use this attribute to specify the project for the backend in start_stress_test_job() and list_stress_test_jobs().

• name: str

• description: str

Firewall

Firewall object wrapper with helpful methods for working with RIME Firewall.

Attributes:

• backend: RIMEBackend

  The RIME backend used to query about the status of the job.

• firewall_id: str

  How to refer to the FW in the backend. Use this attribute to specify the Firewall for tasks in the backend.

update_firewall_stress_test_run()

Update firewall with stress test run id.

Arguments:

• stress_test_run_id: str

Stress Test Run Id to configure new firewall

Return Value:

• None

start_continuous_test()

Start a RIME model firewall test on the backend's ModelTesting service.

This allows you to run Firewall Test job on the RIME backend. This will run firewall on a batch of tabular data.

Arguments:

• test_run_config: dict

  Configuration for the test to be run, which specifies paths to the model and datasets to used for the test.

• custom_image: Optional[CustomImage]

  Specification of a customized container image to use running the model test. The image must have all dependencies required by your model. The image must specify a name for the image and optional a pull secret (of type CustomImage.PullSecret) with the name of the kubernetes pull secret used to access the given image.

• rime_managed_image: Optional[str]

  Name of a managed image to use when running the model test. The image must have all dependencies required by your model. To create new managed images with your desired dependencies, use the client's create_managed_image() method.

• ram_request_megabytes: int

  Megabytes of RAM requested for the stress test job. If none specified, will default to 4000MB. The limit is 2x the megabytes requested.

• cpu_request_millicores: int

  Millicores of CPU requested for the stress test job. If none specified, will default to 1500mi. The limit is 2x the millicores requested.

Return Value: A Job providing information about the model stress test job.

Example:

# This example will likely not work for you because it requires permissions
# to a specific S3 bucket. This demonstrates how you might specify such a
# configuration.
incremental_config = {
    "eval_path": "s3://rime-datasets/
       fraud_continuous_testing/eval_2021_04_30_to_2021_05_01.csv",
    "timestamp_col": "timestamp"
}
# Run the job using the specified config and the default Docker image in
# the RIME backend. Use the RIME Managed Image "tensorflow115".
# This assumes you have already created the Managed Image and waited for it
# to be ready.
firewall = rime_client.get_firewall("foo")
job =
    firewall.start_continuous_test(
        test_run_config=incremental_config,
        rime_managed_image="tensorflow115",
        ram_request_megabytes=8000,
        cpu_request_millicores=2000)

---

RIME CLI

Data Format Check

rime-data-format-check validates data for use with the RIME platform by performing a series of sanity checks, depending on the nature of the data (tabular vs. NLP).

Example (Tabular)

These examples were run using the Kaggle Titanic dataset.

1. Basic Usage (no labels or predictions)

rime-data-format-check -tabular \
  --task "Binary Classification" \
  --ref-path data/titanic/train.csv \
  --eval-path data/titanic/test.csv

Output (PASSING):

Inspecting 'data/titanic/train.csv'
Done!

Inspecting 'data/titanic/test.csv'
Done!

WARNING: No prediction column is provided. Although you can still run RIME without predictions, it will not be as powerful as if you run it WITH predictions.

WARNING: No label column is provided. Although you can still run RIME without labels, it will not be as powerful as if you run it WITH labels.


---


Your data should work with RIME!

Example (NLP)

These examples were run on a formatted version of the arXiv dataset, where input files contains data points with the following structure:

(in data/classification/arxiv/train.json)
[
  ...
  {
    "text": "On maxima and ladder processes for a dense class of Levy processes",
    "timestamp": "2007-05-23 00:00:00",
    "label": 4,
    "probabilities": [0.11, 0.0, 0.03, 0.01, 0.82, 0.0, 0.03, 0.0, 0.0, 0.0]
  },
  ...
]

1. Basic Usage (predictions included with inputs)

rime-data-format-check -nlp \
  --task "Text Classification" \
  --ref-path data/classification/arxiv/train.json \
  --eval-path data/classification/arxiv/val.json

Output (PASSING)

Inspecting '../../../rime-data-format-check/data/classification/arxiv/train.json':
100%|█████████████████████████████████████████████████████████████████████████████| 50000/50000 [00:01<00:00, 33219.50it/s]

Inspecting '../../../rime-data-format-check/data/classification/arxiv/val.json':
100%|███████████████████████████████████████████████████████████████████████████| 147087/147087 [00:04<00:00, 33319.64it/s]

---


Your data should work with RIME!

Appendix

Tabular Data Checks

• accepted file formats (.csv or .parquet)
• appropriate dataset split
  • (train and evaluation sets)
• presence and format of header column(s)
  • strings corresponding to feature names
• presence and format of label column
  • numeric values for regression tasks
  • [0,1] for binary classification tasks
  • integers for multi-class classification tasks
• (recommended) presence and format of prediction column
  • numeric values for regression tasks
  • float values between [0,1] for binary classification tasks
  • integers for multi-class classification tasks*
• (for RIME Continuous Testing only) presence and format of timestamp column
  • “YYYY-MM-DD” or “YYYY-MM-DD HH:TT:SS"

*Prediction values for multi-class classification must be provided in a separate dataset.

NLP Data Checks

Note: Extraneous keys in the entity dictionaries are permitted.

• accepted file formats (.json or .jsonl)
  • can be compressed via gzip
• appropriate dataset split
  • (train and evaluation sets)
• presence and structure of keys (based on task)
  • Text Classification
    • "text:": string
    • "label": int
    • "probabilities": List[float]
  • Named Entity Recognition
    • "text": string
    • "entities": List[dict]
      • entity dicts should have:
        • "type": string
        • "mentions": List[dict]
          • mention dicts should have:
            • "start_offset": int
            • "end_offset": int
    • "predicted_entities": List[dict]
      • entity dicts should have:
        • "type": string
        • "mentions": List[dict]
          • mention dicts should have:
            • "start_offset": int
            • "end_offset": int

More Examples (Tabular)

1. With Labels (and a sample failure)

rime-data-format-check -tabular \
  --task "Binary Classification" \
  --ref-path data/titanic/train.csv \
  --eval-path data/titanic/test.csv \
  --label-col-name "Survived"

Output (ERROR):

Inspecting 'data/titanic/train.csv'
Done!

Inspecting 'data/titanic/test.csv'

---

Error:

Label column (Survived) not found in data (data/titanic/test.csv). If a label column exists in one dataset, it MUST exist in the other.

In this case, it looks like we're missing the label column from our evaluation set. For the Titanic dataset, these values are provided in the gender_submission.csv.

Adding the label column to the evaluation set should resolve the issue:

import pandas as pd
df_test = pd.read_csv("titanic/test.csv")
df_labels = pd.read_csv("titanic/gender_submission.csv")
df_test_with_labels = pd.merge(df_test, df_labels, on=["PassengerId"])
df_test_with_labels.to_csv("titanic/test_with_labels.csv")

Using the updated evaluation set, the data checks should pass:

rime-data-format-check -tabular \
  --task "Binary Classification" \
  --ref-path data/titanic/train.csv \
  --eval-path data/titanic/test_with_labels.csv \
  --label-col-name "Survived"

Output (PASSING):

Inspecting 'data/titanic/train.csv'
Done!

Inspecting 'data/titanic/test_with_labels.csv'
Done!

WARNING: No prediction column is provided. Although you can still run RIME without predictions, it will not be as powerful as if you run it WITH predictions.


---


Your data should work with RIME!

The --pred-col-name flag operates identically to --label-col-name.

More Examples (NLP)

1. With Predictions Provided Separately (and a sample failure)

In this example, predictions are omitted from the inputs and provided in a separate file, data/classification/arxiv/train_preds.json.

(in data/classification/arxiv/train.json)
[
  ...
  {
    "text": "On maxima and ladder processes for a dense class of Levy processes"
    "timestamp": "2007-05-23 00:00:00",
    "label": 4
  },
  ...
]

(in data/classification/arxiv/train_preds.json)
[
  ...
  {
    "probabilities": [0.11, 0.0, 0.03, 0.01, 0.82, 0.0, 0.03, 0.0, 0.0, 0.0]
  },
  ...
]

rime-data-format-check -nlp \
  --task "Text Classification" \
  --ref-path data/classification/arxiv/train.json \
  --preds-ref-path data/classification/arxiv/train_preds.json \
  --eval-path data/classification/arxiv/val.json \
  --preds-eval-path data/classification/arxiv/val_preds.json

Output (ERROR):

Inspecting 'data/classification/arxiv/train_preds.json':
  0%|                                                                             | 35/50000 [00:00<00:02, 24377.39it/s]

---

Error:

File 'data/classification/arxiv/train_preds.json', Index 35:

Key 'probabilities' error:
Or(<class 'float'>) did not validate 24
24 should be instance of 'float'

---

Inputs for task 'Text Classification' must adhere to the following structure:

{'probabilities': [<class 'float'>]}

In this case, it looks like one of our prediction entities has an invalid value of 24, an int instead of the expected float.

The element at the specified file (data/classification/arxiv/train_preds.json) and index (35) has the following values:

{
  "probabilities": [24, 0.01, 0.03, 0.03, 0.68, 0.01, 0.0, 0.0, 0.0, 0.0]
}

It appears that the first value, 24, should actually be 0.24. Adusting the value resolves the issue:

Inspecting 'data/classification/arxiv/train_preds.json':
100%|█████████████████████████████████████████████████████████████████████████████|50000/50000 [00:01<00:00, 26954.06it/s]

Inspecting 'data/classification/arxiv/val_preds.json':
100%|███████████████████████████████████████████████████████████████████████████|147087/147087 [00:05<00:00, 27215.18it/s]

Inspecting 'data/classification/arxiv/train.json':
100%|█████████████████████████████████████████████████████████████████████████████|50000/50000 [00:01<00:00, 32684.76it/s]

Inspecting 'data/classification/arxiv/val.json':
100%|███████████████████████████████████████████████████████████████████████████|147087/147087 [00:04<00:00, 32876.84it/s]

---


Your data should work with RIME!