MOSTLY AI Cheat Sheet¶
Initialization¶
from mostlyai.sdk import MostlyAI
# local mode (with TCP port)
mostly = MostlyAI(
local=True,
local_dir='~/mostlyai',
local_port=8080,
)
# client mode
mostly = MostlyAI(
base_url='https://app.mostly.ai', # or set env var `MOSTLY_BASE_URL`
api_key='INSERT_YOUR_API_KEY', # or set env var `MOSTLY_API_KEY`
)
Generators¶
# shorthand syntax to train a new single-table generator
g = mostly.train(name: str, data: pd.DataFrame)
# train a new generator
g = mostly.train(config: dict | GeneratorConfig, start: bool, wait: bool)
# configure a new generator and launch separately
g = mostly.generators.create(config: dict | GeneratorConfig)
g.training.start()
g.training.wait()
g.training.progress() # includes loss messages
g.training.logs()
# download a ZIP of quality assurance reports
g.reports()
# iterate over all your available generators
for g in mostly.generators.list():
print(g.id, g.name)
# fetch a generator by id
g = mostly.generators.get(id: str)
# fetch a generator's configuration
config = g.config()
# open a generator in a new browser tab
g.open()
# update a generator
g.update(name: str, ...)
# delete a generator
g.delete()
# export a generator as a ZIP file
fn = g.export_to_file()
# import a generator from a ZIP file
g = mostly.generators.import_from_file(fn)
# continue training of an existing generator (with connectors)
g2 = g.clone("continue")
g2.training.start()
g2.training.wait()
Synthetic Datasets¶
# shorthand syntax for generating a new synthetic dataset
sd = mostly.generate(g, size: int)
# shorthand syntax for conditionally generating a new synthetic dataset
sd = mostly.generate(g, seed: pd.DataFrame)
# generate a new synthetic dataset
sd = mostly.generate(g, config: dict | SyntheticDatasetConfig, start: bool, wait: bool)
# configure a new synthetic dataset and launch separately
sd = mostly.synthetic_datasets.create(config: dict | SyntheticDatasetConfig)
sd.generation.start()
sd.generation.wait()
sd.generation.progress()
sd.generation.logs()
# download a ZIP of quality assurance reports
sd.reports()
# iterate over all your available synthetic datasets
for sd in mostly.synthetic_datasets.list():
print(sd.id, sd.name)
# fetch a synthetic dataset by id
sd = mostly.synthetic_datasets.get(id: str)
# fetch a synthetic dataset's configuration
config = sd.config()
# open a synthetic dataset in a new browser tab
sd.open()
# download a synthetic dataset
sd.download(file: str, format: str)
# fetch the synthetic dataset's data
syn_df = sd.data()
# update a synthetic dataset
sd.update(name: str, ...)
# delete a synthetic dataset
sd.delete()
Synthetic probes allow to instantly generate synthetic samples on demand, without storing these on the platform. This feature depends on the availability of Live Probing on the platform. The syntax is similar to generating a synthetic dataset, with the notable difference that its return value is already the synthetic data as pandas DataFrame(s).
# shorthand syntax for probing for synthetic samples
syn_df = mostly.probe(g, size: int)
# shorthand syntax for conditionally probing for synthetic samples
syn_df = mostly.probe(g, seed: pd.DataFrame)
# probe for synthetic samples
syn_df = mostly.probe(g, config: dict | SyntheticDatasetConfig)
Connectors¶
Connectors can be used both as a source of original data for training a generator, as well as a destination for delivering the generated synthetic data samples to. See ConnectorConfig for the full list of available connectors, and their corresponding configuration parameters.
# create a new connector
c = mostly.connect(config: dict | ConnectorConfig)
# fetch a connector by id
c = mostly.connectors.get(id: str)
# list all locations of a connector
c.locations(prefix: str)
# fetch schema for a specific location
c.schema(location: str)
# iterate over all your available connectors
for c in mostly.connectors.list():
print(c.id, c.name)
# update a connector
c.update(name: str, ...)
# open a connector in a new browser tab
c.open()
# delete a connector
c.delete()
Datasets¶
Datasets can be used to train generators or create artifacts and may be created with or without a corresponding connector. Datasets are only available in client mode.
# create a new dataset
ds = mostly.datasets.create(config: dict | DatasetConfig)
# retrieve a dataset by id
ds = mostly.datasets.get(dataset_id: str)
# list datasets (iterator with optional filters)
for ds in mostly.datasets.list(
offset: int = 0,
limit: int | None = None,
search_term: str | None = None,
owner_id: str | list[str] | None = None,
visibility: str | list[str] | None = None, # e.g., PUBLIC, PRIVATE, UNLISTED
created_from: str | None = None, # YYYY-MM-DD
created_to: str | None = None, # YYYY-MM-DD
sort_by: str | list[str] | None = None, # NO_OF_THREADS | NO_OF_LIKES | RECENCY
):
print(ds.id, ds.name)
# update an existing dataset (partial patch)
ds = mostly.datasets._update(dataset_id: str, config: dict | DatasetPatchConfig)
# delete a dataset
mostly.datasets._delete(dataset_id: str)
# download a file from a dataset
content_bytes, filename = mostly.datasets._download_file(
dataset_id: str,
file_path: str, # path inside the dataset
)
# upload a file to a dataset
mostly.datasets._upload_file(
dataset_id: str,
file_path: str | Path, # local path to file
)
# delete a file from a dataset
mostly.datasets._delete_file(
dataset_id: str,
file_path: str | Path, # path inside the dataset
)