Iris commands

An overview of the iris API.

These docs are outdated! Please check out https://docs.titanml.co for the latest information on the TitanML platform. If there's anything that's not covered there, please contact us on our discord.

Also see the API

Contents

• iris distil

• iris get

• iris status

• iris upload

• iris download

• iris pull

• iris makesafe

By default, these commands will send a request to the backend, then print a table with the response to the command line. If you would like to receive the response as a JSON object, simply add the tag --json after the command

iris distil

Dispatches knowledge distillation jobs to the TitanML platform. You can look at all of the iris distil arguments by using this command: iris distil --help.

iris finetune

Dispatches fine-tuning jobs to the TitanML platform.

For the above two commands, both model and dataset can be any of:

• A HuggingFace model/dataset

• A UUID generated by iris upload

• The filepath of a model/dataset folder (which will then be uploaded as part of the job).

More explanation here, here on how to launch a distillation or fine-tuning job.

The following arguments apply to both iris distil and iris finetune:

Command	Shortcut	Type	Description
--model	-m	TEXT	The model to optimize. [default: None] [required]
--dataset	-d	TEXT	The dataset to optimize the model with. [default: None] [required]
--task	-t	[sequence_classification | question_answering | token_classification]	The task to optimize the model for.[default: None] [required]
--subset	-ss	TEXT	The subset of the dataset to use
--name	-n	TEXT	The name to use for this job. Visible in the titan web interface.
--file	-f	PATH	The .yaml file containing experiment parameters.
--short-run	-s	flag	Truncates the run after 1 batch and 1 epoch. Will provide poor results, but useful to check that the model and dataset choices are valid.
--num-labels	-nl	INTEGER	Number of labels. Required for task sequence_classification [default: None]
--text-fields	-tf	TEXT	Text fields. Required for task sequence_classification [default: None]
--has-negative	-hn	flag	Has negative. Required for question_answering
--label-name	-ln	int:TEXT	The label names used for token classification
--help		flag	Show this message and exit.

You can use these additional arguments to specify hyperparameters when running a fine-tuning job:

Command	Shortcut	Type	Description
--batch-size	-bs	INTEGER	The batch size to use for training. Default is 16.
--learning-rate	-lr	FLOAT	The learning rate to use for training. Default is 2e-5.
--num-epochs	-ne	INTEGER	Number of epochs to finetune for. Default is 1.

iris get

Gets objects from the TitanAPI. You can look at all of the Iris get commands by using this command: iris get --help

iris status

Retrieves the status of an experiment. You can look at all of the Iris status commands by using this command: iris status --help

iris upload

Uploads local models and datasets by local filepath. You can look at all of the Iris upload commands by using iris upload --help.

iris delete

Deletes a model or dataset (i.e. an artefact), or an entire experiment, from the TitanML store.

iris download

Downloads a finished, Titan-optimised model directly onto your machine in the ONNX format.

iris pull

Downloads a finished, Titan-optimised model as a Docker image (usage is equivalent to docker pull).

iris infer

Runs inference by requesting the Triton Inference Server.

iris makesafe

Converts a model to the safetensors format, including those models which the transformers save method can't convert.