How to use mml interactively

MML provides the mml.interactive module with the mml.interactive.planning and mml.interactive.loading submodules to support direct usage within jupyter notebooks. This notebook itself demonstrates the usage.

from pathlib import Path

import mml.interactive

# your mml.env location might differ
mml.interactive.init(env_path=Path(mml.__file__).parent / "mml.env")

data exploration

# this example shows how to retrieve all installed tasks
with mml.interactive.default_file_manager() as fm:
    print(sorted(list(fm.task_index.keys())))

# this example loads task information
tasks = ["caltech256_object_classification", "cifar10_object_classification", "svhn"]
mml.interactive.get_task_infos(tasks)

experiment planning

# this example creates some MML calls to run experiments
base_reqs = mml.interactive.DefaultRequirements()
all_cmds = list()
# create task data
all_cmds.append(
    mml.interactive.MMLJobDescription(
        prefix_req=base_reqs, mode="create", config_options={"task_list": tasks, "proj": "my_exp"}
    )
)
# preprocess tasks
all_cmds.append(
    mml.interactive.MMLJobDescription(
        prefix_req=base_reqs, mode="pp", config_options={"task_list": tasks, "proj": "my_exp"}
    )
)
# create tagged task variants
all_cmds.append(
    mml.interactive.MMLJobDescription(
        prefix_req=base_reqs,
        mode="info",
        config_options={"task_list": tasks, "proj": "my_exp", "tagging.all": "+subset?0_05+confuse?0_1"},
    )
)
# you may also loop over
for t in tasks:
    all_cmds.append(
        mml.interactive.MMLJobDescription(
            prefix_req=base_reqs,
            mode="train",
            config_options={"pivot.name": f"{t}+subset?0_05+confuse?0_1", "proj": "my_exp", "sampling.batch_size": 500},
        )
    )
# either render jobs directly
for cmd in all_cmds:
    print(cmd.render())
# or put them into a file
mml.interactive.write_out_commands(cmd_list=all_cmds, name="my_exp")

start planned jobs from within a notebook

runner = mml.interactive.EmbeddedJobRunner()
job = mml.interactive.MMLJobDescription(
    prefix_req=base_reqs,
    mode="info",
    config_options={"task_list": tasks, "proj": "my_exp", "tagging.all": "+subset?0_05+confuse?0_1"},
)
runner.run(job=job)

experiment evaluation

# load experiment models
models = mml.interactive.load_project_models("test")

# inspect and evaluate
print(models["lapgyn4_instrument_count_miccai"][0].training_time)
print(models["lapgyn4_instrument_count_miccai"][0].performance)
print(models["lapgyn4_instrument_count_miccai"][0].metrics[-2])