Getting started / Lab Network
Your first shared study,
one clear step at a time.
Bring the people you already work with and a question you can test together. Start with the shared workspace, then connect local learning when your labs are ready.
01 / A useful starting point
Start with one shared question.
Open Studies and create a study with a recognizable name, its purpose and a written protocol. Say what you will compare, how you will review observations, and what evidence would answer the question.
The study owner sponsors the workspace through Lab Network. You can join an existing study with a free account.
Make the first step small.
You can use the protocol, discussion and reviewed summaries before you connect any learning clients. Decide together when the study is ready for training.
02 / Your collaborators
Bring in the right people.
In the study’s People section, create an invitation and share its single-use code with your collaborator. They sign in, choose their display name and submit the code to join. Invitations expire after seven days.
- Owner: maintains the protocol, manages invitations and membership, and starts or stops learning.
- Contributor: publishes their own reviewed summaries and takes part in the discussion.
- Viewer: reads the study and downloads its exports.
A study includes up to 10 members, including the owner. A learning run can connect up to 5 laboratory clients. Training and held-out participation are separate from these workspace roles.
Each study holds up to 500 published summaries and 1,000 discussion messages at one time. Browse older pages to review or withdraw your records. These are retained-content capacities, not monthly allowances; downloading an export does not free space. Export and then deliberately remove records, or archive a completed study and start a new one within your active-study allowance.
04 / Connected research
Connect the data where it lives.
Open Learning in your study. Agree which labs will train and which will provide held-out evaluation. A run requires at least two distinct training accounts and one separate held-out account.
Prepare the annotated images locally using the research kit. Each data owner must explicitly approve the intended use and sharing, with an expiry date. Merely joining the study or uploading photos to DishFlow does not give this approval.
Stays on the lab’s computer
Training images, individual image features, local file paths and per-image ground truth.
Sent to the coordinator
Enrollment commitments, model updates, sample counts, local baselines and aggregate evaluation results.
Install the kit and connect your client
Use Python 3.11 or newer. Download and extract the research kit, then open a terminal in its extracted folder. Install the same pinned dependencies at every lab:
python -m pip install -r training/federated/requirements.txtFollow the kit’s preparation instructions to create private/prepared/local.json from your manifest, images and consent. Copy the Study split salt from the Learning page; every connected lab uses that exact value. The preparation tool keeps related dishes and acquisition sessions together.
In the study’s Learning page, choose your local site and participation role, then download the one-time connection file. Keep it private and save it as assaypilot-connection.json in the kit folder. Run:
python -m training.federated_network run \
--token-file assaypilot-connection.json \
--local private/prepared/local.json \
--work-dir private/connected-studyThe connection file supplies the server and credential. The client uses HTTPS and waits for the owner to freeze the plan. Keep the terminal and computer running for your lab’s required contributions. Use the same files and command to reconnect after a connection interruption.
Download the research kitReview the sharing and permission boundary
The coordinator can read submitted updates. Authorized study participants can access the shared plan commitments, models and aggregate evaluations. The service operator can access hosted records.
Updates and commitments may reveal information about the underlying data. This release uses ordinary model averaging; it does not provide secure aggregation or differential privacy. Each lab is responsible for its dataset rights and the information it shares.
The kit’s execution minimums are technical prerequisites, not a recommended sample size or evidence of independent laboratory validation. Agree your experimental design and data handling before sending real updates.
Freeze together, then let the clients work.
The owner reviews the enrolled sites and freezes the roster, settings and number of rounds. Connected clients then train, exchange updates, fit local baselines and evaluate the final candidate. The study waits for every required training participant in each round.
Each run supports up to 20 rounds. The owner’s subscription includes 100 completed aggregation rounds per UTC month across their studies; a pending round waits when that allowance is reached.
05 / Evidence you can inspect
Review what the study found.
Read the final candidate’s results alongside the local-only baselines. Check every site’s sample counts, count error, countability results and coverage. Give the entirely held-out lab particular attention: its images were not used for training.
Download the study record, research report and candidate. The candidate can make local research predictions through the client’s prediction command. It remains a research candidate marked HOLD and does not replace the active DishFlow detector.
If a study pauses or participation changes
A disconnected required client leaves the study waiting for its contribution. A monthly round limit or expired sponsor subscription pauses new learning. The Learning page shows the current waiting reason.
Leaving a study, being removed or deleting your account removes your shared summaries and discussion. Hosted learning artifact contents are also removed for every frozen run involving your lab, including completed runs; only stop metadata and round usage remain. Copies already downloaded by members cannot be recalled.
Archiving a study stops active execution and keeps its history. Consent expiry stops the affected frozen run. Use the study’s withdrawal controls when your lab’s permission changes.
If the owner’s subscription ends, existing study content and exports remain readable. New content, invitations and learning wait for Lab Network access to be restored.
Start with the people and the question.
Your first study can begin with a protocol and a conversation.