FEDn is an open-source, modular and ML-framework agnostic framework for Federated Machine Learning

Status

• [x] Ready
• [ ] Draft
• [ ] Hold

Description

Included in this pr:

• A token is generated (encoded) by json web tokens (jwt) based on the payload (including expire date and current date)
• Expire date is set to 90 days
• Adding the boolean flag --token will generate the token, by default false
• If a token has been generated, combiner and clients must send the token via header in the request
• If no token is used, the reducer will ignore the header.
• The encryption alg. for encoding and decoding is set to HS256
• obs the cli/tests does not currently work due to wrong reference import (no parent package, cli is not included in fedn module)
• Obs. merge with develop needed because of PR #341

Types of changes

What types of changes does your code introduce to FEDn?

• [ ] Hotfix (fixing a critical bug in production)
• [ ] Bugfix
• [x] New feature
• [ ] Documentation update

Checklist

If you're unsure about any of the items below, don't hesitate to ask. We're here to help! This is simply a reminder of what we are going to look for before merging your code.

• [x ] This pull request is against develop branch (not applicable for hotfixes)
• [ ] I have included a link to the issue on GitHub or JIRA (if any)
• [ ] I have included migration files (if there are changes to the model classes)
• [ ] I have read the CONTRIBUTING doc
• [ ] I have included tests to complement my changes
• [ ] I have updated the related documentation (if necessary)
• [ ] I have added a reviewer for this pull request
• [ ] I have added myself as an author for this pull request

Further comments

Anything else you think we should know before merging your code!

Status

• [ ] Ready
• [x] Draft
• [ ] Hold

Description

The Sphinx autodoc system is in, along with some PEP-8 fixes and setting docstrings in the right format. Please note that a build scheme to compile and deploy the API docs to a webpage will need to be setup in due course.

Types of changes

What types of changes does your code introduce to FEDn?

• [ ] Hotfix (fixing a critical bug in production)
• [ ] Bugfix
• [ ] New feature
• [x] Documentation update

Checklist

If you're unsure about any of the items below, don't hesitate to ask. We're here to help! This is simply a reminder of what we are going to look for before merging your code.

• [x] This pull request is against develop branch (not applicable for hotfixes)
• [ ] I have included a link to the issue on GitHub or JIRA (if any)
• [ ] I have included migration files (if there are changes to the model classes)
• [ ] I have read the CONTRIBUTING doc
• [ ] I have included tests to complement my changes
• [x] I have updated the related documentation (if necessary)
• [x] I have added a reviewer for this pull request
• [x] I have added myself as an author for this pull request

Further comments

Anything else you think we should know before merging your code!

I am trying to follow the example given in the readme, but get an error when I come to the section "Attach two Clients to the FEDn network". I get the following message when the keras-clients are starting client_1 | Traceback (most recent call last): client_1 | File "train.py", line 54, in client_1 | from fedn.utils.kerassequential import KerasSequentialHelper client_1 | ModuleNotFoundError: No module named 'fedn.utils.kerassequential' client_2 | Traceback (most recent call last): client_2 | File "train.py", line 54, in client_2 | from fedn.utils.kerassequential import KerasSequentialHelper client_2 | ModuleNotFoundError: No module named 'fedn.utils.kerassequential' Curret directory is "test/mnist-keras". As far as I can see there is no KerasSequentialHelper, but a KerasHelper that should be called. I use Linux Mint 18.3.

Is your feature request related to a problem? Please describe.

In Tensorflow 2.6 there is no longer a "predict_classes", so validate.py in test/mnist-keras is not working.

Describe the solution you'd like An overhaul of the client code to work with latest TF.

How would the solution positively affect the functionality?

Describe any drawbacks (if any) A clear and concise description of the negative outcomes of the suggested solution.

Contact Details An e-mail address in case we need to contact you for further details regarding this request.

To streamline better test projects In this suggested commit there are a new structure to allow for overriding of parameters.

version: '3.7'
services:
  client:
    build:
      context: .
      dockerfile: components/client/Dockerfile
      args:
        baseimage: "tensorflow/tensorflow:latest"

Illustrated by an example

docker-compose -f docker-compose.yaml -f reducer.yaml -f combiner.yaml -f client.yaml **-f tensorflow.yaml** --build --scale client=5

where tensorflow.yaml in this example is overriding the BASEIMAGE variable to the Dockerimage and instructs which base image to initiate the client image from.

In this sense the tests we construct can have the same basic structure but depending on framework selected can inherit base image from for example tensorflow to ensure the execution context is present.

Severity

• [] Critical/Blocker (select if the issue makes the application unusable or causes serious data loss)
• [x] High (select if the issue affects a major feature and there is no workaround or the available workaround is very complex)
• [ ] Medium (select if the issue affects a minor feature or affects a major feature but has an easy enough workaround to not cause any major inconvenience)
• [ ] Low (select if the issue doesn't significantly affect the user experience, like minor visual bugs)

Describe the bug The pre-built compute package for mnist-keras in current master does not match the actual client code (has not been updated in the release)

Environment:

• OS: [e.g. mac OS]
• Version: [e.g. macOS Catalina]
• Browser [e.g. chrome, safari]
• Version [e.g. 22]

Reproduction Steps Steps to reproduce the behavior:

1. Go to '...'
2. Click on '....'
3. Scroll down to '....'
4. See error

Expected behavior A clear and concise description of what you expected to happen.

Screenshots If applicable, add screenshots to help explain your problem.

Contact Details An e-mail address in case we need to contact you for further details regarding this issue.

Severity Low (select if the issue doesn't significantly affect the user experience, like minor visual bugs)

Describe the bug Got the following warnings when start the minio service. Version 0.2.4

minio_1 | WARNING: MINIO_ACCESS_KEY and MINIO_SECRET_KEY are deprecated. minio_1 | Please use MINIO_ROOT_USER and MINIO_ROOT_PASSWORD minio_1 | API: http://172.19.0.4:9000 http://127.0.0.1:9000 minio_1 | minio_1 | Console: http://172.19.0.4:46547 http://127.0.0.1:46547 minio_1 | minio_1 | Documentation: https://docs.min.io minio_1 | minio_1 | WARNING: Console endpoint is listening on a dynamic port (46547), please use --console-address ":PORT" to choose a static port.

Is your feature request related to a problem? Please describe. After starting the reducer and combiners we will get this error after starting clients:

client_1 | Asking for assignment client_1 | 09/25/2021 11:50:08 AM [connectionpool.py:971] Starting new HTTPS connection (1): 130.238.29.53:8090 client_1 | 2021-09-25 11:50:08,297 - urllib3.connectionpool - DEBUG - Starting new HTTPS connection (1): 130.238.29.53:8090 client_1 | 09/25/2021 11:50:08 AM [connectionpool.py:452] https://130.238.29.53:8090 "GET /assign?name=client00ba921f HTTP/1.1" 200 24

This will confuse the user, however, it's just a package problem, one needs to upload package before starting clients So, the suggestion is to have a clear message on the client log that asks the user to upload the package.

Describe the solution you'd like A clear and concise description of what you want to happen.

How would the solution positively affect the functionality? A clear and concise description of the positive outcomes of the suggested solution.

Describe any drawbacks (if any) A clear and concise description of the negative outcomes of the suggested solution.

Contact Details An e-mail address in case we need to contact you for further details regarding this request.

Is your feature request related to a problem? Please describe. Why not adding a Timestamp to the combiner log like reducer, then at least it will help to track errors for long training

Describe the solution you'd like A clear and concise description of what you want to happen. Instead of having this: combiner_1 | COMBINER(FEDn_Combiner_addi):0 COMBINER: waiting for model updates: 0 of 4 completed.

it's better to have this: combiner_1 | COMBINER(FEDn_Combiner_addi): 09/24/2021 09:38:58 AM 0 COMBINER: waiting for model updates: 0 of 4 completed.

Describe any drawbacks (if any) None

Contact Details An e-mail address in case we need to contact you for further details regarding this request.

Is your feature request related to a problem? Please describe.

Is is often helpful to have a graphical representation of the FEDn network. We have a simple version now in the /network view, but this should be improved to better handle multiple combiners. Also, it would be nice if the actual client names/ids were shown, as well as their status (online/offline).

Severity

• [x ] Critical/Blocker (select if the issue makes the application unusable or causes serious data loss)
• [ ] High (select if the issue affects a major feature and there is no workaround or the available workaround is very complex)
• [ ] Medium (select if the issue affects a minor feature or affects a major feature but has an easy enough workaround to not cause any major inconvenience)
• [ ] Low (select if the issue doesn't significantly affect the user experience, like minor visual bugs)

Describe the bug When I try to upload the mnist-keras or mnist-pytorch models at 'https://localhost:8090/control' , it gives a http 500 server error. The log of the reducer contains the following error message:

model = self.load_model(path) reducer_1 | File "/app/fedn/fedn/utils/pytorchhelper.py", line 31, in load_model reducer_1 | b = np.load(path) reducer_1 | File "/usr/local/lib/python3.8/site-packages/numpy/lib/npyio.py", line 457, in load reducer_1 | raise ValueError("Cannot load file containing pickled data " reducer_1 | ValueError: Cannot load file containing pickled data when allow_pickle=False

It seems that numpy tries to load a pickled file with the allow_pickle=False option. This is for PyTorch, but the same error comes with the Keras model as well.

Environment: Ubuntu 18.04

Contact Details [email protected]

Severity

• [ ] Critical/Blocker (select if the issue makes the application unusable or causes serious data loss)
• [ ] High (select if the issue affects a major feature and there is no workaround or the available workaround is very complex)
• [x] Medium (select if the issue affects a minor feature or affects a major feature but has an easy enough workaround to not cause any major inconvenience)
• [ ] Low (select if the issue doesn't significantly affect the user experience, like minor visual bugs)

Describe the bug After moving the validation metric plot to the /models view the dropdown where a user can choose which validation metric to plot is broken.

Is your feature request related to a problem? Please describe. Since we are now basing the quickstart tutorial on the pytorch version of the MINST example, it would reduce the risk for mistakes if we made the pytorch helper the default in the UI.

Is your feature request related to a problem? Please describe. User understanding of how FEDn works might be obscured by the level of automation in the current examples that relies on Docker and docker-compose.

Describe the solution you'd like To complement the existing, developer-oriented examples, we should add a step-by-step guide that only uses venv.

How would the solution positively affect the functionality? Easier to grasp what each services is doing.

Is your feature request related to a problem? Please describe.

We have been going back and forth on this one, trading off the ability to easily automate launching an arbitrary number of clients for the minst quickstart examples. Right now it is automated, but this comes at the price that we are using docker library in the entrypoint, and a relatively complex setup with docker-compose (overriding a base file). To strike a better balance here we will make the number of clients more static in the default configuration, with examples / docs for how to automate scaling of the clients for testing purposes.

Describe the solution you'd like When provide a docker-compose file that "manually" creates 3 clients (each with different, human readable name), and the default data-partition should be 3.

How would the solution positively affect the functionality? It will make the different steps involved more transparent and we can remove the docker library dependency in the entrypoint.

Describe any drawbacks (if any) It will create more work to automate launching arbitrary number of clients.

Severity

• [ ] Critical/Blocker (select if the issue makes the application unusable or causes serious data loss)
• [ ] High (select if the issue affects a major feature and there is no workaround or the available workaround is very complex)
• [ ] Medium (select if the issue affects a minor feature or affects a major feature but has an easy enough workaround to not cause any major inconvenience)
• [x] Low (select if the issue doesn't significantly affect the user experience, like minor visual bugs)

Describe the bug The graph is not shown, due to this error:

fedn-reducer-1 | ValueError: failed to validate StaticLayoutProvider(id='p1074', ...).graph_layout: expected an element of Dict(Int, Seq(Any)), got {'combiner': array([-0.0008935, 0.0017451]), 'reducer': array([ 0.45759661, -0.88965251]), '637d40c4e93300a65d6c1bb7': array([0.54329689, 0.84086893]), '637d40c4e93300a65d6c1bb9': array([-1. , 0.04703848])}

Environment:

• OS: [e.g. mac OS]
• Version: [e.g. macOS Catalina]
• Browser [e.g. chrome, safari]
• Version [e.g. 22]