Using custom mods π§ͺ#
π§ͺ = This example covers experimental features that might change in future versions of Flower Please consult the regular PyTorch code examples (quickstart, advanced) to learn how to use Flower with PyTorch.
The following steps describe how to write custom Flower Mods and use them in a simple example.
Writing custom Flower Mods#
Flower Mods basics#
As described here, Flower Mods in their simplest form can be described as:
def basic_mod(msg: Message, context: Context, app: ClientApp) -> Message:
# Do something with incoming Message (or Context)
# before passing to the inner ``ClientApp``
reply = app(msg, context)
# Do something with outgoing Message (or Context)
# before returning
return reply
and used when defining the ClientApp
:
app = fl.client.ClientApp(
client_fn=client_fn,
mods=[basic_mod],
)
Note that in this specific case, this mod wonβt modify anything, and perform FL as usual.
WandB Flower Mod#
If we want to write a mod to monitor our client-side training using Weights & Biases, we can follow the steps below.
First, we need to initialize our W&B project with the correct parameters:
wandb.init(
project=...,
group=...,
name=...,
id=...,
resume="allow",
reinit=True,
)
In our case, the group should be the run_id
, specific to a ServerApp
run, and the name
should be the node_id
. This will make it easy to navigate our W&B project, as for each run we will be able to see the computed results as a whole or for each individual client.
The id
needs to be unique, so it will be a combination of run_id
and node_id
.
In the end we have:
def wandb_mod(msg: Message, context: Context, app: ClientAppCallable) -> Message:
run_id = msg.metadata.run_id
group_name = f"Run ID: {run_id}"
node_id = str(msg.metadata.dst_node_id)
run_name = f"Node ID: {node_id}"
wandb.init(
project="Mod Name",
group=group_name,
name=run_name,
id=f"{run_id}_{node_id}",
resume="allow",
reinit=True,
)
Now, before the message is processed by the server, we will store the starting time and the round number, in order to compute the time it took the client to perform its fit step.
server_round = int(msg.metadata.group_id)
start_time = time.time()
And then, we can send the message to the client:
reply = app(msg, context)
And now, with the message we got back, we can gather our metrics:
if reply.metadata.message_type == MessageType.TRAIN and reply.has_content():
time_diff = time.time() - start_time
metrics = reply.content.configs_records
results_to_log = dict(metrics.get("fitres.metrics", ConfigsRecord()))
results_to_log["fit_time"] = time_diff
Note that we store our metrics in the results_to_log
variable and that we only initialize this variable when our client is sending back fit results (with content in it).
Finally, we can send our results to W&B using:
wandb.log(results_to_log, step=int(server_round), commit=True)
The complete mod becomes:
def wandb_mod(msg: Message, context: Context, app: ClientAppCallable) -> Message:
server_round = int(msg.metadata.group_id)
if reply.metadata.message_type == MessageType.TRAIN and server_round == 1:
run_id = msg.metadata.run_id
group_name = f"Run ID: {run_id}"
node_id = str(msg.metadata.dst_node_id)
run_name = f"Node ID: {node_id}"
wandb.init(
project="Mod Name",
group=group_name,
name=run_name,
id=f"{run_id}_{node_id}",
resume="allow",
reinit=True,
)
start_time = time.time()
reply = app(msg, context)
if reply.metadata.message_type == MessageType.TRAIN and reply.has_content():
time_diff = time.time() - start_time
metrics = reply.content.configs_records
results_to_log = dict(metrics.get("fitres.metrics", ConfigsRecord()))
results_to_log["fit_time"] = time_diff
wandb.log(results_to_log, step=int(server_round), commit=True)
return reply
And it can be used like:
app = fl.client.ClientApp(
client_fn=client_fn,
mods=[wandb_mod],
)
If we want to pass an argument to our mod, we can use a wrapper function:
def get_wandb_mod(name: str) -> Mod:
def wandb_mod(msg: Message, context: Context, app: ClientAppCallable) -> Message:
server_round = int(msg.metadata.group_id)
run_id = msg.metadata.run_id
group_name = f"Run ID: {run_id}"
node_id = str(msg.metadata.dst_node_id)
run_name = f"Node ID: {node_id}"
wandb.init(
project=name,
group=group_name,
name=run_name,
id=f"{run_id}_{node_id}",
resume="allow",
reinit=True,
)
start_time = time.time()
reply = app(msg, context)
if reply.metadata.message_type == MessageType.TRAIN and reply.has_content():
time_diff = time.time() - start_time
metrics = reply.content.configs_records
results_to_log = dict(metrics.get("fitres.metrics", ConfigsRecord()))
results_to_log["fit_time"] = time_diff
wandb.log(results_to_log, step=int(server_round), commit=True)
return reply
return wandb_mod
And use it like:
app = fl.client.ClientApp(
client_fn=client_fn,
mods=[
get_wandb_mod("Custom mods example"),
],
)
TensorBoard Flower Mod#
The TensorBoard Mod will only differ in the initialization and how the data is sent to TensorBoard:
def get_tensorboard_mod(logdir) -> Mod:
os.makedirs(logdir, exist_ok=True)
def tensorboard_mod(
msg: Message, context: Context, app: ClientAppCallable
) -> Message:
logdir_run = os.path.join(logdir, str(msg.metadata.run_id))
node_id = str(msg.metadata.dst_node_id)
server_round = int(msg.metadata.group_id)
start_time = time.time()
reply = app(msg, context)
time_diff = time.time() - start_time
if reply.metadata.message_type == MessageType.TRAIN and reply.has_content():
writer = tf.summary.create_file_writer(os.path.join(logdir_run, node_id))
metrics = dict(
reply.content.configs_records.get("fitres.metrics", ConfigsRecord())
)
with writer.as_default(step=server_round):
tf.summary.scalar(f"fit_time", time_diff, step=server_round)
for metric in metrics:
tf.summary.scalar(
f"{metric}",
metrics[metric],
step=server_round,
)
writer.flush()
return reply
return tensorboard_mod
For the initialization, TensorBoard uses a custom directory path, which can, in this case, be passed as an argument to the wrapper function.
It can be used in the following way:
app = fl.client.ClientApp(
client_fn=client_fn,
mods=[get_tensorboard_mod(".runs_history/")],
)
Running the example#
Preconditions#
Letβs assume the following project structure:
$ tree .
.
βββ client.py # <-- contains `ClientApp`
βββ server.py # <-- contains `ServerApp`
βββ task.py # <-- task-specific code (model, data)
βββ requirements.txt # <-- dependencies
Install dependencies#
pip install -r requirements.txt
For W&B you will also need a valid account.
Start the long-running Flower server (SuperLink)#
flower-superlink --insecure
Start the long-running Flower client (SuperNode)#
In a new terminal window, start the first long-running Flower client using:
flower-client-app client:wandb_app --insecure
for W&B monitoring, or:
flower-client-app client:tb_app --insecure
for TensorBoard.
In yet another new terminal window, start the second long-running Flower client (with the mod of your choice):
flower-client-app client:{wandb,tb}_app --insecure
Run the Flower App#
With both the long-running server (SuperLink) and two clients (SuperNode) up and running, we can now run the actual Flower App:
flower-server-app server:app --insecure
Check the results#
For W&B, you will need to login to the website.
For TensorBoard, you will need to run the following command in your terminal:
tensorboard --logdir <LOG_DIR>
Where <LOG_DIR>
needs to be replaced by the directory passed as an argument to the wrapper function (.runs_history/
by default).