Submit a training job (SDK)

Long story short

Your model class, two configs and a folder of shard zips go into one call:

from resontech import TrainingConfig, FederationConfig

job = sdk.rt_submit(
    model=MyResNet,
    name="my-fl-job",
    shards_dir="./shards",
    training=TrainingConfig(num_classes=10, local_epochs=2, batch_size=32, learning_rate=1e-3, img_size=224),
    federation=FederationConfig(num_rounds=5, min_clients=1, job_name="my-fl-job"),
)

print(job.dashboard_url)

Full path

1. Define your model

A standard nn.Module with a num_classes argument in __init__ and a forward(B, 3, H, W) -> logits. The SDK copies the class source into the job as model_def.py.

import torch.nn as nn
from torchvision import models

class MyResNet(nn.Module):
    def __init__(self, num_classes: int):
        super().__init__()
        self.backbone = models.resnet18(weights=None)
        self.backbone.fc = nn.Linear(self.backbone.fc.in_features, num_classes)

    def forward(self, x):
        return self.backbone(x)

To override the executor, the persistor or the data loader, see Custom training classes.

2. Describe training and federation

from resontech import TrainingConfig, FederationConfig

training = TrainingConfig(
    num_classes=10,
    local_epochs=2,
    batch_size=32,
    learning_rate=1e-3,
    img_size=224,
)

federation = FederationConfig(
    num_rounds=5,
    min_clients=1,
    job_name="my-fl-job",
)

Settings without a typed field (mixup, weight decay, a custom scheduler) go into TrainingConfig.extra. Every field: Training config reference.

3. Submit

job = sdk.rt_submit(
    model=MyResNet,
    name="my-fl-job",
    shards_dir="./shards",                     # one shard_*.zip per worker
    requirements_txt="./requirements.txt",     # optional
    training=training,
    federation=federation,
)

print(f"Submitted: {job.id}")
print(f"Track at: {job.dashboard_url}")

rt_submit() returns once your shards are uploaded: a minute or two for a few GB. The number of zips in shards_dir is the number of GPU workers; how to split your data: Dataset format & sharding. You get back a Job with .id and .dashboard_url.

4. Track progress

Open job.dashboard_url to watch the round counter, browse the job’s files and download the trained model. From Python:

• ›sdk.jobs.get(job.id) returns the job’s current state.
• ›sdk.storage.list("jobs/<name>/") lists its files in your bucket.
• ›sdk.storage.presign_get(key) gives a download link for any of them.

More patterns: Track jobs & download results.

What’s next

• ›Tune the configs: Training config reference lists every TrainingConfig and FederationConfig field.
• ›Port a training script you already have: Port a centralized recipe to FL.
• ›Start from a working example: job_training_sdk trains a small CNN on MNIST with a complete submit script.
• ›Browse more recipes: FL examples gallery covers vision, language, speech, embeddings, medical imaging and image generation.