In this tutorial, we build an advanced federated learning experiment with NVIDIA FLARE. We compare FedAvg and FedProx on a non-IID CIFAR-10 setup, where client data is split using a Dirichlet distribution to simulate realistic label imbalance across federated sites. We use the NVFlare Job API to define and launch federated jobs, while the Client API handles local training, model exchange, and communication between simulated clients and the server. Finally, we run both algorithms on the same partitioned dataset and visualize how their global model accuracy evolves across communication rounds.
We install the required libraries and import the main packages needed for the federated learning experiment. We define the experiment settings, including the number of clients, training rounds, batch size, learning rate, and non-IID Dirichlet alpha value. We also create the required data and results folders, then download CIFAR-10 once, so that all simulated clients can reuse the same dataset safely.
Copy CodeCopiedUse a different Browser
CLIENT_SCRIPT = r''' import argparse, os, csv import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from torch.utils.data import DataLoader, Subset import torchvision import torchvision.transforms as T import nvflare.client as flare class Net(nn.Module): """Small CNN for CIFAR-10 (no batchnorm -> clean state_dict for FedAvg).""" def __init__(self): super().__init__() self.conv1 = nn.Conv2d(3, 32, 3, padding=1) self.conv2 = nn.Conv2d(32, 64, 3, padding=1) self.pool = nn.MaxPool2d(2, 2) self.fc1 = nn.Linear(64 * 8 * 8, 128) self.fc2 = nn.Linear(128, 10) def forward(self, x): x = self.pool(F.relu(self.conv1(x))) x = self.pool(F.relu(self.conv2(x))) x = x.flatten(1) x = F.relu(self.fc1(x)) return self.fc2(x) def dirichlet_partition(labels, num_sites, alpha, seed=42): """Deterministic non-IID label-skew partition. All client processes use the same seed, so they independently agree on the same global split.""" rng = np.random.default_rng(seed) num_classes = int(labels.max()) + 1 site_idx = [[] for _ in range(num_sites)] for c in range(num_classes): idx_c = np.where(labels == c)[0] rng.shuffle(idx_c) props = rng.dirichlet([alpha] * num_sites) cuts = (np.cumsum(props) * len(idx_c)).astype(int)[:-1] for s, part in enumerate(np.split(idx_c, cuts)): site_idx[s].extend(part.tolist()) return [np.array(s) for s in site_idx] @torch.no_grad() def evaluate(model, loader, device): model.eval() correct = total = 0 for x, y in loader: x, y = x.to(device), y.to(device) pred = model(x).argmax(1) correct += (pred == y).sum().item() total += y.size(0) return correct / total '''
In this tutorial, we build an advanced federated learning experiment with NVIDIA FLARE. We compare FedAvg and FedProx on a non-IID CIFAR-10 setup, where client data is split using a Dirichlet distribution to simulate realistic label imbalance across federated sites. We use the NVFlare Job API to define and launch federated jobs, while the Client API handles local training, model exchange, and communication between simulated clients and the server. Finally, we run both algorithms on the same partitioned dataset and visualize how their global model accuracy evolves across communication rounds.
Copy CodeCopiedUse a different Browser
We create the client-side training script as a separate Python file so NVFlare can import and run it inside each simulated client process. We define a small CNN model for CIFAR-10 classification and add a deterministic Dirichlet partitioning function to create non-IID client datasets. We also define an evaluation function that measures the global model’s accuracy on the shared CIFAR-10 test set.
Copy CodeCopiedUse a different Browser
CLIENT_SCRIPT += r''' def main(): p = argparse.ArgumentParser() p.add_argument("--num_sites", type=int, default=3) p.add_argument("--alpha", type=float, default=0.3) p.add_argument("--local_epochs", type=int, default=1) p.add_argument("--mu", type=float, default=0.0) p.add_argument("--max_samples", type=int, default=4000) p.add_argument("--batch_size", type=int, default=64) p.add_argument("--lr", type=float, default=0.01) p.add_argument("--data_root", type=str, default="/tmp/nvflare/data") p.add_argument("--results_dir", type=str, default="/tmp/nvflare/results") p.add_argument("--tag", type=str, default="fedavg") args = p.parse_args() device = "cuda" if torch.cuda.is_available() else "cpu" tf = T.Compose([T.ToTensor(), T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]) train_set = torchvision.datasets.CIFAR10(args.data_root, train=True, download=False, transform=tf) test_set = torchvision.datasets.CIFAR10(args.data_root, train=False, download=False, transform=tf) flare.init() site_name = flare.get_site_name() site_id = int(site_name.split("-")[-1]) - 1 labels = np.array(train_set.targets) my_idx = dirichlet_partition(labels, args.num_sites, args.alpha)[site_id] if len(my_idx) > args.max_samples: my_idx = my_idx[:args.max_samples] train_loader = DataLoader(Subset(train_set, my_idx), batch_size=args.batch_size, shuffle=True) test_loader = DataLoader(test_set, batch_size=512, shuffle=False) print(f"[{site_name}] mu={args.mu} local samples={len(my_idx)}", flush=True) model = Net().to(device) optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=0.9) criterion = nn.CrossEntropyLoss() while flare.is_running(): input_model = flare.receive() rnd = input_model.current_round global_state = {k: torch.as_tensor(v) for k, v in input_model.params.items()} model.load_state_dict(global_state) acc = evaluate(model, test_loader, device) if site_id == 0: with open(os.path.join(args.results_dir, args.tag + ".csv"), "a", newline="") as f: csv.writer(f).writerow([rnd, acc]) print(f"[{site_name}] round {rnd}: global test acc = {acc:.4f}", flush=True) global_w = [w.detach().clone() for w in model.parameters()] model.train() steps = 0 for _ in range(args.local_epochs): for x, y in train_loader: x, y = x.to(device), y.to(device) optimizer.zero_grad() loss = criterion(model(x), y) if args.mu > 0: prox = sum(((w - g) ** 2).sum() for w, g in zip(model.parameters(), global_w)) loss = loss + (args.mu / 2.0) * prox loss.backward() optimizer.step() steps += 1 out = flare.FLModel( params={k: v.cpu().numpy() for k, v in model.state_dict().items()}, metrics={"test_accuracy": acc}, meta={"NUM_STEPS_CURRENT_ROUND": steps}, ) flare.send(out) if __name__ == "__main__": main() ''' with open("client_train.py", "w") as f: f.write(CLIENT_SCRIPT) sys.path.insert(0, os.getcwd()) from client_train import Net
We build the main client training workflow that runs inside every NVFlare site. We initialize the NVFlare Client API, identify the current client site, load that client’s non-IID data shard, and prepare the local model, optimizer, and loss function. We then receive global weights from the server, evaluate the model, train locally with either FedAvg or FedProx, and send the updated model back for aggregation.
Copy CodeCopiedUse a different Browser
from nvflare.app_opt.pt.job_config.fed_avg import FedAvgJob from nvflare.job_config.script_runner import ScriptRunner def run_experiment(tag, mu): csv_path = os.path.join(RESULTS_DIR, tag + ".csv") if os.path.exists(csv_path): os.remove(csv_path) workspace = f"/tmp/nvflare/{tag}_ws" shutil.rmtree(workspace, ignore_errors=True) job = FedAvgJob( name=f"cifar10_{tag}", n_clients=NUM_SITES, num_rounds=NUM_ROUNDS, initial_model=Net(), ) args = (f"--num_sites {NUM_SITES} --alpha {ALPHA} --local_epochs {LOCAL_EPOCHS} " f"--mu {mu} --max_samples {MAX_SAMPLES} --batch_size {BATCH_SIZE} --lr {LR} " f"--data_root {DATA_ROOT} --results_dir {RESULTS_DIR} --tag {tag}") for i in range(NUM_SITES): job.to(ScriptRunner(script="client_train.py", script_args=args), target=f"site-{i+1}") gpu = "0" if torch.cuda.is_available() else None print(f"\n===== Running {tag} (mu={mu}) on {'GPU' if gpu else 'CPU'} =====") job.simulator_run(workspace, gpu=gpu) return workspace run_experiment("fedavg", mu=0.0) run_experiment("fedprox", mu=0.1)
We define the server-side federated experiment using the NVFlare Job API. We create a FedAvg job, attach the client training script to each simulated site, and pass all experimental arguments, such as alpha, learning rate, local epochs, and the FedProx mu value. We then run two experiments on the same non-IID setup: one with standard FedAvg and another with FedProx.
Copy CodeCopiedUse a different Browser
def load_curve(tag): rounds, accs = [], [] with open(os.path.join(RESULTS_DIR, tag + ".csv")) as f: for line in f: r, a = line.strip().split(",") rounds.append(int(r)); accs.append(float(a)) order = np.argsort(rounds) return np.array(rounds)[order], np.array(accs)[order] plt.figure(figsize=(7, 4.5)) for tag, label in [("fedavg", "FedAvg"), ("fedprox", "FedProx (mu=0.1)")]: r, a = load_curve(tag) plt.plot(r, a, marker="o", label=label) plt.title(f"Global model accuracy on non-IID CIFAR-10 (Dirichlet alpha={ALPHA}, {NUM_SITES} clients)") plt.xlabel("Federated round"); plt.ylabel("Global test accuracy") plt.grid(alpha=0.3); plt.legend(); plt.tight_layout(); plt.show() for tag in ["fedavg", "fedprox"]: try: ckpt = glob.glob(f"/tmp/nvflare/{tag}_ws/**/FL_global_model.pt", recursive=True)[0] obj = torch.load(ckpt, map_location="cpu", weights_only=False) state = obj.get("model", obj) if isinstance(obj, dict) else obj print(f"[{tag}] final global model checkpoint: {ckpt}") except Exception as e: print(f"[{tag}] could not locate final checkpoint ({e})")
We load the saved accuracy logs for both FedAvg and FedProx and sort them by communication round. We plot the global test accuracy curves to visually compare how the two methods perform on non-IID CIFAR-10 data. We also try to locate the final aggregated global model checkpoint from each NVFlare workspace for optional inspection or reuse.
In conclusion, we completed an end-to-end federated learning workflow using NVIDIA FLARE, from preparing non-IID CIFAR-10 client shards to training and comparing FedAvg and FedProx. We saw how the Job API helps us configure server-side orchestration, while the Client API lets each simulated client receive global weights, train locally, and send updated models back for aggregation. We also tracked global test accuracy across rounds and plotted learning curves to understand how FedProx differs from standard FedAvg under heterogeneous data conditions.
Check out the Full Codes here. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
The post Step by Step Guide to Build and Compare FedAvg and FedProx Federated Learning on Non-IID CIFAR-10 with NVIDIA FLARE appeared first on MarkTechPost.
We install the required libraries and import the main packages needed for the federated learning experiment. We define the experiment settings, including the number of clients, training rounds, batch size, learning rate, and non-IID Dirichlet alpha value. We also create the required data and results folders, then download CIFAR-10 once, so that all simulated clients can reuse the same dataset safely.
Copy CodeCopiedUse a different Browser
CLIENT_SCRIPT = r''' import argparse, os, csv import numpy as np import torch import torch.nn as nn import torch.nn.functional as F from torch.utils.data import DataLoader, Subset import torchvision import torchvision.transforms as T import nvflare.client as flare class Net(nn.Module): """Small CNN for CIFAR-10 (no batchnorm -> clean state_dict for FedAvg).""" def __init__(self): super().__init__() self.conv1 = nn.Conv2d(3, 32, 3, padding=1) self.conv2 = nn.Conv2d(32, 64, 3, padding=1) self.pool = nn.MaxPool2d(2, 2) self.fc1 = nn.Linear(64 * 8 * 8, 128) self.fc2 = nn.Linear(128, 10) def forward(self, x): x = self.pool(F.relu(self.conv1(x))) x = self.pool(F.relu(self.conv2(x))) x = x.flatten(1) x = F.relu(self.fc1(x)) return self.fc2(x) def dirichlet_partition(labels, num_sites, alpha, seed=42): """Deterministic non-IID label-skew partition. All client processes use the same seed, so they independently agree on the same global split.""" rng = np.random.default_rng(seed) num_classes = int(labels.max()) + 1 site_idx = [[] for _ in range(num_sites)] for c in range(num_classes): idx_c = np.where(labels == c)[0] rng.shuffle(idx_c) props = rng.dirichlet([alpha] * num_sites) cuts = (np.cumsum(props) * len(idx_c)).astype(int)[:-1] for s, part in enumerate(np.split(idx_c, cuts)): site_idx[s].extend(part.tolist()) return [np.array(s) for s in site_idx] @torch.no_grad() def evaluate(model, loader, device): model.eval() correct = total = 0 for x, y in loader: x, y = x.to(device), y.to(device) pred = model(x).argmax(1) correct += (pred == y).sum().item() total += y.size(0) return correct / total '''
We create the client-side training script as a separate Python file so NVFlare can import and run it inside each simulated client process. We define a small CNN model for CIFAR-10 classification and add a deterministic Dirichlet partitioning function to create non-IID client datasets. We also define an evaluation function that measures the global model’s accuracy on the shared CIFAR-10 test set.
Copy CodeCopiedUse a different Browser
CLIENT_SCRIPT += r''' def main(): p = argparse.ArgumentParser() p.add_argument("--num_sites", type=int, default=3) p.add_argument("--alpha", type=float, default=0.3) p.add_argument("--local_epochs", type=int, default=1) p.add_argument("--mu", type=float, default=0.0) p.add_argument("--max_samples", type=int, default=4000) p.add_argument("--batch_size", type=int, default=64) p.add_argument("--lr", type=float, default=0.01) p.add_argument("--data_root", type=str, default="/tmp/nvflare/data") p.add_argument("--results_dir", type=str, default="/tmp/nvflare/results") p.add_argument("--tag", type=str, default="fedavg") args = p.parse_args() device = "cuda" if torch.cuda.is_available() else "cpu" tf = T.Compose([T.ToTensor(), T.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))]) train_set = torchvision.datasets.CIFAR10(args.data_root, train=True, download=False, transform=tf) test_set = torchvision.datasets.CIFAR10(args.data_root, train=False, download=False, transform=tf) flare.init() site_name = flare.get_site_name() site_id = int(site_name.split("-")[-1]) - 1 labels = np.array(train_set.targets) my_idx = dirichlet_partition(labels, args.num_sites, args.alpha)[site_id] if len(my_idx) > args.max_samples: my_idx = my_idx[:args.max_samples] train_loader = DataLoader(Subset(train_set, my_idx), batch_size=args.batch_size, shuffle=True) test_loader = DataLoader(test_set, batch_size=512, shuffle=False) print(f"[{site_name}] mu={args.mu} local samples={len(my_idx)}", flush=True) model = Net().to(device) optimizer = torch.optim.SGD(model.parameters(), lr=args.lr, momentum=0.9) criterion = nn.CrossEntropyLoss() while flare.is_running(): input_model = flare.receive() rnd = input_model.current_round global_state = {k: torch.as_tensor(v) for k, v in input_model.params.items()} model.load_state_dict(global_state) acc = evaluate(model, test_loader, device) if site_id == 0: with open(os.path.join(args.results_dir, args.tag + ".csv"), "a", newline="") as f: csv.writer(f).writerow([rnd, acc]) print(f"[{site_name}] round {rnd}: global test acc = {acc:.4f}", flush=True) global_w = [w.detach().clone() for w in model.parameters()] model.train() steps = 0 for _ in range(args.local_epochs): for x, y in train_loader: x, y = x.to(device), y.to(device) optimizer.zero_grad() loss = criterion(model(x), y) if args.mu > 0: prox = sum(((w - g) ** 2).sum() for w, g in zip(model.parameters(), global_w)) loss = loss + (args.mu / 2.0) * prox loss.backward() optimizer.step() steps += 1 out = flare.FLModel( params={k: v.cpu().numpy() for k, v in model.state_dict().items()}, metrics={"test_accuracy": acc}, meta={"NUM_STEPS_CURRENT_ROUND": steps}, ) flare.send(out) if __name__ == "__main__": main() ''' with open("client_train.py", "w") as f: f.write(CLIENT_SCRIPT) sys.path.insert(0, os.getcwd()) from client_train import Net
We build the main client training workflow that runs inside every NVFlare site. We initialize the NVFlare Client API, identify the current client site, load that client’s non-IID data shard, and prepare the local model, optimizer, and loss function. We then receive global weights from the server, evaluate the model, train locally with either FedAvg or FedProx, and send the updated model back for aggregation.
Copy CodeCopiedUse a different Browser
from nvflare.app_opt.pt.job_config.fed_avg import FedAvgJob from nvflare.job_config.script_runner import ScriptRunner def run_experiment(tag, mu): csv_path = os.path.join(RESULTS_DIR, tag + ".csv") if os.path.exists(csv_path): os.remove(csv_path) workspace = f"/tmp/nvflare/{tag}_ws" shutil.rmtree(workspace, ignore_errors=True) job = FedAvgJob( name=f"cifar10_{tag}", n_clients=NUM_SITES, num_rounds=NUM_ROUNDS, initial_model=Net(), ) args = (f"--num_sites {NUM_SITES} --alpha {ALPHA} --local_epochs {LOCAL_EPOCHS} " f"--mu {mu} --max_samples {MAX_SAMPLES} --batch_size {BATCH_SIZE} --lr {LR} " f"--data_root {DATA_ROOT} --results_dir {RESULTS_DIR} --tag {tag}") for i in range(NUM_SITES): job.to(ScriptRunner(script="client_train.py", script_args=args), target=f"site-{i+1}") gpu = "0" if torch.cuda.is_available() else None print(f"\n===== Running {tag} (mu={mu}) on {'GPU' if gpu else 'CPU'} =====") job.simulator_run(workspace, gpu=gpu) return workspace run_experiment("fedavg", mu=0.0) run_experiment("fedprox", mu=0.1)
We define the server-side federated experiment using the NVFlare Job API. We create a FedAvg job, attach the client training script to each simulated site, and pass all experimental arguments, such as alpha, learning rate, local epochs, and the FedProx mu value. We then run two experiments on the same non-IID setup: one with standard FedAvg and another with FedProx.
Copy CodeCopiedUse a different Browser
def load_curve(tag): rounds, accs = [], [] with open(os.path.join(RESULTS_DIR, tag + ".csv")) as f: for line in f: r, a = line.strip().split(",") rounds.append(int(r)); accs.append(float(a)) order = np.argsort(rounds) return np.array(rounds)[order], np.array(accs)[order] plt.figure(figsize=(7, 4.5)) for tag, label in [("fedavg", "FedAvg"), ("fedprox", "FedProx (mu=0.1)")]: r, a = load_curve(tag) plt.plot(r, a, marker="o", label=label) plt.title(f"Global model accuracy on non-IID CIFAR-10 (Dirichlet alpha={ALPHA}, {NUM_SITES} clients)") plt.xlabel("Federated round"); plt.ylabel("Global test accuracy") plt.grid(alpha=0.3); plt.legend(); plt.tight_layout(); plt.show() for tag in ["fedavg", "fedprox"]: try: ckpt = glob.glob(f"/tmp/nvflare/{tag}_ws/**/FL_global_model.pt", recursive=True)[0] obj = torch.load(ckpt, map_location="cpu", weights_only=False) state = obj.get("model", obj) if isinstance(obj, dict) else obj print(f"[{tag}] final global model checkpoint: {ckpt}") except Exception as e: print(f"[{tag}] could not locate final checkpoint ({e})")
We load the saved accuracy logs for both FedAvg and FedProx and sort them by communication round. We plot the global test accuracy curves to visually compare how the two methods perform on non-IID CIFAR-10 data. We also try to locate the final aggregated global model checkpoint from each NVFlare workspace for optional inspection or reuse.
In conclusion, we completed an end-to-end federated learning workflow using NVIDIA FLARE, from preparing non-IID CIFAR-10 client shards to training and comparing FedAvg and FedProx. We saw how the Job API helps us configure server-side orchestration, while the Client API lets each simulated client receive global weights, train locally, and send updated models back for aggregation. We also tracked global test accuracy across rounds and plotted learning curves to understand how FedProx differs from standard FedAvg under heterogeneous data conditions.
Check out the Full Codes here. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
The post Step by Step Guide to Build and Compare FedAvg and FedProx Federated Learning on Non-IID CIFAR-10 with NVIDIA FLARE appeared first on MarkTechPost.