From c8f5a46de844dce7606ff3e10753cb1b7d2ffe6f Mon Sep 17 00:00:00 2001
From: leahcimali <michaelbenalipro@gmail.com>
Date: Tue, 22 Oct 2024 14:07:17 +0200
Subject: [PATCH] Correct errors for run_exp and datatransform
---
run_exp.py | 45 +++++++++++++++++++++++--------------------
src/utils_data.py | 41 ++++++++++++++++++++++++---------------
src/utils_training.py | 6 ++++--
3 files changed, 54 insertions(+), 38 deletions(-)
diff --git a/run_exp.py b/run_exp.py
index b0399a6..16c8e31 100644
--- a/run_exp.py
+++ b/run_exp.py
@@ -4,31 +4,34 @@ import subprocess
# Path to your CSV file
csv_file = "exp_configs.csv"
-# Read the second line from the CSV file
+# Open and read the CSV file
with open(csv_file, newline='') as csvfile:
reader = csv.reader(csvfile)
- # Skip the header (if any) and the first row
- next(reader) # Skipping the header
- row = next(reader) # Reading the second row
+ # Skip the header (first row)
+ next(reader)
- # Assigning CSV values to variables
- exp_type, dataset, nn_model, heterogeneity_type, num_clients, num_samples_by_label, num_clusters, centralized_epochs, federated_rounds, seed = row
+ # Iterate over each row (experiment configuration) in the CSV file
+ for row in reader:
+ # Assigning CSV values to variables
+ exp_type, dataset, nn_model, heterogeneity_type, num_clients, num_samples_by_label, num_clusters, centralized_epochs, federated_rounds, seed = row
- # Building the command
- command = [
- "python", "driver.py",
- "--exp_type", exp_type,
- "--dataset", dataset,
+ # Building the command to run the driver.py script with the corresponding arguments
+ command = [
+ "python", "driver.py",
+ "--exp_type", exp_type,
+ "--dataset", dataset,
+ "--nn_model", nn_model,
+ "--heterogeneity_type", heterogeneity_type,
+ "--num_clients", num_clients,
+ "--num_samples_by_label", num_samples_by_label,
+ "--num_clusters", num_clusters,
+ "--centralized_epochs", centralized_epochs,
+ "--federated_rounds", federated_rounds,
+ "--seed", seed]
- "--nn_model", nn_model,
- "--heterogeneity_type", heterogeneity_type,
- "--num_clients", num_clients,
- "--num_samples_by_label", num_samples_by_label,
- "--num_clusters", num_clusters,
- "--centralized_epochs", centralized_epochs,
- "--federated_rounds", federated_rounds,
- "--seed", seed]
+ # Print the command to check it before running (optional)
+ print(f"Running command: {' '.join(command)}")
- # Run the command
- subprocess.run(command)
+ # Run the command
+ subprocess.run(command)
diff --git a/src/utils_data.py b/src/utils_data.py
index db1d33e..1d67a06 100644
--- a/src/utils_data.py
+++ b/src/utils_data.py
@@ -183,7 +183,7 @@ class AddRandomJitter(object):
saturation = self.saturation, hue = self.hue)
return transform(tensor)
-class CifarDataset(Dataset):
+class CustomDataset(Dataset):
def __init__(self, data, labels, transform=None):
# Ensure data is in (N, H, W, C) format
@@ -257,9 +257,9 @@ def data_preparation(client: Client, row_exp: dict) -> None:
# Create datasets with transformations
- train_dataset = CifarDataset(x_train, y_train, transform=train_transform)
- val_dataset = CifarDataset(x_val, y_val, transform=test_val_transform)
- test_dataset = CifarDataset(x_test, y_test, transform=test_val_transform)
+ train_dataset = CustomDataset(x_train, y_train, transform=train_transform)
+ val_dataset = CustomDataset(x_val, y_val, transform=test_val_transform)
+ test_dataset = CustomDataset(x_test, y_test, transform=test_val_transform)
# Create DataLoaders
train_loader = DataLoader(train_dataset, batch_size=128, shuffle=True)
@@ -616,12 +616,14 @@ def swap_labels(labels : list, client : Client, heterogeneity_class : int) -> Cl
return client
-def centralize_data(list_clients: list) -> Tuple[DataLoader, DataLoader]:
+def centralize_data(list_clients: list, row_exp: dict) -> Tuple[DataLoader, DataLoader]:
"""Centralize data of the federated learning setup for central model comparison
Arguments:
list_clients : The list of Client Objects
+ row_exp : The current experiment's global parameters
+
Returns:
Train and test torch DataLoaders with data of all Clients
"""
@@ -630,18 +632,26 @@ def centralize_data(list_clients: list) -> Tuple[DataLoader, DataLoader]:
import torch
from torch.utils.data import DataLoader,TensorDataset
import numpy as np
-
- train_transform = transforms.Compose([
+
+ if row_exp['dataset'] == 'cifar10':
+ train_transform = transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.RandomRotation(20), # Normalize if needed
transforms.RandomCrop(32, padding=4),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
+ ])
+ # Transform for validation and test data (no augmentation, just normalization)
+ test_val_transform = transforms.Compose([
+ transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)), ])
+ else :
+ train_transform = transforms.Compose([
+ transforms.Normalize((0.5,), (0.5,)), # Normalize if needed
])
-
- # Transform for validation and test data (no augmentation, just normalization)
- test_val_transform = transforms.Compose([
- transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)), # Normalize if needed
- ])
+ # Transform for validation and test data (no augmentation, just normalization)
+ test_val_transform = transforms.Compose([
+ transforms.Normalize((0.5,), (0.5,)), # Normalize if needed
+ ])
+
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
@@ -658,9 +668,9 @@ def centralize_data(list_clients: list) -> Tuple[DataLoader, DataLoader]:
y_test = np.concatenate([list_clients[id].train_test['y_test'] for id in range(len(list_clients))], axis=0)
# Create Custom Datasets
- train_dataset = CifarDataset(x_train, y_train, transform=train_transform)
- val_dataset = CifarDataset(x_val, y_val, transform=test_val_transform)
- test_dataset = CifarDataset(x_test, y_test, transform=test_val_transform)
+ train_dataset = CustomDataset(x_train, y_train, transform=train_transform)
+ val_dataset = CustomDataset(x_val, y_val, transform=test_val_transform)
+ test_dataset = CustomDataset(x_test, y_test, transform=test_val_transform)
# Create DataLoaders
train_loader = DataLoader(train_dataset, batch_size=128, shuffle=True)
@@ -672,6 +682,7 @@ def centralize_data(list_clients: list) -> Tuple[DataLoader, DataLoader]:
+
def unbalancing(client : Client ,labels_list : list ,ratio_list: list) -> Client :
""" Downsample the dataset of a client with each elements of the labels_list will be downsampled by the corresponding ration of ratio_list
diff --git a/src/utils_training.py b/src/utils_training.py
index d1cc6d0..8c77d04 100644
--- a/src/utils_training.py
+++ b/src/utils_training.py
@@ -115,7 +115,7 @@ def run_benchmark(main_model : nn.Module, list_clients : list, row_exp : dict) -
if row_exp['exp_type'] == 'pers-centralized':
for heterogeneity_class in list_heterogeneities:
list_clients_filtered = [client for client in list_clients if client.heterogeneity_class == heterogeneity_class]
- train_loader, val_loader, test_loader = centralize_data(list_clients_filtered)
+ train_loader, val_loader, test_loader = centralize_data(list_clients_filtered,row_exp)
model_trained, _ = train_central(curr_model, train_loader, val_loader, row_exp)
global_acc = test_model(model_trained, test_loader)
@@ -129,7 +129,7 @@ def run_benchmark(main_model : nn.Module, list_clients : list, row_exp : dict) -
model_server = copy.deepcopy(curr_model)
model_trained = train_federated(model_server, list_clients, row_exp, use_cluster_models = False)
- _, _,test_loader = centralize_data(list_clients)
+ _, _,test_loader = centralize_data(list_clients,row_exp)
global_acc = test_model(model_trained.model, test_loader)
for client in list_clients :
@@ -271,6 +271,8 @@ def test_model(model: nn.Module, test_loader: DataLoader) -> float:
for batch in test_loader:
inputs, labels = [item.to(device) for item in batch]
+
+ labels = labels.long()
outputs = model(inputs)
# Compute the loss
--
GitLab