MNIST classification with Scikit-Learn Classifier (Perceptron)¶
Overview of the tutorial:
In this tutorial, we are going to train Scikit-Learn Perceptron as a federated model over a Node.
At the end of this tutorial, you will learn:
- how to define a Sklearn classifier in Fed-BioMed (especially
Perceptronmodel) - how to train it
- how to evaluate the resulting model
HINT : to reload the notebook, please click on the following button:
Kernel -> Restart and clear Output
1. Clean your environments¶
Before executing notebook and starting nodes, it is safer to remove all configuration scripts automatically generated by Fed-BioMed. To do so, enter the following in a terminal:
2. Setting the node up¶
fedbiomed node dataset add
- Select option 2 (default) to add MNIST to the node
- Confirm default tags by hitting "y" and ENTER
- Pick the folder where MNIST is downloaded (this is due torch issue https://github.com/pytorch/vision/issues/3549)
- Data must have been added (if you get a warning saying that data must be unique is because it has been already added)
- Check that your data has been added by executing
fedbiomed node dataset list - Run the node using
fedbiomed node start. Wait until you getConnected with result code 0. it means your node is working and ready to participate to a Federated training.
More details are given in tutorial : Installation/setting up environment
3. Create Sklearn Federated Perceptron Training Plan¶
The class FedPerceptron constitutes the Fed-BioMed wrapper for executing Federated Learning using Scikit-Learn Perceptron model based on mini-batch Stochastic Gradient Descent (SGD). As we have done with Pytorch model in previous chapter, we create a new training plan class SkLearnClassifierTrainingPlan that inherits from it. For a refresher on how Training Plans work in Fed-BioMed, please refer to our Training Plan user guide.
In scikit-learn Training Plans, you typically need to define only the training_data function, and optionally an init_dependencies function if your code requires additional module imports.
The training_data function defines how datasets should be loaded in nodes to make them ready for training. It takes a batch_size argument and returns a DataManager class. For scikit-learn, the DataManager must be instantiated with a dataset and a target argument, both np.ndarrays of the same length.
In [ ]:
Copied!
from fedbiomed.common.training_plans import FedPerceptron
from fedbiomed.common.datamanager import DataManager
from fedbiomed.common.dataset import MnistDataset
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import FunctionTransformer
import numpy as np
class SkLearnClassifierTrainingPlan(FedPerceptron):
def init_dependencies(self):
"""Define additional dependencies."""
return ["from sklearn.pipeline import Pipeline",
"from sklearn.preprocessing import FunctionTransformer",
"from fedbiomed.common.dataset import MnistDataset",
"import numpy as np"]
def training_data(self):
"""Prepare data for training.
This function loads a MNIST dataset from the node's filesystem, applies some
preprocessing and converts the full dataset to a numpy array.
Finally, it returns a DataManager created with these numpy arrays.
"""
pipeline = Pipeline([
("norm", FunctionTransformer(
lambda im: (np.asarray(im, dtype=np.float64) / 255.0 - 0.1307) / 0.3081,
validate=False
)),
("flatten", FunctionTransformer(
lambda x: np.ascontiguousarray(x.reshape(-1), dtype=np.float64),
validate=False
)),
])
dataset = MnistDataset(transform=pipeline.transform)
return DataManager(dataset=dataset, shuffle=False)
from fedbiomed.common.training_plans import FedPerceptron from fedbiomed.common.datamanager import DataManager from fedbiomed.common.dataset import MnistDataset from sklearn.pipeline import Pipeline from sklearn.preprocessing import FunctionTransformer import numpy as np class SkLearnClassifierTrainingPlan(FedPerceptron): def init_dependencies(self): """Define additional dependencies.""" return ["from sklearn.pipeline import Pipeline", "from sklearn.preprocessing import FunctionTransformer", "from fedbiomed.common.dataset import MnistDataset", "import numpy as np"] def training_data(self): """Prepare data for training. This function loads a MNIST dataset from the node's filesystem, applies some preprocessing and converts the full dataset to a numpy array. Finally, it returns a DataManager created with these numpy arrays. """ pipeline = Pipeline([ ("norm", FunctionTransformer( lambda im: (np.asarray(im, dtype=np.float64) / 255.0 - 0.1307) / 0.3081, validate=False )), ("flatten", FunctionTransformer( lambda x: np.ascontiguousarray(x.reshape(-1), dtype=np.float64), validate=False )), ]) dataset = MnistDataset(transform=pipeline.transform) return DataManager(dataset=dataset, shuffle=False)
Provide dynamic arguments for the model and training. These may potentially be changed at every round.
Model arguments¶
model_args is a dictionary with the arguments related to the model, that will be passed to the Perceptron constructor.
IMPORTANT For classification tasks, you are required to specify the following two fields:
n_features: the number of features in each input sample (in our case, the number of pixels in the images)n_classes: the number of classes in the target data
Furthermore, the classes may not be represented by arbitrary values: classes must be identified by integers in the range 0..n_classes
Training arguments¶
training_args is a dictionary containing the arguments for the training routine (e.g. batch size, learning rate, epochs, etc.). This will be passed to the routine on the node side.
In [ ]:
Copied!
model_args = {'n_features': 28*28,
'n_classes' : 10,
'eta0':1e-6,
'random_state':1234,
'alpha':0.1 }
training_args = {
'epochs': 3,
'batch_maxnum': 20, # can be used to debugging to limit the number of batches per epoch
# 'log_interval': 1, # output a logging message every log_interval batches
'loader_args': {
'batch_size': 4,
},
}
model_args = {'n_features': 28*28, 'n_classes' : 10, 'eta0':1e-6, 'random_state':1234, 'alpha':0.1 } training_args = { 'epochs': 3, 'batch_maxnum': 20, # can be used to debugging to limit the number of batches per epoch # 'log_interval': 1, # output a logging message every log_interval batches 'loader_args': { 'batch_size': 4, }, }
4. Train your model on MNIST dataset¶
MNIST dataset is composed of handwritten digits images, from 0 to 9. The purpose of our classifier is to associate an image to the corresponding represented digit
In [ ]:
Copied!
from fedbiomed.researcher.federated_workflows import Experiment
from fedbiomed.researcher.aggregators.fedavg import FedAverage
tags = ['#MNIST', '#dataset']
rounds = 3
# select nodes participating in this experiment
exp = Experiment(tags=tags,
model_args=model_args,
training_plan_class=SkLearnClassifierTrainingPlan,
training_args=training_args,
round_limit=rounds,
aggregator=FedAverage(),
node_selection_strategy=None)
from fedbiomed.researcher.federated_workflows import Experiment from fedbiomed.researcher.aggregators.fedavg import FedAverage tags = ['#MNIST', '#dataset'] rounds = 3 # select nodes participating in this experiment exp = Experiment(tags=tags, model_args=model_args, training_plan_class=SkLearnClassifierTrainingPlan, training_args=training_args, round_limit=rounds, aggregator=FedAverage(), node_selection_strategy=None)
In [ ]:
Copied!
exp.run(increase=True)
exp.run(increase=True)
Save trained model to file
In [ ]:
Copied!
exp.training_plan().export_model('./trained_model')
exp.training_plan().export_model('./trained_model')
5. Testing on MNIST test dataset¶
Let's assess performance of our classifier with MNIST testing dataset
In [ ]:
Copied!
import numpy as np
from fedbiomed.common.dataset_types import DataReturnFormat
from fedbiomed.researcher.config import config
pipeline = Pipeline([
("norm", FunctionTransformer(
lambda im: (np.asarray(im, dtype=np.float64) / 255.0 - 0.1307) / 0.3081,
validate=False
)),
("flatten", FunctionTransformer(
lambda x: np.ascontiguousarray(x.reshape(-1), dtype=np.float64),
validate=False
)),
])
testing_MNIST_dataset = MnistDataset(transform=pipeline.transform)
testing_MNIST_dataset.complete_initialization(controller_kwargs={'root': config.root + '/notebooks/data', 'train': False}, to_format=DataReturnFormat.SKLEARN)
testing_MNIST_data = []
testing_MNIST_targets = []
for i in range(len(testing_MNIST_dataset)):
testing_MNIST_data.append(testing_MNIST_dataset[i][0])
testing_MNIST_targets.append(testing_MNIST_dataset[i][1])
testing_MNIST_data = np.array(testing_MNIST_data)
testing_MNIST_targets = np.array(testing_MNIST_targets)
import numpy as np from fedbiomed.common.dataset_types import DataReturnFormat from fedbiomed.researcher.config import config pipeline = Pipeline([ ("norm", FunctionTransformer( lambda im: (np.asarray(im, dtype=np.float64) / 255.0 - 0.1307) / 0.3081, validate=False )), ("flatten", FunctionTransformer( lambda x: np.ascontiguousarray(x.reshape(-1), dtype=np.float64), validate=False )), ]) testing_MNIST_dataset = MnistDataset(transform=pipeline.transform) testing_MNIST_dataset.complete_initialization(controller_kwargs={'root': config.root + '/notebooks/data', 'train': False}, to_format=DataReturnFormat.SKLEARN) testing_MNIST_data = [] testing_MNIST_targets = [] for i in range(len(testing_MNIST_dataset)): testing_MNIST_data.append(testing_MNIST_dataset[i][0]) testing_MNIST_targets.append(testing_MNIST_dataset[i][1]) testing_MNIST_data = np.array(testing_MNIST_data) testing_MNIST_targets = np.array(testing_MNIST_targets)
6. Getting Loss function¶
Here we use the aggregated_params() getter to access all model weights at the end of each round to plot the evolution of Perceptron loss function, as well as its accuracy.
In [ ]:
Copied!
# retrieve Sklearn model and losses at the end of each round
from sklearn.linear_model import SGDClassifier
from sklearn.metrics import accuracy_score, confusion_matrix, hinge_loss
fed_perceptron_model = exp.training_plan().model()
perceptron_args = {key: model_args[key] for key in model_args.keys() if key in fed_perceptron_model.get_params().keys()}
losses = []
accuracies = []
for r in range(rounds):
fed_perceptron_model = fed_perceptron_model.set_params(**perceptron_args)
fed_perceptron_model.classes_ = np.unique(testing_MNIST_targets)
fed_perceptron_model.coef_ = exp.aggregated_params()[r]['params']['coef_'].copy()
fed_perceptron_model.intercept_ = exp.aggregated_params()[r]['params']['intercept_'].copy()
prediction = fed_perceptron_model.decision_function(testing_MNIST_data)
losses.append(hinge_loss(testing_MNIST_targets, prediction))
accuracies.append(fed_perceptron_model.score(testing_MNIST_data, testing_MNIST_targets))
# retrieve Sklearn model and losses at the end of each round from sklearn.linear_model import SGDClassifier from sklearn.metrics import accuracy_score, confusion_matrix, hinge_loss fed_perceptron_model = exp.training_plan().model() perceptron_args = {key: model_args[key] for key in model_args.keys() if key in fed_perceptron_model.get_params().keys()} losses = [] accuracies = [] for r in range(rounds): fed_perceptron_model = fed_perceptron_model.set_params(**perceptron_args) fed_perceptron_model.classes_ = np.unique(testing_MNIST_targets) fed_perceptron_model.coef_ = exp.aggregated_params()[r]['params']['coef_'].copy() fed_perceptron_model.intercept_ = exp.aggregated_params()[r]['params']['intercept_'].copy() prediction = fed_perceptron_model.decision_function(testing_MNIST_data) losses.append(hinge_loss(testing_MNIST_targets, prediction)) accuracies.append(fed_perceptron_model.score(testing_MNIST_data, testing_MNIST_targets))
7. Comparison with a local Perceptron model¶
In this section, we implement a local Perceptron model, so we can compare remote and local models accuracies.
You can use this section as an insight on how things are implemented within the Fed-BioMed network. In particular, looking at the code in the next few cells you may learn how:
- we implement mini-batch gradient descent for scikit-learn models
- we implement
Perceptronbased onSGDClassifier
In [ ]:
Copied!
pipeline = Pipeline([
("norm", FunctionTransformer(
lambda im: (np.asarray(im, dtype=np.float64) / 255.0 - 0.1307) / 0.3081,
validate=False
)),
("flatten", FunctionTransformer(
lambda x: np.ascontiguousarray(x.reshape(-1), dtype=np.float64),
validate=False
)),
])
training_MNIST_dataset = MnistDataset(transform=pipeline.transform)
training_MNIST_dataset.complete_initialization(controller_kwargs={'root': config.root + '/notebooks/data', 'train': True}, to_format=DataReturnFormat.SKLEARN)
training_MNIST_data = []
training_MNIST_targets = []
for i in range(len(testing_MNIST_dataset)):
training_MNIST_data.append(training_MNIST_dataset[i][0])
training_MNIST_targets.append(training_MNIST_dataset[i][1])
training_MNIST_data = np.array(training_MNIST_data)
training_MNIST_targets = np.array(training_MNIST_targets)
pipeline = Pipeline([ ("norm", FunctionTransformer( lambda im: (np.asarray(im, dtype=np.float64) / 255.0 - 0.1307) / 0.3081, validate=False )), ("flatten", FunctionTransformer( lambda x: np.ascontiguousarray(x.reshape(-1), dtype=np.float64), validate=False )), ]) training_MNIST_dataset = MnistDataset(transform=pipeline.transform) training_MNIST_dataset.complete_initialization(controller_kwargs={'root': config.root + '/notebooks/data', 'train': True}, to_format=DataReturnFormat.SKLEARN) training_MNIST_data = [] training_MNIST_targets = [] for i in range(len(testing_MNIST_dataset)): training_MNIST_data.append(training_MNIST_dataset[i][0]) training_MNIST_targets.append(training_MNIST_dataset[i][1]) training_MNIST_data = np.array(training_MNIST_data) training_MNIST_targets = np.array(training_MNIST_targets)
Local Model training loop : a new model is trained locally, then compared with the remote FedPerceptron model
In [ ]:
Copied!
fed_perceptron_model.get_params()
fed_perceptron_model.get_params()
In [ ]:
Copied!
local_perceptron_losses = []
local_perceptron_accuracies = []
classes = np.unique(training_MNIST_targets)
batch_size = training_args["loader_args"]["batch_size"]
# model definition
local_perceptron_model = SGDClassifier()
perceptron_args = {key: model_args[key] for key in model_args.keys() if key in fed_perceptron_model.get_params().keys()}
local_perceptron_model.set_params(**perceptron_args)
model_param_list = ['coef_', 'intercept_']
# Model initialization
local_perceptron_model.intercept_ = np.zeros((model_args["n_classes"],))
local_perceptron_model.coef_ = np.zeros((model_args["n_classes"], model_args["n_features"]))
local_perceptron_losses = [] local_perceptron_accuracies = [] classes = np.unique(training_MNIST_targets) batch_size = training_args["loader_args"]["batch_size"] # model definition local_perceptron_model = SGDClassifier() perceptron_args = {key: model_args[key] for key in model_args.keys() if key in fed_perceptron_model.get_params().keys()} local_perceptron_model.set_params(**perceptron_args) model_param_list = ['coef_', 'intercept_'] # Model initialization local_perceptron_model.intercept_ = np.zeros((model_args["n_classes"],)) local_perceptron_model.coef_ = np.zeros((model_args["n_classes"], model_args["n_features"]))
Implementation of mini-batch SGD
In [ ]:
Copied!
for r in range(rounds):
for e in range(training_args["epochs"]):
tot_samples_processed = 0
for idx_batch in range(training_args["batch_maxnum"]):
param = {k: getattr(local_perceptron_model, k) for k in model_param_list}
grads = {k: np.zeros_like(v) for k, v in param.items()}
# for each sample: 1) call partial_fit 2) accumulate the gradients 3) reset the model parameters
for sample_idx in range(tot_samples_processed, tot_samples_processed+batch_size):
local_perceptron_model.partial_fit(training_MNIST_data[sample_idx:sample_idx+1,:],
training_MNIST_targets[sample_idx:sample_idx+1],
classes=classes)
for key in model_param_list:
grads[key] += getattr(local_perceptron_model, key)
setattr(local_perceptron_model, key, param[key])
tot_samples_processed += batch_size
# after each epoch, we update the model with the averaged gradients over the batch
for key in model_param_list:
setattr(local_perceptron_model, key, grads[key] / batch_size)
predictions = local_perceptron_model.decision_function(testing_MNIST_data)
local_perceptron_losses.append(hinge_loss(testing_MNIST_targets, predictions))
local_perceptron_accuracies.append(local_perceptron_model.score(testing_MNIST_data,
testing_MNIST_targets))
for r in range(rounds): for e in range(training_args["epochs"]): tot_samples_processed = 0 for idx_batch in range(training_args["batch_maxnum"]): param = {k: getattr(local_perceptron_model, k) for k in model_param_list} grads = {k: np.zeros_like(v) for k, v in param.items()} # for each sample: 1) call partial_fit 2) accumulate the gradients 3) reset the model parameters for sample_idx in range(tot_samples_processed, tot_samples_processed+batch_size): local_perceptron_model.partial_fit(training_MNIST_data[sample_idx:sample_idx+1,:], training_MNIST_targets[sample_idx:sample_idx+1], classes=classes) for key in model_param_list: grads[key] += getattr(local_perceptron_model, key) setattr(local_perceptron_model, key, param[key]) tot_samples_processed += batch_size # after each epoch, we update the model with the averaged gradients over the batch for key in model_param_list: setattr(local_perceptron_model, key, grads[key] / batch_size) predictions = local_perceptron_model.decision_function(testing_MNIST_data) local_perceptron_losses.append(hinge_loss(testing_MNIST_targets, predictions)) local_perceptron_accuracies.append(local_perceptron_model.score(testing_MNIST_data, testing_MNIST_targets))
Compare the local and federated models. The two curves should overlap almost identically, although slight numerical errors are acceptable.
In [ ]:
Copied!
import matplotlib.pyplot as plt
plt.figure(figsize=(10,5))
plt.subplot(1,2,1)
plt.plot(losses, label="federated Perceptron losses")
plt.plot(local_perceptron_losses, "--", color='r', label="local Perceptron losses")
plt.ylabel('Perceptron Cost Function (Hinge)')
plt.xlabel('Number of Rounds')
plt.title('Perceptron loss evolution on test dataset')
plt.legend()
plt.subplot(1,2,2)
plt.plot(accuracies, label="federated Perceptron accuracies")
plt.plot(local_perceptron_accuracies, "--", color='r',
label="local Perceptron accuracies")
plt.ylabel('Accuracy')
plt.xlabel('Number of Rounds')
plt.title('Perceptron accuracy over rounds (on test dataset)')
plt.legend()
import matplotlib.pyplot as plt plt.figure(figsize=(10,5)) plt.subplot(1,2,1) plt.plot(losses, label="federated Perceptron losses") plt.plot(local_perceptron_losses, "--", color='r', label="local Perceptron losses") plt.ylabel('Perceptron Cost Function (Hinge)') plt.xlabel('Number of Rounds') plt.title('Perceptron loss evolution on test dataset') plt.legend() plt.subplot(1,2,2) plt.plot(accuracies, label="federated Perceptron accuracies") plt.plot(local_perceptron_accuracies, "--", color='r', label="local Perceptron accuracies") plt.ylabel('Accuracy') plt.xlabel('Number of Rounds') plt.title('Perceptron accuracy over rounds (on test dataset)') plt.legend()
In this example, plots appear to be the same: this means that Federated and local Perceptron models are performing equivalently!
8. Getting accuracy and confusion matrix¶
In [ ]:
Copied!
# federated model predictions
fed_prediction = fed_perceptron_model.predict(testing_MNIST_data)
acc = accuracy_score(testing_MNIST_targets, fed_prediction)
print('Federated Perceptron Model accuracy :', acc)
# local model predictions
local_prediction = local_perceptron_model.predict(testing_MNIST_data)
acc = accuracy_score(testing_MNIST_targets, local_prediction)
print('Local Perceptron Model accuracy :', acc)
# federated model predictions fed_prediction = fed_perceptron_model.predict(testing_MNIST_data) acc = accuracy_score(testing_MNIST_targets, fed_prediction) print('Federated Perceptron Model accuracy :', acc) # local model predictions local_prediction = local_perceptron_model.predict(testing_MNIST_data) acc = accuracy_score(testing_MNIST_targets, local_prediction) print('Local Perceptron Model accuracy :', acc)
In [ ]:
Copied!
def plot_confusion_matrix(fig, ax, conf_matrix, title, xlabel, ylabel, n_image=0):
im = ax[n_image].imshow(conf_matrix)
ax[n_image].set_xticks(np.arange(10))
ax[n_image].set_yticks(np.arange(10))
for i in range(conf_matrix.shape[0]):
for j in range(conf_matrix.shape[1]):
text = ax[n_image].text(j, i, conf_matrix[i, j],
ha="center", va="center", color="w")
ax[n_image].set_xlabel(xlabel)
ax[n_image].set_ylabel(ylabel)
ax[n_image].set_title(title)
def plot_confusion_matrix(fig, ax, conf_matrix, title, xlabel, ylabel, n_image=0): im = ax[n_image].imshow(conf_matrix) ax[n_image].set_xticks(np.arange(10)) ax[n_image].set_yticks(np.arange(10)) for i in range(conf_matrix.shape[0]): for j in range(conf_matrix.shape[1]): text = ax[n_image].text(j, i, conf_matrix[i, j], ha="center", va="center", color="w") ax[n_image].set_xlabel(xlabel) ax[n_image].set_ylabel(ylabel) ax[n_image].set_title(title)
In [ ]:
Copied!
fed_conf_matrix = confusion_matrix(testing_MNIST_targets, fed_prediction)
local_conf_matrix = confusion_matrix(testing_MNIST_targets, local_prediction)
fig, axs = plt.subplots(nrows=1, ncols=2,figsize=(10,5))
plot_confusion_matrix(fig, axs, fed_conf_matrix,
"Federated Perceptron Confusion Matrix",
"Actual values", "Predicted values", n_image=0)
plot_confusion_matrix(fig, axs, local_conf_matrix,
"Local Perceptron Confusion Matrix",
"Actual values", "Predicted values", n_image=1)
fed_conf_matrix = confusion_matrix(testing_MNIST_targets, fed_prediction) local_conf_matrix = confusion_matrix(testing_MNIST_targets, local_prediction) fig, axs = plt.subplots(nrows=1, ncols=2,figsize=(10,5)) plot_confusion_matrix(fig, axs, fed_conf_matrix, "Federated Perceptron Confusion Matrix", "Actual values", "Predicted values", n_image=0) plot_confusion_matrix(fig, axs, local_conf_matrix, "Local Perceptron Confusion Matrix", "Actual values", "Predicted values", n_image=1)
Congrats !¶
You have figured out how to train your first Federated Sklearn classifier model !
If you want to practise more, you can try to deploy such classifier on two or more nodes. As you can see, Perceptron is a limited model: its generalization is SGDCLassifier, provided by Fed-BioMed as a FedSGDCLassifier Training Plan. You can thus try to apply SGDCLassifier, providing more feature such as different cost functions, regularizations and learning rate decays.