MNIST Dataset#
MNIST is a dataset for classification of image inputs.
We will use HuggingFace
dataset hub for convenience, but you can
also download the data from another source on your side and update the tutorial accordingly.
The goal of this task is to classify a given image of a handwritten digit into one of 10 classes representing integer values from 0 to 9, inclusively.
Please follow this end-to-end tutorial to prepare the dataset, create and train the model, and finally compute explanations.
Prepare the Dataset#
1. Split and Convert your Raw Data#
The first step consists in creating your train, test and validation splits as StandardDataset
.
We load the dataset from HuggingFace
datasets hub for convenience.
The dataset is then split into train, test and validation set.
from datasets import DatasetDict
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict(
{
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
}
)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
As stated in the doc, Xpdeep requires a ".parquet" file to create the dataset.
The original data is stored under a DatasetDict
object, therefore each split must be converted to a ".parquet" file.
Tip
To get your ".parquet" files, you can easily convert each split from pandas.DataFrame
to pyarrow.Table
first.
Warning
Here with set preserve_index
to False in order to remove the DataFrame "index" column from the resulting Pyarrow Table.
import pyarrow as pa
import pyarrow.parquet as pq
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
2. Upload your Converted Data#
Warning
Don't forget to set up a Project
and initialize the API with your credentials !
from xpdeep import init, set_project
from xpdeep.project import Project
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
With your Project
set up, you can upload the converted parquet files into Xpdeep server.
from xpdeep.dataset.upload import upload
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
3. Instantiate a Dataset#
Here we instantiate a ParquetDataset
for the train set only. We will create the validation and test dataset later.
from xpdeep.dataset.parquet_dataset import ParquetDataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
4. Find a schema#
We use the AutoAnalyzer
to get a schema proposal on the train set.
The only requirement is to specify the target name, here the "income" feature. It takes two values: "<=50K" and ">50K".
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
+------------------------------------------------+
| Schema Contents |
+--------------------+---------------+-----------+
| Type | Name | Is Target |
+--------------------+---------------+-----------+
| ImageFeature | image | ❌ |
| CategoricalFeature | label | ✅ |
| IndexMetadata | index_xp_deep | |
+--------------------+---------------+-----------+
Note
Please note that the index_xp_deep
column is automatically computed and stored as a IndexMetadata
in the Schema
.
However, we would like the feature "image" to contain a scaler preprocessing, to scale the pixel values from [0, 255] to [-1, 1]. We will also add data augmentation on preprocessed images by applying a 90° random rotation.
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from torchvision.transforms import Compose, RandomRotation
from xpdeep.dataset.schema.preprocessor import TorchPreprocessor
import torch
class ScaleImage(TorchPreprocessor):
"""Given an image in range [0, 256], scale the pixel values to [-1 ,1]"""
def transform(self, inputs: torch.Tensor) -> torch.Tensor:
"""Transform."""
return inputs / 128.0 - 1.0
def inverse_transform(self, output: torch.Tensor) -> torch.Tensor:
"""Apply inverse transform."""
return (output + 1.0) * 128.0
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image")
analyzed_train_dataset.analyzed_schema["image"] = image_feature
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
5. Fit the schema#
With your Schema
analyzed on the train set, you can now fit the schema to fit each feature preprocessor on the train set.
We use the same FittedSchema
to create a FittedParquetDataset
corresponding to the validation and test set.
from xpdeep.dataset.parquet_dataset import FittedParquetDataset
fit_test_dataset = FittedParquetDataset(split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema)
fit_val_dataset = FittedParquetDataset(split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
And that's all for the dataset preparation. We now have three FittedParquetDataset
, each with its FittedSchema
,
ready to be used.
Prepare the Model#
We need now to create an explainable model XpdeepModel
.
1. Create the required torch models#
We have a multi-class classification task with image input data. We will use a Multi Layer Perceptron (MLP) for this task, in combination with a CNN backbone model.
Tip
Model input and output sizes (including the batch dimension) can be easily retrieved from the fitted schema.
Therefore:
- The
FeatureExtractionModel
will embed input data into a 50 dimension space. - The
TaskLearnerModel
will use aSoftmax
output layer for the output which is of size 2. - The
BackboneModel
will be a CNN with a serie of residual blocks.
Warning
Torch export
requires batch normalization layer torch.nn.BatchNorm1d
to be given as partial with
track_running_stats
False. However, without learnt parameters, the behaviour in inference is not stable. A single
sample may have a different prediction alone or within a batch as the batch is scaled without learnt parameters.
Batch normalization layer will be available when fully compatible with torch export.
from functools import partial
import torch
from xpdeep.model.zoo.mlp import MLP
from xpdeep.model.zoo.cnn import ResidualCNN2D
class MnistCNN(ResidualCNN2D):
"""CNN model definition for dit dataset (tome series)."""
def __init__(self, output_size: int, out_channels: tuple[int, ...] = (16, 32, 64)):
"""Initialize a basic CNN model for classification."""
super().__init__(in_channels=1, out_channels=out_channels, output_size=output_size)
def forward(self, x: torch.Tensor) -> torch.Tensor:
"""MNIST is grayscale."""
return super().forward(x.unsqueeze(1))
feature_extraction = MLP(input_size=128, flatten_input=True, hidden_channels=[128, 64],
activation_layer=partial(torch.nn.ReLU))
task_learner = MLP(input_size=64, activation_layer=partial(torch.nn.ReLU), flatten_input=True,
hidden_channels=[target_size], last_activation=partial(torch.nn.Softmax, dim=1))
backbone = MnistCNN(output_size=128)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
2. Explainable Model Specifications#
Here comes the crucial part: we need to specify model specifications under ModelDecisionGraphParameters
to get the best explanations (Model Decision Graph and Inference Graph).
from xpdeep.model.model_builder import ModelDecisionGraphParameters
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
For further details, see docs
Note
All parameters have a default value, you can start by using those default value, then iterate and update the configuration to find suitable explanations.
3. Create the Explainable Model#
Given the model architecture and configuration, we can finally instantiate the explainable model XpdeepModel
.
from xpdeep.model.xpdeep_model import XpdeepModel
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
Train#
The train step is straightforward: we need to specify the Trainer
parameters.
from xpdeep.trainer.callbacks import EarlyStopping, Scheduler, ModelCheckpoint
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.trainer.trainer import Trainer
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassF1Score, MulticlassConfusionMatrix
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from functools import partial
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2), step_method="epoch",
monitoring_metric="Total loss"),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
Note
Here, the loss is a custom loss compatible with our output format, based on the default torch
loss.
Note
For multiclass metrics, torchmetrics
expects the target to be an index vector and not onehot vector.
As our targets are onehot vectors and not indexes, we add the target_as_indexes
parameter which convert targets
from onehot to indexes prior to the metric computation.
Warning
Here, we set foreach
and fused
to False as currently it may lead to unstable behaviour in the training process.
We can now train the model:
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
The training logs are displayed in the console:
Once the model trained, it can be used to get explanations.
Explain#
Similarly to the Trainer
, explanations are computed with an Explainer
interface.
1. Build the Explainer#
We provide the Explainer
quality metrics to get insights on the explanation quality. In addition, we compute
along with the explanations the distribution on targets and predictions. Finally, we set description_representativeness
to 1000.
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
Tip
Here we reuse metrics
from the train stage for convenience, but they can be adapted to your needs !
2. Model Functioning Explanations#
Model Functioning Explanations are computed with the global_explain
method.
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
We can visualize explanations with XpViz
, using the link in model_explanations.visualisation_link
, if you already
have requested the correct credentials.
3. Inference and their Causal Explanations#
We need a subset of samples to compute Causal Explanations on. Here we filter the test set on the image label, selecting samples with label 1 and 2. It represents 1083 samples.
from xpdeep.filtering.filter import Filter
from xpdeep.filtering.criteria import CategoricalCriterion
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
Explanation can then be computed using the local_explain
method from the Explainer
.
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
👀 Full file preview
"""MNIST workflow, classification, image data."""
from functools import partial
import pyarrow as pa
import pyarrow.parquet as pq
import torch
from datasets import DatasetDict, load_dataset
from model import MnistCNN
from preprocessor import ScaleImage
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torchmetrics.classification import MulticlassAccuracy, MulticlassConfusionMatrix, MulticlassF1Score
from xpdeep import init, set_project
from xpdeep.dataset.parquet_dataset import FittedParquetDataset, ParquetDataset
from xpdeep.dataset.schema.feature.augmentation import ImageFeatureAugmentation
from xpdeep.dataset.schema.feature.feature import ImageFeature
from xpdeep.dataset.upload import upload
from xpdeep.explain.explainer import Explainer
from xpdeep.explain.quality_metrics import Infidelity, Sensitivity
from xpdeep.explain.statistic import DictStats, DistributionStat
from xpdeep.filtering.criteria import CategoricalCriterion
from xpdeep.filtering.filter import Filter
from xpdeep.metric import DictMetrics, TorchGlobalMetric, TorchLeafMetric
from xpdeep.model.model_builder import ModelDecisionGraphParameters
from xpdeep.model.xpdeep_model import XpdeepModel
from xpdeep.model.zoo.cross_entropy_loss_from_proba import CrossEntropyLossFromProbabilities
from xpdeep.model.zoo.mlp import MLP
from xpdeep.project import Project, get_project
from xpdeep.trainer.callbacks import EarlyStopping, ModelCheckpoint, Scheduler
from xpdeep.trainer.trainer import Trainer
from torchvision.transforms import Compose, RandomRotation
def main():
"""Process the dataset, train, and explain the model."""
torch.random.manual_seed(5)
# ##### Prepare the Dataset #######
# 1. Split and Convert your Raw Data
# Load the dataset from HuggingFace datasets hub for convenience.
dataset = load_dataset("mnist", trust_remote_code=True)
test_eval = dataset["test"].train_test_split(test_size=0.5, stratify_by_column="label", seed=1225)
splits = DatasetDict({
"train": dataset["train"],
"val": test_eval["train"], # HuggingFace requires the "train" keyword.
"test": test_eval["test"],
})
# Convert to pyarrow Table format
train_table = pa.Table.from_pandas(splits["train"].to_pandas(), preserve_index=False)
val_table = pa.Table.from_pandas(splits["val"].to_pandas(), preserve_index=False)
test_table = pa.Table.from_pandas(splits["test"].to_pandas(), preserve_index=False)
# Save each split as ".parquet" file
pq.write_table(train_table, "train.parquet")
pq.write_table(val_table, "val.parquet")
pq.write_table(test_table, "test.parquet")
# 2. Upload your Converted Data
directory = upload(
directory_name="mnist_uploaded",
train_set_path="train.parquet",
test_set_path="test.parquet",
val_set_path="val.parquet",
)
# 3. Instantiate a Dataset
train_dataset = ParquetDataset(
split_name="train",
identifier_name="my_local_dataset",
path=directory["train_set_path"],
)
# 4. Find a schema
analyzed_train_dataset = train_dataset.analyze(target_names=["label"])
print(analyzed_train_dataset.analyzed_schema)
# Define augmentation
augmentation = Compose([RandomRotation(90)])
image_rotation_augmentation = ImageFeatureAugmentation(augment_preprocessed=augmentation)
# Correct the image feature type as it cannot be inferred automatically yet.
image_feature = ImageFeature(
preprocessor=ScaleImage(input_size=(28, 28)), feature_augmentation=image_rotation_augmentation, name="image"
)
analyzed_train_dataset.analyzed_schema["image"] = image_feature
print(analyzed_train_dataset.analyzed_schema)
# 5. Fit the schema
fit_train_dataset = analyzed_train_dataset.fit()
fit_test_dataset = FittedParquetDataset(
split_name="test",
identifier_name="my_local_dataset",
path=directory["test_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
fit_val_dataset = FittedParquetDataset(
split_name="val",
identifier_name="my_local_dataset",
path=directory["val_set_path"],
fitted_schema=fit_train_dataset.fitted_schema,
)
# ##### Prepare the Model #######
# 1. Create the required torch models
input_size = fit_train_dataset.fitted_schema.input_size[1:] # 28 x 28
target_size = fit_train_dataset.fitted_schema.target_size[1] # 10
print(f"input_size: {input_size} - target_size: {target_size}")
feature_extraction = MLP(
input_size=128, flatten_input=True, hidden_channels=[128, 64], activation_layer=partial(torch.nn.ReLU)
)
task_learner = MLP(
input_size=64,
activation_layer=partial(torch.nn.ReLU),
flatten_input=True,
hidden_channels=[target_size],
last_activation=partial(torch.nn.Softmax, dim=1),
)
backbone = MnistCNN(output_size=128)
# 2. Explainable Model Specifications
model_specifications = ModelDecisionGraphParameters(
graph_depth=4,
target_homogeneity_pruning_threshold=0.8,
population_pruning_threshold=0.01,
prune_step=20,
target_homogeneity_weight=1.0,
discrimination_weight=0.1,
balancing_weight=0.0,
)
# 3. Create the Explainable Model
xpdeep_model = XpdeepModel.from_torch(
fitted_schema=fit_train_dataset.fitted_schema,
feature_extraction=feature_extraction,
task_learner=task_learner,
backbone=backbone,
decision_graph_parameters=model_specifications,
)
# ##### Train #######
# Metrics to monitor the training.
metrics = DictMetrics(
global_multi_class_accuracy=TorchGlobalMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
leaf_multi_class_accuracy=TorchLeafMetric(
partial(MulticlassAccuracy, num_classes=target_size, average="micro"), target_as_indexes=True
),
global_multi_class_F1_score=TorchGlobalMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
leaf_multi_class_F1_score=TorchLeafMetric(
partial(MulticlassF1Score, num_classes=target_size, average="macro"), target_as_indexes=True
),
global_confusion_matrix=TorchGlobalMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
leaf_confusion_matrix=TorchLeafMetric(
partial(MulticlassConfusionMatrix, num_classes=target_size, normalize="all"), target_as_indexes=True
),
)
callbacks = [
EarlyStopping(monitoring_metric="global_multi_class_accuracy", mode="maximize", patience=10),
Scheduler(
pre_scheduler=partial(ReduceLROnPlateau, mode="min", patience=2),
step_method="epoch",
monitoring_metric="Total loss",
),
ModelCheckpoint(monitoring_metric="global_multi_class_accuracy", mode="maximize"),
]
# Optimizer is a partial object as pytorch needs to give the model as optimizer parameter.
optimizer = partial(torch.optim.AdamW, lr=0.001, foreach=False, fused=False)
trainer = Trainer(
loss=CrossEntropyLossFromProbabilities(reduction="none"),
optimizer=optimizer,
callbacks=callbacks,
start_epoch=0,
max_epochs=30,
metrics=metrics,
)
trained_model = trainer.train(
model=xpdeep_model,
train_set=fit_train_dataset,
validation_set=fit_val_dataset,
batch_size=128,
)
# ##### Explain #######
# 1. Build the Explainer
statistics = DictStats(
distribution_target=DistributionStat(on="target"), distribution_prediction=DistributionStat(on="prediction")
)
quality_metrics = [Sensitivity(), Infidelity()]
explainer = Explainer(
description_representativeness=1000, quality_metrics=quality_metrics, metrics=metrics, statistics=statistics
)
# 2. Model Functioning Explanations
model_explanations = explainer.global_explain(
trained_model,
train_set=fit_train_dataset,
test_set=fit_test_dataset,
validation_set=fit_val_dataset,
)
print(model_explanations.visualisation_link)
# 3. Inference and their Causal Explanations
my_filter = Filter("testing_filter", fit_test_dataset)
my_filter.add_criteria(
CategoricalCriterion(fit_test_dataset.fitted_schema["label"], categories=[1, 2]),
)
causal_explanations = explainer.local_explain(trained_model, fit_test_dataset, my_filter)
print(causal_explanations.visualisation_link)
if __name__ == "__main__":
init(api_key="api_key", api_url="api_url")
set_project(Project.create_or_get(name="MNIST Tutorial"))
try:
main()
finally:
get_project().delete()
We can again visualize causal explanations using the visualisation_link
.