Skip to main content

Practice

In order to utilise TiRex classification, make sure to either locally install it in your preferred Python environment or use a hosted Jupyter Notebook service like Google Colab.

1. Install Tirex​

# install with the extra 'classification' for classification support
pip install 'tirex-ts[classification]'

2. Import TiRex and supporting libraries​

Install additional packages (used only for example data).

pip install aeon

Either in a Jupyter Notebook, or in a local Python file in your Python Environment:

# General imports
import torch
from aeon.datasets import load_italy_power_demand
from sklearn.preprocessing import LabelEncoder

# Import TiRex classifiers
from tirex.models.classification import TirexRFClassifier, TirexClassifierTorch

3. Preprocessing of the Data​

# Load train data
train_X, train_y = load_italy_power_demand(split="train")

# Load test data
test_X, test_y = load_italy_power_demand(split="test")

# Encode string labels -> integers
label_encoder = LabelEncoder()
train_y = label_encoder.fit_transform(train_y)
test_y = label_encoder.transform(test_y)

# Convert to torch tensors
train_X = torch.tensor(train_X, dtype=torch.float32)
test_X = torch.tensor(test_X, dtype=torch.float32)

train_y = torch.tensor(train_y, dtype=torch.long)
test_y = torch.tensor(test_y, dtype=torch.long)

4. Initialize TiRex Classifier​

Model weights of the TiRex backbone model are automatically fetched from HuggingFace.

Option A: Random Forest Classifier​

classifier = TirexRFClassifier(data_augmentation=False, device="cuda:0", n_estimators=50, max_depth=10, random_state=42)
  • data_augmentation (bool): Whether to use additional data augmentation concatenated to the embeddings. Defaults to False.
  • device (str): Device used for embedding computation (for example, "cuda:0" for GPU or "cpu"). Note: Random Forest itself always runs on CPU (uses scikit-learn).
  • The rest of the parameters are kwargs to RandomForest of sklearn. For more details see scikit-learn RandomForestClassifier documentation.

Option B: Torch Linear Classifier​

classifier = TirexClassifierTorch(
    data_augmentation=False,
    device="cuda:0"
)
  • data_augmentation (bool): Whether to use additional data augmentation concatenated to the embeddings. Defaults to False.
  • device (str): Device used for training and inference (for example, "cuda:0" for GPU or "cpu").

The rest of the parameters you can see in the API description.

5. Fit the model to the data​

classifier.fit((train_X, train_y))

For Linear classifier, you have the possibility to provide a validation set.

classifier.fit(train_data=(train_X, train_y),val_data=(val_X, val_y))

If you don't provide it, the train dataset will be split into train and validation datasets internally. Two parameters are responsible for the train/validation split by model initialization:

  • val_split_ratio (float): is responsible for the size of the validation set (between 0 and 1). Defaults to: 0.2
  • stratify (bool): whether to stratify the train/validation split by class labels. Default: True

6. Prediction results​

Analyze your prediction results

from sklearn.metrics import accuracy_score, f1_score

pred_y = classifier.predict(test_X)
# Convert to numpy if needed (predict returns torch.Tensor)

pred_y = pred_y.cpu().numpy()
test_y_np = test_y.cpu().numpy()

print("Accuracy: " + str(accuracy_score(test_y_np, pred_y)))
print("F1 Score: " + str(f1_score(test_y_np, pred_y, average="macro")))