Classifier

CNNClassifierTorch

Time Convolutional Neural Network (CNN) in PyTorch, as described in [1].

Zhao et al. 2017 uses MSE loss and sigmoid activation in all layers. To obtain same behaviour as Zhao et al. 2017, set activation and criterion to “sigmoid” and “MSELoss” respectively.

Adapted from the implementation from Fawaz et. al https://github.com/hfawaz/dl-4-tsc/blob/master/classifiers/cnn.py

Schnellstart

python

from sktime.classification.deep_learning.cnn import CNNClassifierTorch

estimator = CNNClassifierTorch(num_epochs: int=2000, batch_size: int=16, kernel_sizes: tuple [int, ... ]=(7, 7), avg_pool_size: int=3, filter_sizes: tuple [int, ... ]=(6, 12), padding: str='auto', use_bias: bool=True, activation: str | Callable | None=None, activation_hidden: str | Callable='Sigmoid', optimizer: str | None | Callable='Adam', optimizer_kwargs: dict | None=None, criterion: str | None | Callable='CrossEntropyLoss', criterion_kwargs: dict | None=None, callbacks: None | str | tuple [str, ... ]='ReduceLROnPlateau', callback_kwargs: dict | None=None, lr: float=0.01, verbose: bool=False, init_weights: str | None=None, random_state: int | None=None)

Parameter(19)

num_epochsint, default = 2000

Number of epochs to train the model.

batch_sizeint, default = 16

Size of each mini-batch.

kernel_sizestuple of int, default = (7, 7)

A tuple of length equal to the number of conv layers with each entry in the tuple specifies the kernel size for the corresponding convolutional layer. The length of kernel_sizes must be equal to the length of filter_sizes.

avg_pool_sizeint, default = 3

Size of the average pooling window.

filter_sizestuple of int, default = (6, 12)

A tuple of length equal to the number of conv layers with each entry in the tuple specifies the filter size for the corresponding convolutional layer. The length of filter_sizes must be equal to the length of kernel_sizes.

paddingstr, default = “auto”

Padding for conv layers. “auto”: “same” if series_length < 60 else “valid”; “valid” or “same” otherwise.

use_biasbool, default = True

Whether to use bias in output layer.

activationstr, Callable, or None, default=None

Activation applied to the output layer.

Permitted values:

None: no activation is applied to the output layer and the network returns raw outputs (logits). This is typically required when using CrossEntropyLoss, which expects logits as input.
str: name of a class in torch.nn. Case-sensitive names are recommended and must match PyTorch (e.g., "ReLU", "LeakyReLU"). Lowercase aliases for common activations are also accepted (e.g., "relu" is resolved to "ReLU"). The class is instantiated with default constructor arguments. Must be a valid torch.nn activation; see https://pytorch.org/docs/stable/nn.html#non-linear-activations-weighted-sum-nonlinearity
torch.nn.Module: an instance of a torch.nn.Module subclass, for example torch.nn.ReLU(). Arbitrary callables are not supported.

activation_hiddenstr, Callable, or None, default=”Sigmoid”

Activation applied to the hidden layers.

Permitted values:

None: no activation is applied to the hidden layers.
str: name of a class in torch.nn. Case-sensitive names are recommended and must match PyTorch (e.g., "ReLU", "LeakyReLU"). Lowercase aliases for common activations are also accepted (e.g., "relu" is resolved to "ReLU"). The class is instantiated with default constructor arguments. Must be a valid torch.nn activation; see https://pytorch.org/docs/stable/nn.html#non-linear-activations-weighted-sum-nonlinearity
torch.nn.Module: an instance of a torch.nn.Module subclass, for example torch.nn.ReLU(). Arbitrary callables are not supported.

Recommended activations: Sigmoid, ReLU, Tanh, Softmax, LogSoftmax.

optimizerstr or callable, default = “Adam”

Optimizer to use. Same as TF default (Adam).

optimizer_kwargsdict or None, default = None

Additional keyword arguments for the optimizer.

criterionstr or callable, default = “CrossEntropyLoss”

Loss function for training.

criterion_kwargsdict or None, default = None

Additional keyword arguments for the criterion.

callbacksNone or str or tuple of str, default = “ReduceLROnPlateau”

Learning rate schedulers as callbacks.

callback_kwargsdict or None, default = None

Keyword arguments for callbacks.

lrfloat, default = 0.01

Learning rate (TF CNN uses Adam(lr=0.01)).

verbosebool, default = False

Whether to print progress during training.

init_weights: str or None, default = None

The method to initialize the weights of the conv layers. Supported values are ‘kaiming_uniform’, ‘kaiming_normal’, ‘xavier_uniform’, ‘xavier_normal’, or None for default PyTorch initialization.

random_stateint or None, default = None

Seed for reproducibility.

Beispiele

>>> from sktime.classification.deep_learning.cnn import CNNClassifierTorch
>>> from sktime.datasets import load_unit_test
>>> X_train, y_train = load_unit_test (split = "train")
>>> X_test, y_test = load_unit_test (split = "test")
>>> clf = CNNClassifierTorch (num_epochs = 20, batch_size = 4)
>>> clf. fit (X_train, y_train) CNNClassifierTorch(
... )

Referenzen

[1]

Zhao et al. Convolutional neural networks for time series classification, Journal of Systems Engineering and Electronics, 28(1):2017.