@JonathanHuangC Good question. It comes from this repo. I've transferred your question.

laiguokun/multivariate-time-series-data#7

from keras-tcn.

@philipperemy
Is there any example of using TCN to extract features?
For example, compressing the exchange rate of [1000 * 8] into [1 * 8], where 1000 represents the day and 8 is the feature of the data.
Thank you very much for your reply.

from keras-tcn.

You mean some form of auto encoding? You can just search for LSTM features extraction and you swap the LSTM class with the TCN class and it should work.

from keras-tcn.

You mean some form of auto encoding? -> Yes, it seems to be using TCN to implement the auto encoder.

Do you mean that it is enough to change LSTM to TCN? Sorry for there are a lot of questions.

Thank you again for your reply.

LSTM autoencoder

define model

model = Sequential()

Encoder step

model.add(LSTM(15, input_shape=(X_train.shape[1], X_train.shape[2]), activation='relu'))
model.add(RepeatVector(X_train.shape[1]))

Decoder step

model.add(LSTM(15, activation='relu', return_sequences=True))
model.add(TimeDistributed(Dense(X_train.shape[2])))

model.compile(optimizer='adam', loss='mse')

history = model.fit(X_train, X_train, epochs=_epochs, batch_size = _batch_size,
validation_split=_validation_split, callbacks=callback)

TCN autoencoder

define model

model = Sequential()

Encoder step

model.add(TCN(15, input_shape=(X_train.shape[1], X_train.shape[2]), activation='relu'))
model.add(RepeatVector(X_train.shape[1]))

Decoder step

model.add(TCN(15, activation='relu', return_sequences=True))
model.add(TimeDistributed(Dense(X_train.shape[2])))

model.compile(optimizer='adam', loss='mse')

history = model.fit(X_train, X_train, epochs=_epochs, batch_size = _batch_size,
validation_split=_validation_split, callbacks=callback)

from keras-tcn.

yeah it's as easy as swapping the class.

from keras-tcn.