Autoencoder with variable length data input

TensorFlow
,
1

Hi all,
I’m trying to create a model for anomaly detection like those described in the link below:

which makes use of ECG5000 dataset.

In the above-mentioned example, they make use of a dense layer autoencoder; they train the model with only normal ECG curves; each ECG curvers are made of 140 samples.
In my case, the dataset is made of a time series in which each curve has a different number of samples (they could be 140, 145, 160, 130 and so on).

In your opinion, what could be the best approach to face this problem?
As a possible solution, I thought of padding each curve at the maximum length, but the problem is that I don’t know in advance what could be the maximum duration.

thank for your suggestion,
regards

3

Hi @savino_giusto,

You can try in the reverse scenario. Fix a minimum length, for example 140 and if its less than that do padding else prune the extra length and then try to train the model.

Thanks.

5

Thanks for your reply. I’ll try your suggest. Padding and prouning should be a good advice.

1 Like