the relationship between receptive field and the input sequence timestep
Kitaharakasusa opened this issue · 2 comments
I'm confused about this problem, does the receptive field have to be larger than the input timestep? if my input is [batch_size, 1, 6]
can the TCN work?
@Kitaharakasusa for a given sequence, the prediction at time t will only see from t-receptive_field to t.
So if you have a sequence of 100 and a receptive field of 20, the TCN will predict 100 values where each one depends on the last 20 values. That means the predict at index 48 for example will only see from 27 up to 47.
If you have return_sequences=True, then only the last value at 99 will be returned. In that case yes you will need to have a receptive field greater than 100 if you want to include all the sequence for predicting the value at index 99.
I'll close this question. If it's not clear, please feel free to re-open it. Thank you!