lovelytt0
commented
1 year ago We trained our model with 3-component seismograms (X; dimension: 36000) to extract the latent representations (V; dimension: 3264). The training and test sets are processed in the same format. The model will not run correctly if your input is a different sequence length (like 318000), because the encoder we use is from EQTransformer. The output dimension will scale with the input sequence, it will not be 3264 anymore.
The performance of the model is not guaranteed due to the different physical meanings of input sequences (like, sampling rate, and duration of seismograms).
If you want to use our model without training and fine-tuning, I suggest you resample your waveform first, then use a sliding window to get the same format of inputs.
If you want to change the input and output dimensions, you might need to add several pooling layers to the encoder. Also, training is necessary.
Dear Dr.Feng,
I would like to ask you a question. In your paper, you mentioned that the encoder encodes 3-component seismograms (X; dimension: 36000) into the latent representations (V; dimension: 3264) , I would like to know how you converted, is there any basis for such transformation, because I input three-component seismic map (X :3 18000), should my latent representations also be converted to 3264?