markusthill / bioma-tcn-ae Goto Github PK

Conv1DTranspose layer of Keras shall be used instead of UpSampling1D since it only does upsampling, not deconvolution.

Place in the code:
https://github.com/MarkusThill/bioma-tcn-ae/blob/main/src/tcnae.py#L127

Conv1DTranspose:
https://keras.io/api/layers/convolution_layers/convolution1d_transpose/

As for the Feature map reduction mentioned in the article[1], I did not find it in your code or keras_tcn. So I have a problem with adding this layer, which is where to add it:

About whether "Feature map reduction" should be added before or after Residual Connections in TCN, like:

out = self.net(x)  # net is TCN

res = x if self.downsample is None else self.downsample(x)

return self.relu(out + res), out

Should I go for reduction ‘out’ or ‘self.relu(out + res)’？
Looking forward to your response!

[1]Thill M, Konen W, Wang H, et al. Temporal convolutional autoencoder for unsupervised anomaly detection in time series[J]. Applied Soft Computing, 2021, 112: 107751.

Hi, I want to test TCN-AE on my own dataset, and there is a problem, could you help me see how to solve it?

My dataset shape is (1000,300,300), I changed the model input parameter ts_dimension to 300, and run model.fit but it got an error ValueError: Dimensions must be equal, but are 294 and 300 for 'loss/dense_13_loss/sub' (op: 'Sub') with input shapes: [?,294,300], [?,300,300].

I checked your dataset shape and it was (19791, 1050, 1), so when I changed the data shape to (1000,1050,300) it can run well. I guess there's something wrong with the convolution layer, maybe the latent_sample_rate? But I haven't found a solution so far. Could you help me, thank you very much!

Using padding=same for time series data will refer to future data.
Can you elaborate on this part?
Thank you very much for your answer.

Hi, I am studying your research and have a few questions regarding your code.

In Section 2.6.3, "Utilizing hidden representations for the anomaly detection task," you took the blue bar from Fig. 3 and apply a 1 x 1 convolution layer to reduce the channel size from 16 (as shown in the figure) to 1.

However, from the provided explanation, it seems the 1 x 1 convolution layer is not included in the training loop. I am curious if using an untrained convolution layer to reduce the feature dimension is feasible or if it might negatively impact performance.

Additionally, in Section 2.6.4, "Feature map reduction," you place 1 x 1 convolutions after each dilated convolutional layer to reduce the feature map dimension. Based on Issue #3, my understanding is that these layers are optionally included in the training loop. Typically, adding additional layers (such as 1 x 1 convolutions) would increase training time and the number of parameters.

Could you please explain how adding these additional layers (1 x 1 convolutions) helps in reducing trainable parameters and training time?

Could you also provide the code for this part?

Thank you for your help :)