Beginning Anomaly Detection Using

Encoding 

Stage

Conv 2

Pool 1

Conv 1

Pool 2

Output of Encoding Stage

Input of Decoding Stage

Decoding 

Stage

Output

Conv 1

Upsample 2

Conv 2

Upsample 1

Softmax

Figure 7-46.  In both the encoding and decoding stages, the model is comprised of 

causal convolutional layers and is structured so that the layers are always causal

Chapter 7   temporal Convolutional networks

---

285

The decoding stage is a bit different in this case, since you make use of what is called 

upsampling. Upsampling is a technique in which you repeat the data n number of times 

to scale it up by a factor 

n. In the max pooling layers, the data is reduced by a factor of 

two. So, to upsample and increase the data by the same factor of two, you repeat the 

data twice. In this case, you are using one-dimensional upsampling, so the layer repeats 

each step 

n times with respect to the axis of time. To get a better understanding of what 

upsampling does, let’s apply one-dimensional upsampling to Figure 

7-47

 and Figure 

7- 48

.

Keeping in mind that each individual temporal step is repeated twice, you would see 

something like Figure 

7-49

, Figure 

7-50

, and Figure 

7-51

.

4   2   6   7   1   6   9

x = 

Figure 7-47.  A vector x defined with the corresponding values

n = 2

So data increases by factor of 2 /

repeat each step two mes