Visualize LSTM for time series sequential data I am trying to visualize LSTM that is applied to sequential data available here: https://its-ml.de/index.php/plattscaling/
There are 25 features with 50 sequence length.
LSTM layer 1 has 100 cells, LSTM layer 2 has 50 cells with droput as 0.2.
The output/dense layer has 1 cell which contains a single value.
This can be visualized as shown in figure below:

I am trying to connect the features with respective LSTM layers by referring Fig. 2 as as shown here: https://towardsdatascience.com/how-to-reshape-data-and-do-regression-for-time-series-using-lstm-133dad96cd00
However, from cell-51 in LSTM layer-1, there is no input coming from the feature because the sequence length itself is 50.
What is the input for cell-51 in LSTM layer-1?
Also, how is LSTM1, LSTM2 and output layers connected?
How do we represent the dropout as 0.2 in this figure?
What will be the dimensions of hidden and cell state in Layer1 (H,C) and layer2 (h,c) ?
Can somebody please help me out in visualizing this network?
 A: It seems you're confusing "number of output neurons" in the 1st linked webpage with "number of LSTM cells". The former refers to the number of the hidden units, i.e. the dimension of the LSTM cell output vector. It is also known by other names such as the size of the hidden layer, the hidden size, the hidden dimension, etc. An LSTM only has a single cell that consumes a sequential input step by step. This computation is often unrolled when visualised, which is why it looks as if the LSTM has a fixed number of cells; it doesn't. So in your figure, there shouldn't be cell 51, 52, and so on. In fact, the cells shouldn't be numbered at all as it gives the impression of multiple cells. See the 2nd figure by Christopher Olah depicting an unrolled RNN for reference (LSTMs are a special kind of RNNs). This explanation also answers your question about the dimension of the hidden and cell state: they're both equal to the number of hidden units. As for visualising the dropout layer, there isn't a standard way. You're allowed to just draw a box labelled "Dropout 0.2" similar to that on the 1st linked webpage.
If the LSTM has multiple layers, the output  $\mathbf{h}^{(\ell)}_t$ at timestep $t$ of the $\ell$-th layer becomes the input $\mathbf{x}^{(\ell+1)}_t$ of the same timestep in the next layer.
