How exactly keras LSTM layer works?

Question

I try to create a sentiment analysis that have 7 classification. Let's say, I have 100.000 unique word (already converted into 100.000 integer) which have the longest input is 41. I created 3 layer embedding, LSTM and Dense Layer.

model = Sequential(); 
model.add(Embedding(100000, 50, input_length = 41))
model.add(LSTM(75, dropout=0.2, recurrent_dropout=0.0)) 
model.add(Dense(7, activation='softmax'))
optimizer = Adam(lr=0.00001)
model.compile(loss='sparse_categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])

Based on code above, I tried to create the illustration model as below :

From my understanding, X_t is 41 integer that will be an input. The embedding layer will convert every integer into real-valued vector of length 50. Every vector that has been converted will be an input of LSTM layer (X₁ - X₄₁)

1. How many LSTM units created based on the code (as my understanding is 41) ?
   Because of return_sequences & return_state by default false for LSTM layer, so it will become many to one ?
   If so, h₄₁ is the only output value that will be used for classification ?
   

Answer 1

Interestingly I searched for this answer for a long time and then I got an answer. It is actually one LSTM unit that is repeated by the length of the sentence or collection of records. This was excellently explained here

https://zhuanlan.zhihu.com/p/58854907'

64 is the number of dense units inside - Isn't that interesting?