why does masking requires us to sample (take only 15%) of the words?
I think this guarantees that only about 1/7 of the words are masked, which is just like the window size in word2vec. That is on average in BERT we use 7 words as context to predict one word. The more words we mask the smaller the "window size" and the smaller the context.
Ane model would be to estimate the probability of getting a certain diagnosis (per time frame) for each of the categories of interest. Then just draw samples from these distribution to generate diagnosises.
A simple version could take just agegroup, gender as basis for categories. More advanced versions could include past medical history.
A high cross entropy and a high accuracy are not necessarily inconsistent results.
Here's a simple example. Suppose you have a problem with 10 classes, and a data set with 1 sample of each class. In 9 of the samples, the correct class has prediction $0.85$, but in one class it has prediction $0.001$. The cross-entropy loss in this case is almost $0.84$. This ...
In keras.layers.LSTM(units, activation='tanh', ....), the units refers to the dimensionality or length of the hidden state or the length of the activation vector passed on the next LSTM cell/unit - the next LSTM cell/unit is the "green picture above with the gates etc from http://colah.github.io/posts/2015-08-Understanding-LSTMs/
The next LSTM cell/unit ...
In addition to the tools mentioned by @BCjuan, there is also.
a platform independent inference engine for neural networks. It takes a Keras model as input.
X-CUBE-AI from ST Microelectronics. It is however proprietary and only supports their devices.