I'm using an
LSTM to achieve a classification problem. I have a dataset composed by sentences, each sentence is composed by a variable number of words and I have to predict a label for each word of each sentence.
For example, I have a dataset of shape
(300000, 800), so 300000 words and each word is made of 800 features (word embeddings, etc...).
Moreover, each sentence is divided in subtrees, which each subtree is a syntactic dependency tree between words. My task is to learn the dependencies between words of a subtree.
I thought to train the network using
timesteps=1 and the
train_on_batch function, where the batch size is variable and equal to the subtree cardinality.
Is it a reasonable process?