[QUESTION] Best practice for initializing HMM language model?

[ivnle]

Thanks for putting together this great package! I'm trying wrap my head around Pomegranate by training an HMM language model. Based on what i've read in the documentation https://pomegranate.readthedocs.io/en/latest/tutorials/B_Model_Tutorial_4_Hidden_Markov_Models.html and https://pomegranate.readthedocs.io/en/latest/api.html, I've put together the following implementation: https://colab.research.google.com/drive/1wbG2mvQEpqaMm9poxclIq15_xg5e_wrq?usp=sharing

Can you confirm if this is the correct way to initialize an HMM where all parameters (emission, transition, initial state) are updated?

# 1. each hidden state is a categorical distribution over the vocabulary
ds = [] # [h, v]
for _ in range(n_hidden_states):
    d = torch.ones(1, vocab_size)
    noise = torch.randn_like(d) * std
    d = d + noise
    d = (d / d.sum()) # [v]
    ds.append(Categorical(d)) 
# 2. define transition matrix
edges = torch.ones(n_hidden_states, n_hidden_states)
noise = torch.randn_like(edges) * std
edges = edges + noise
edges = edges / edges.sum(dim=1, keepdim=True) # [h, h]
edges = edges
# 3. initial states
starts = torch.ones(n_hidden_states)
noise = torch.randn_like(starts) * std
starts = starts + noise
starts = starts / starts.sum() # [h]
starts = starts
model = DenseHMM(distributions=ds, edges=edges, starts=starts, verbose=True, tol=0.00001, max_iter=10)