def train_models(model, x, y):
"""
Train a Gaussian HMM for each class in the input matrix.
For this, a default of 7 states is used. Hence the size of startprob and transmat.
As this is a Left-Right HMM, the start probability is 100% for state 1, and 0 for all others.
Args:
model: HMM class object
x: input image feature vector
y: labels for images
Return:
Trained Model
"""
classes = len(np.unique(y))
if classes < 1:
raise ValueError("Need at least 1 class to train HMMs with.")
startprob = np.array([1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0])
transmat = np.array([[0.8, 0.2, 0.0, 0.0, 0.0, 0.0, 0.0],
[0.0, 0.8, 0.2, 0.0, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.8, 0.2, 0.0, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.8, 0.2, 0.0, 0.0],
[0.0, 0.0, 0.0, 0.0, 0.8, 0.2, 0.0],
[0.0, 0.0, 0.0, 0.0, 0.0, 0.8, 0.2],
[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.1]])
for image_class in range(classes):
# Partition the training data into arrays for each class
# Then train a Hidden Markov Model for each
covar = "diag" if model.diag else "full"
hmm = GaussianHMM(model.states,
covar,
n_iter=20,
tol=0.00001,
random_state=np.random.RandomState())
hmm.startprob_ = startprob
hmm.transmat_ = transmat
# Loop iterator corresponds to label of data
hmm.label = image_class
features = functions.get_labelled_train_data(x, y, image_class)
# Fit the HMM for that label
old_block_num = features.shape[1]
features_size = features.shape[2]
length = features.shape[0]
# Need to flatten out the first 2 dimensions of feature vector
# and pass an array of length into the HMM
# old_block_num = number of observation blocks for a single image.
lengths = []
for i in range(length):
lengths.append(old_block_num)
feat = features.reshape((length * old_block_num, features_size))
hmm.fit(feat, lengths=lengths) # Train the model for 1 class.
model.hmms.append(hmm)
return model
