def test_categorical_smoothed_layer_eval():
alpha = 0.1
# creating input layer
input_layer = CategoricalSmoothedLayer(vars=vars,
node_dicts=dicts,
alpha=alpha)
# evaluate it
input_layer.eval(obs)
# getting values
layer_evals = input_layer.node_values()
print('layer eval nodes')
print(layer_evals)
# crafting by hand
logvals = []
for node_dict in dicts:
var_id = node_dict['var']
freqs = node_dict['freqs'] if 'freqs' in node_dict else None
logvals.append(compute_smoothed_ll(obs[var_id],
freqs,
vars[var_id],
alpha))
print('log vals')
print(logvals)
assert logvals == layer_evals
# now changing alphas
print('\nCHANGING ALPHAS\n')
alphas = [0., 0.1, 1., 10.]
for alpha_new in alphas:
print('alpha', alpha_new)
input_layer.smooth_probs(alpha_new)
# evaluating again
input_layer.eval(obs)
# getting values
layer_evals = input_layer.node_values()
print('layer evals')
print(layer_evals)
logvals = []
for node_dict in dicts:
var_id = node_dict['var']
freqs = node_dict['freqs'] if 'freqs' in node_dict else None
logvals.append(compute_smoothed_ll(obs[var_id],
freqs,
vars[var_id],
alpha_new))
print('logvals')
print(logvals)
assert_array_almost_equal(logvals, layer_evals)
def test_categorical_smoothed_node_create_and_eval():
for alpha in alphas:
for i, var in enumerate(vars):
var_freq = freqs[i]
smo = CategoricalSmoothedNode(i, var, alpha, var_freq)
smo.eval(obs)
print('smo values')
print(smo.log_val)
ll = compute_smoothed_ll(obs[i], var_freq, var, alpha)
print('log values')
print(ll)
assert_almost_equal(ll, smo.log_val, 15)
def test_categorical_smoothed_node_resmooth():
for i, var in enumerate(vars):
alpha = alphas[0]
var_freq = freqs[i]
smo = CategoricalSmoothedNode(i, var, alpha, var_freq)
smo.eval(obs[i])
print('smo values')
print(smo.log_val)
# checking the right value
ll = compute_smoothed_ll(obs[i], var_freq, var, alpha)
print('log values')
print(ll)
assert_almost_equal(ll, smo.log_val, 15)
# now setting another alpha
print('Changing smooth level')
for alpha_new in alphas:
smo.smooth_probs(alpha_new)
smo.eval(obs[i])
print('smo values')
print(smo.log_val)
ll = compute_smoothed_ll(obs[i], var_freq, var, alpha_new)
print('log values')
print(ll)
assert_almost_equal(ll, smo.log_val, 15)
def test_categorical_input_layer():
print('categorical input layer')
# I could loop through alpha as well
alpha = 0.1
for var_id1 in range(len(vars)):
for var_id2 in range(len(vars)):
for var_val1 in range(vars[var_id1]):
print('varid1, varid2, varval1',
var_id1, var_id2, var_val1)
# var_id1 = 0
# var_val1 = 0
node1 = CategoricalIndicatorNode(var_id1,
var_val1)
# var_id2 = 0
var_vals2 = vars[var_id2]
node2 = CategoricalSmoothedNode(
var_id2, var_vals2, alpha, freqs[var_id2])
# creating the generic input layer
input_layer = CategoricalInputLayer([node1,
node2])
# evaluating according to an observation
input_layer.eval(obs)
layer_evals = input_layer.node_values()
print('layer eval nodes')
print(layer_evals)
# computing evaluation by hand
val1 = 1 if var_val1 == obs[var_id1] or obs[
var_id1] == MARG_IND else 0
logval1 = log(val1) if val1 == 1 else LOG_ZERO
logval2 = compute_smoothed_ll(
obs[var_id2], freqs[var_id2], vars[var_id2], alpha)
logvals = [logval1, logval2]
print('log vals')
print(logvals)
for logval, eval in zip(logvals, layer_evals):
if logval == LOG_ZERO:
# for zero log check this way for correctness
assert IS_LOG_ZERO(eval) is True
else:
assert_almost_equal(logval, eval, PRECISION)