def validate_model(sda,
valid_names,
read_window,
read_algo,
read_rank,
window_size):
valid_reader = ICHISeqDataReader(valid_names)
valid_errors = []
for i in xrange (len(valid_names)):
valid_x, valid_y = valid_reader.read_next_doc(
algo = read_algo,
rank = read_rank,
window = read_window,
divide = False
)
valid_x = valid_x.get_value()
valid_y = valid_y.eval()
n_valid_times = valid_x.shape[0] - window_size + 1
new_valid_x = numpy.array(
[sda.get_da_output(
valid_x[time: time + window_size]
).ravel()
for time in xrange(n_valid_times)]
)
half_window_size = int(window_size/2)
new_valid_y = valid_y[
half_window_size: n_valid_times + half_window_size
]
#compute mean error value for patients in validation set
pat_error = mean_error(
gen_hmm = sda.hmm1,
obs_seq = new_valid_x,
actual_states = new_valid_y
)
valid_errors.append(pat_error)
return numpy.mean(valid_errors)
def test_sda(
sda,
test_names,
read_window,
read_algo,
read_rank,
window_size,
posttrain_rank,
posttrain_algo,
predict_algo='viterbi'):
test_reader = ICHISeqDataReader(test_names)
posttrain_window = sda.da_layers_output_size
index = T.lscalar('index')
test_set_x = T.vector('test_set_x')
test_set_y = T.ivector('test_set_y')
y = T.iscalar('y') # labels, presented as int label
hmm1_error_array = []
hmm2_error_array = []
log_reg_errors = []
test_log_reg = theano.function(
inputs=[
index,
test_set_x,
test_set_y
],
outputs=[sda.logLayer.errors(y), sda.logLayer.predict(), y],
givens={
sda.x: test_set_x[index: index + window_size],
y: test_set_y[index + window_size - 1]
}
)
for test_patient in test_names:
test_set_x, test_set_y = test_reader.read_next_doc(
algo = read_algo,
window = read_window,
rank = read_rank
)
test_set_x = test_set_x.get_value(borrow=True)
test_set_y = test_set_y.eval()
n_test_times = test_set_x.shape[0] - window_size + 1
test_result = [test_log_reg(
index = i,
test_set_x = test_set_x,
test_set_y = test_set_y) for i in xrange(n_test_times)
]
test_result = numpy.asarray(test_result)
test_losses = test_result[:,0]
test_score = float(numpy.mean(test_losses))*100
log_reg_errors.append(test_score)
test_visible_after_sda = numpy.array(
[sda.get_da_output(
test_set_x[time: time+window_size]
).ravel()
for time in xrange(n_test_times)]
)
half_window_size = int(window_size/2)
test_y_after_sda = test_set_y[
half_window_size : n_test_times + half_window_size
]
#compute mean error value for patients in validation set hmm1
pat_error = mean_error(
gen_hmm = sda.hmm1,
obs_seq = test_visible_after_sda,
actual_states = test_y_after_sda
)
hmm1_error_array.append(pat_error)
new_test_visible = create_labels_after_das(
da_output_matrix = test_visible_after_sda,
algo = posttrain_algo,
rank = posttrain_rank,
window = posttrain_window
)
n_patient_samples = len(new_test_visible)
new_test_hidden = test_set_y[half_window_size:n_patient_samples+half_window_size]
patient_error = get_error_on_patient(
model = sda.hmm2,
visible_set = new_test_visible,
hidden_set = new_test_hidden,
algo = predict_algo,
pat = test_patient,
all_labels = True
)
hmm2_error_array.append(patient_error)
print(patient_error, ' error (hmm) for patient ' + test_patient)
print(test_score, ' error (log_reg) for patient ' + test_patient)
gc.collect()
return hmm1_error_array, hmm2_error_array, log_reg_errors
