def test_mutual_information_and_min_nonzero_count():
print('\n' + "*" * 20 + '\ntest_mutual_information_and_min_nonzero_count')
with tf.device('/gpu:0'):
# rand_vec = tf.random_normal([1, 100000000])
# rand_vec_2 = tf.random_normal([1, 100000000])
# max_mutual_info_activations = tf.concat([rand_vec, -rand_vec], 0)
# min_mutual_info_activations = tf.concat([rand_vec, rand_vec_2], 0)
real_activatons = tf.random_normal([10, 16 * 10**6])
activations = real_activatons
value_axis = -1
cross_axis = -2
num_bins = 100
range_ = [-2., 2.]
mutual_info, min_num_events = tensors.mutual_information_and_min_nonzero_count(
activations,
value_axis,
cross_axis,
num_bins,
range_,
max_sample_size_per_iteration=2 * 10**4,
)
hists = tensors.hist_1d_loop(activations, num_bins, [-2., 2.], -1,
10**6)
entropy = tensors.entropy_MM_from_hist(hists, value_axis)
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
run_options = tf.RunOptions(report_tensor_allocations_upon_oom=True)
with tf.Session(config=config) as sess:
entropy, mutual_info, min_num_events = sess.run(
[entropy, mutual_info, min_num_events],
options=run_options,
)
print('entropy:\n', entropy)
print('\nmutual_info:\n', mutual_info)
print('\nmin_num_events:\n', min_num_events)
def test_hist_1d_loop():
print('\n' + "*" * 20 + '\ntest_hist_1d_loop')
values = tf.random_normal([10, 1000, 3])
num_bins = 20
range_ = [-3., 3.]
axis = 1
max_sample_size_per_iteration = 10
hist = tensors.hist_1d_loop(values, num_bins, range_, axis,
max_sample_size_per_iteration)
with tf.Session() as sess:
print(sess.run(hist))
def test_mean_mutual_information_and_min_nonzero_count_4():
print('\n' + "*" * 20 +
'\ntest_mean_mutual_information_and_min_nonzero_count')
distr = [.0, .25, .25, .25, .25]
borders = [[float(i), float(i + 1)] for i in range(5)]
with tf.device('/gpu:0'):
# rand_vec = tf.random_normal([1, 100000000])
# rand_vec_2 = tf.random_normal([1, 100000000])
# max_mutual_info_activations = tf.concat([rand_vec, -rand_vec], 0)
# min_mutual_info_activations = tf.concat([rand_vec, rand_vec_2], 0)
activations = tf.placeholder(tf.float32)
value_axis = -2
cross_axis = -1
num_bins = 5
range_ = [-0., 5.]
mutual_info, mean_mutual_info, min_num_events = tensors.mean_mutual_information_and_min_nonzero_count(
activations,
value_axis,
cross_axis,
num_bins,
range_,
max_sample_size_per_iteration=2 * 10**4,
)
hists = tensors.hist_1d_loop(activations, num_bins, range_, value_axis,
10**6)
entropy = tensors.entropy_MM_from_hist(hists, value_axis)
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
run_options = tf.RunOptions(report_tensor_allocations_upon_oom=True)
with tf.Session(config=config) as sess:
for i in [50, 100, 200, 400, 1000, 2000, 5000, 10000]:
real_activations = 0.5 * np.array([[
tensors.sample_from_distribution_continuous(distr, borders)
for _ in range(i)
] for _ in range(4)] + np.array([[
tensors.sample_from_distribution_continuous(distr, borders)
for _ in range(i)
]] * 4)).transpose()
entropy_r, mutual_info_r, mean_mutual_info_r, min_num_events_r = sess.run(
[entropy, mutual_info, mean_mutual_info, min_num_events],
options=run_options,
feed_dict={activations: real_activations})
print('\n' + '#' * 20)
print("number of points:", i)
print('entropy:\n', entropy_r)
print('\nmutual_info:\n', mutual_info_r)
print('\nmean_mutual_info:\n', mean_mutual_info_r)
print('\nmin_num_events:\n', min_num_events_r)