def test_binarize(self) -> None:
x = int64factory.tensor(tf.constant([
2**62 + 3,
2**63 - 1,
2**63 - 2,
-3
], shape=[2, 2], dtype=tf.int64))
y = x.to_bits()
expected = np.array([
[1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0],
[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0],
[0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0],
[1, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
]).reshape([2, 2, 64])
with tf.Session() as sess:
actual = sess.run(y.to_native())
np.testing.assert_array_equal(actual, expected)
def test_forward(self) -> None:
# input
batch_size, channels_in, channels_out = 32, 3, 64
img_height, img_width = 28, 28
input_shape = (batch_size, channels_in, img_height, img_width)
input_conv = np.random.normal(size=input_shape).astype(np.int64)
# filters
h_filter, w_filter, strides = 2, 2, 2
filter_shape = (h_filter, w_filter, channels_in, channels_out)
filter_values = np.random.normal(size=filter_shape).astype(np.int64)
x_in = int64factory.tensor(input_conv)
out = x_in.conv2d(int64factory.tensor(filter_values), strides)
with tf.Session() as sess:
actual = sess.run(out.to_native())
# reset graph
tf.reset_default_graph()
# convolution tensorflow
with tf.Session() as sess:
# conv input
x = tf.Variable(input_conv, dtype=tf.float32)
x_nhwc = tf.transpose(x, (0, 2, 3, 1))
# convolution Tensorflow
filters_tf = tf.Variable(filter_values, dtype=tf.float32)
conv_out_tf = tf.nn.conv2d(
x_nhwc,
filters_tf,
strides=[1, strides, strides, 1],
padding="SAME",
)
sess.run(tf.global_variables_initializer())
out_tensorflow = sess.run(conv_out_tf).transpose(0, 3, 1, 2)
np.testing.assert_array_almost_equal(actual, out_tensorflow, decimal=3)
def test_random_binarize(self) -> None:
input = np.random.uniform(low=2**63 + 1, high=2**63 - 1, size=2000).astype(np.int64).tolist()
x = int64factory.tensor(tf.constant(input, dtype=tf.int64))
y = x.to_bits()
with tf.Session() as sess:
actual = sess.run(y.to_native())
j = 0
for i in input:
if i < 0:
binary = bin(((1 << 64) - 1) & i)[2:][::-1]
else:
binary = bin(i)
binary = binary[2:].zfill(64)[::-1]
bin_list = np.array(list(binary)).astype(np.int64)
np.testing.assert_equal(actual[j], bin_list)
j += 1