def test_quantized_relu(bits, integer, use_sigmoid, test_values,
expected_values):
"""Test quantized_relu function."""
x = K.placeholder(ndim=2)
f = K.function([x], [quantized_relu(bits, integer, use_sigmoid)(x)])
result = f([test_values])[0]
assert_allclose(result, expected_values, rtol=1e-05)
def test_sequential_qnetwork():
model = tf.keras.Sequential()
model.add(Input((28, 28, 1), name='input'))
model.add(
QConv2D(32, (2, 2),
strides=(2, 2),
kernel_quantizer=quantized_bits(4, 0, 1),
bias_quantizer=quantized_bits(4, 0, 1),
name='conv2d_0_m'))
model.add(QActivation(quantized_relu(4, 0), name='act0_m'))
model.add(
QConv2D(64, (3, 3),
strides=(2, 2),
kernel_quantizer=quantized_bits(4, 0, 1),
bias_quantizer=quantized_bits(4, 0, 1),
name='conv2d_1_m'))
model.add(QActivation(quantized_relu(4, 0), name='act1_m'))
model.add(
QConv2D(64, (2, 2),
strides=(2, 2),
kernel_quantizer=quantized_bits(4, 0, 1),
bias_quantizer=quantized_bits(4, 0, 1),
name='conv2d_2_m'))
model.add(QActivation(quantized_relu(4, 0), name='act2_m'))
model.add(Flatten())
model.add(
QDense(10,
kernel_quantizer=quantized_bits(4, 0, 1),
bias_quantizer=quantized_bits(4, 0, 1),
name='dense'))
model.add(Activation('softmax', name='softmax'))
# Check that all model operation were found correctly
model_ops = extract_model_operations(model)
for layer in model_ops.keys():
assert model_ops[layer]['type'][0] != 'null'
return model
def main():
# check the mean value of samples from stochastic_rounding for po2
np.random.seed(42)
count = 100000
val = 42
a = K.constant([val] * count)
b = quantized_po2(use_stochastic_rounding=True)(a)
res = np.sum(K.eval(b)) / count
print(res, "should be close to ", val)
b = quantized_relu_po2(use_stochastic_rounding=True)(a)
res = np.sum(K.eval(b)) / count
print(res, "should be close to ", val)
a = K.constant([-1] * count)
b = quantized_relu_po2(use_stochastic_rounding=True)(a)
res = np.sum(K.eval(b)) / count
print(res, "should be all ", 0)
# non-stochastic rounding quantizer.
a = K.constant([-3.0, -2.0, -1.0, -0.5, 0.0, 0.5, 1.0, 2.0, 3.0])
a = K.constant([0.194336])
print(" a =", K.eval(a).astype(np.float16))
print("qa =", K.eval(quantized_relu(6,2)(a)).astype(np.float16))
print("ss =", K.eval(smooth_sigmoid(a)).astype(np.float16))
print("hs =", K.eval(hard_sigmoid(a)).astype(np.float16))
print("ht =", K.eval(hard_tanh(a)).astype(np.float16))
print("st =", K.eval(smooth_tanh(a)).astype(np.float16))
c = K.constant(np.arange(-1.5, 1.51, 0.3))
print(" c =", K.eval(c).astype(np.float16))
print("qb_111 =", K.eval(quantized_bits(1,1,1)(c)).astype(np.float16))
print("qb_210 =", K.eval(quantized_bits(2,1,0)(c)).astype(np.float16))
print("qb_211 =", K.eval(quantized_bits(2,1,1)(c)).astype(np.float16))
print("qb_300 =", K.eval(quantized_bits(3,0,0)(c)).astype(np.float16))
print("qb_301 =", K.eval(quantized_bits(3,0,1)(c)).astype(np.float16))
c_1000 = K.constant(np.array([list(K.eval(c))] * 1000))
b = np.sum(K.eval(bernoulli()(c_1000)).astype(np.int32), axis=0) / 1000.0
print(" hs =", K.eval(hard_sigmoid(c)).astype(np.float16))
print(" b_all =", b.astype(np.float16))
T = 0.0
t = K.eval(stochastic_ternary(alpha="auto")(c_1000))
for i in range(10):
print("stochastic_ternary({}) =".format(i), t[i])
print(" st_all =", np.round(
np.sum(t.astype(np.float32), axis=0).astype(np.float16) /
1000.0, 2).astype(np.float16))
print(" ternary =", K.eval(ternary(threshold=0.5)(c)).astype(np.int32))
c = K.constant(np.arange(-1.5, 1.51, 0.3))
print(" c =", K.eval(c).astype(np.float16))
print(" b_10 =", K.eval(binary(1)(c)).astype(np.float16))
print("qr_10 =", K.eval(quantized_relu(1,0)(c)).astype(np.float16))
print("qr_11 =", K.eval(quantized_relu(1,1)(c)).astype(np.float16))
print("qr_20 =", K.eval(quantized_relu(2,0)(c)).astype(np.float16))
print("qr_21 =", K.eval(quantized_relu(2,1)(c)).astype(np.float16))
print("qr_101 =", K.eval(quantized_relu(1,0,1)(c)).astype(np.float16))
print("qr_111 =", K.eval(quantized_relu(1,1,1)(c)).astype(np.float16))
print("qr_201 =", K.eval(quantized_relu(2,0,1)(c)).astype(np.float16))
print("qr_211 =", K.eval(quantized_relu(2,1,1)(c)).astype(np.float16))
print("qt_200 =", K.eval(quantized_tanh(2,0)(c)).astype(np.float16))
print("qt_210 =", K.eval(quantized_tanh(2,1)(c)).astype(np.float16))
print("qt_201 =", K.eval(quantized_tanh(2,0,1)(c)).astype(np.float16))
print("qt_211 =", K.eval(quantized_tanh(2,1,1)(c)).astype(np.float16))
set_internal_sigmoid("smooth"); print("with smooth sigmoid")
print("qr_101 =", K.eval(quantized_relu(1,0,1)(c)).astype(np.float16))
print("qr_111 =", K.eval(quantized_relu(1,1,1)(c)).astype(np.float16))
print("qr_201 =", K.eval(quantized_relu(2,0,1)(c)).astype(np.float16))
print("qr_211 =", K.eval(quantized_relu(2,1,1)(c)).astype(np.float16))
print("qt_200 =", K.eval(quantized_tanh(2,0)(c)).astype(np.float16))
print("qt_210 =", K.eval(quantized_tanh(2,1)(c)).astype(np.float16))
print("qt_201 =", K.eval(quantized_tanh(2,0,1)(c)).astype(np.float16))
print("qt_211 =", K.eval(quantized_tanh(2,1,1)(c)).astype(np.float16))
set_internal_sigmoid("real"); print("with real sigmoid")
print("qr_101 =", K.eval(quantized_relu(1,0,1)(c)).astype(np.float16))
print("qr_111 =", K.eval(quantized_relu(1,1,1)(c)).astype(np.float16))
print("qr_201 =", K.eval(quantized_relu(2,0,1)(c)).astype(np.float16))
print("qr_211 =", K.eval(quantized_relu(2,1,1)(c)).astype(np.float16))
print("qt_200 =", K.eval(quantized_tanh(2,0)(c)).astype(np.float16))
print("qt_210 =", K.eval(quantized_tanh(2,1)(c)).astype(np.float16))
print("qt_201 =", K.eval(quantized_tanh(2,0,1)(c)).astype(np.float16))
print("qt_211 =", K.eval(quantized_tanh(2,1,1)(c)).astype(np.float16))
set_internal_sigmoid("hard")
print(" c =", K.eval(c).astype(np.float16))
print("q2_31 =", K.eval(quantized_po2(3,1)(c)).astype(np.float16))
print("q2_32 =", K.eval(quantized_po2(3,2)(c)).astype(np.float16))
print("qr2_21 =", K.eval(quantized_relu_po2(2,1)(c)).astype(np.float16))
print("qr2_22 =", K.eval(quantized_relu_po2(2,2)(c)).astype(np.float16))
print("qr2_44 =", K.eval(quantized_relu_po2(4,1)(c)).astype(np.float16))
# stochastic rounding
c = K.constant(np.arange(-1.5, 1.51, 0.3))
print("q2_32_2 =", K.eval(quantized_relu_po2(32,2)(c)).astype(np.float16))
b = K.eval(stochastic_binary()(c_1000)).astype(np.int32)
for i in range(5):
print("sbinary({}) =".format(i), b[i])
print("sbinary =", np.round(np.sum(b, axis=0) / 1000.0, 2).astype(np.float16))
print(" binary =", K.eval(binary()(c)).astype(np.int32))
print(" c =", K.eval(c).astype(np.float16))
for i in range(10):
print(" s_bin({}) =".format(i),
K.eval(binary(use_stochastic_rounding=1)(c)).astype(np.int32))
for i in range(10):
print(" s_po2({}) =".format(i),
K.eval(quantized_po2(use_stochastic_rounding=1)(c)).astype(np.int32))
for i in range(10):
print(
" s_relu_po2({}) =".format(i),
K.eval(quantized_relu_po2(use_stochastic_rounding=1)(c)).astype(
np.int32))
def test_qnetwork():
x = x_in = Input((28, 28, 1), name='input')
x = QSeparableConv2D(
32, (2, 2),
strides=(2, 2),
depthwise_quantizer="binary",
pointwise_quantizer=quantized_bits(4, 0, 1),
depthwise_activation=quantized_bits(6, 2, 1),
bias_quantizer=quantized_bits(4, 0, 1),
name='conv2d_0_m')(
x)
x = QActivation('quantized_relu(6,2,1)', name='act0_m')(x)
x = QConv2D(
64, (3, 3),
strides=(2, 2),
kernel_quantizer="ternary",
bias_quantizer=quantized_bits(4, 0, 1),
name='conv2d_1_m',
activation=quantized_relu(6, 3, 1))(
x)
x = QConv2D(
64, (2, 2),
strides=(2, 2),
kernel_quantizer=quantized_bits(6, 2, 1),
bias_quantizer=quantized_bits(4, 0, 1),
name='conv2d_2_m')(
x)
x = QActivation('quantized_relu(6,4,1)', name='act2_m')(x)
x = Flatten(name='flatten')(x)
x = QDense(
10,
kernel_quantizer=quantized_bits(6, 2, 1),
bias_quantizer=quantized_bits(4, 0, 1),
name='dense')(
x)
x = Activation('softmax', name='softmax')(x)
model = Model(inputs=[x_in], outputs=[x])
# reload the model to ensure saving/loading works
json_string = model.to_json()
clear_session()
model = quantized_model_from_json(json_string)
# generate same output for weights
np.random.seed(42)
for layer in model.layers:
all_weights = []
for i, weights in enumerate(layer.get_weights()):
input_size = np.prod(layer.input.shape.as_list()[1:])
if input_size is None:
input_size = 576 * 10 # to avoid learning sizes
shape = weights.shape
assert input_size > 0, 'input size for {} {}'.format(layer.name, i)
# he normal initialization with a scale factor of 2.0
all_weights.append(
10.0 * np.random.normal(0.0, np.sqrt(2.0 / input_size), shape))
if all_weights:
layer.set_weights(all_weights)
# apply quantizer to weights
model_save_quantized_weights(model)
all_weights = []
for layer in model.layers:
for i, weights in enumerate(layer.get_weights()):
w = np.sum(weights)
all_weights.append(w)
all_weights = np.array(all_weights)
# test_qnetwork_weight_quantization
all_weights_signature = np.array(
[2., -6.75, -0.625, -2., -0.25, -56., 1.125, -1.625, -1.125])
assert all_weights.size == all_weights_signature.size
assert np.all(all_weights == all_weights_signature)
# test_qnetwork_forward:
expected_output = np.array([[0.e+00, 0.e+00, 0.e+00, 0.e+00, 0.e+00,
0.e+00, 1.e+00, 0.e+00, 0.e+00, 0.e+00],
[0.e+00, 0.e+00, 0.e+00, 0.e+00, 0.e+00,
0.e+00, 1.e+00, 0.e+00, 0.e+00, 0.e+00],
[0.e+00, 0.e+00, 0.e+00, 0.e+00, 0.e+00,
0.e+00, 0.e+00, 0.e+00, 6.e-08, 1.e+00],
[0.e+00, 0.e+00, 0.e+00, 0.e+00, 0.e+00,
0.e+00, 1.e+00, 0.e+00, 0.e+00, 0.e+00],
[0.e+00 ,0.e+00, 0.e+00, 0.e+00, 0.e+00,
0.e+00, 1.e+00, 0.e+00, 0.e+00, 0.e+00],
[0.e+00, 0.e+00, 0.e+00, 0.e+00, 0.e+00,
0.e+00, 0.e+00, 0.e+00, 5.e-07, 1.e+00],
[0.e+00, 0.e+00, 0.e+00, 0.e+00, 0.e+00,
0.e+00 ,1.e+00, 0.e+00, 0.e+00, 0.e+00],
[0.e+00, 1.e+00, 0.e+00, 0.e+00, 0.e+00,
0.e+00 ,0.e+00, 0.e+00, 0.e+00, 0.e+00],
[0.e+00, 0.e+00, 0.e+00, 0.e+00, 1.e+00,
0.e+00, 0.e+00, 0.e+00, 0.e+00, 0.e+00],
[0.e+00, 0.e+00, 0.e+00, 0.e+00, 0.e+00,
1.e+00, 0.e+00, 0.e+00, 0.e+00, 0.e+00]]).astype(np.float16)
inputs = 2 * np.random.rand(10, 28, 28, 1)
actual_output = model.predict(inputs).astype(np.float16)
assert_allclose(actual_output, expected_output, rtol=1e-4)
def test_quantized_relu_range(bits, integer, expected_values):
"""Test quantized_relu range function."""
q = quantized_relu(bits, integer)
result = q.range()
assert_allclose(result, expected_values, rtol=1e-05)
def main():
np.random.seed(42)
a = K.constant([-3.0, -2.0, -1.0, -0.5, 0.0, 0.5, 1.0, 2.0, 3.0])
a = K.constant([0.194336])
print(" a =", K.eval(a).astype(np.float16))
print("qa =", K.eval(quantized_relu(6,2)(a)).astype(np.float16))
print("ss =", K.eval(smooth_sigmoid(a)).astype(np.float16))
print("hs =", K.eval(hard_sigmoid(a)).astype(np.float16))
print("ht =", K.eval(hard_tanh(a)).astype(np.float16))
print("st =", K.eval(smooth_tanh(a)).astype(np.float16))
c = K.constant(np.arange(-1.5, 1.51, 0.3))
print(" c =", K.eval(c).astype(np.float16))
print("qb_111 =", K.eval(quantized_bits(1,1,1)(c)).astype(np.float16))
print("qb_210 =", K.eval(quantized_bits(2,1,0)(c)).astype(np.float16))
print("qb_211 =", K.eval(quantized_bits(2,1,1)(c)).astype(np.float16))
print("qb_300 =", K.eval(quantized_bits(3,0,0)(c)).astype(np.float16))
print("qb_301 =", K.eval(quantized_bits(3,0,1)(c)).astype(np.float16))
c_1000 = K.constant(np.array([list(K.eval(c))] * 1000))
b = np.sum(K.eval(bernoulli()(c_1000)).astype(np.int32), axis=0) / 1000.0
print(" hs =", K.eval(hard_sigmoid(c)).astype(np.float16))
print(" b_all =", b.astype(np.float16))
T = 0.0
t = K.eval(stochastic_ternary(threshold=T)(c_1000)).astype(np.int32)
for i in range(10):
print("sternary({}) =".format(i), t[i])
print(" st_all =", np.round(
np.sum(t.astype(np.float32), axis=0).astype(np.float16) /
1000.0, 2).astype(np.float16))
print(" ternary =", K.eval(ternary(threshold=0.5)(c)).astype(np.int32))
b = K.eval(stochastic_binary()(c_1000)).astype(np.int32)
for i in range(5):
print("sbinary({}) =".format(i), b[i])
print("sbinary =", np.round(np.sum(b, axis=0) / 1000.0, 2).astype(np.float16))
print(" binary =", K.eval(binary()(c)).astype(np.int32))
c = K.constant(np.arange(-1.5, 1.51, 0.3))
print(" c =", K.eval(c).astype(np.float16))
print(" b_10 =", K.eval(binary(1)(c)).astype(np.float16))
print("qr_10 =", K.eval(quantized_relu(1,0)(c)).astype(np.float16))
print("qr_11 =", K.eval(quantized_relu(1,1)(c)).astype(np.float16))
print("qr_20 =", K.eval(quantized_relu(2,0)(c)).astype(np.float16))
print("qr_21 =", K.eval(quantized_relu(2,1)(c)).astype(np.float16))
print("qr_101 =", K.eval(quantized_relu(1,0,1)(c)).astype(np.float16))
print("qr_111 =", K.eval(quantized_relu(1,1,1)(c)).astype(np.float16))
print("qr_201 =", K.eval(quantized_relu(2,0,1)(c)).astype(np.float16))
print("qr_211 =", K.eval(quantized_relu(2,1,1)(c)).astype(np.float16))
print("qt_200 =", K.eval(quantized_tanh(2,0)(c)).astype(np.float16))
print("qt_210 =", K.eval(quantized_tanh(2,1)(c)).astype(np.float16))
print("qt_201 =", K.eval(quantized_tanh(2,0,1)(c)).astype(np.float16))
print("qt_211 =", K.eval(quantized_tanh(2,1,1)(c)).astype(np.float16))
set_internal_sigmoid("smooth"); print("with smooth sigmoid")
print("qr_101 =", K.eval(quantized_relu(1,0,1)(c)).astype(np.float16))
print("qr_111 =", K.eval(quantized_relu(1,1,1)(c)).astype(np.float16))
print("qr_201 =", K.eval(quantized_relu(2,0,1)(c)).astype(np.float16))
print("qr_211 =", K.eval(quantized_relu(2,1,1)(c)).astype(np.float16))
print("qt_200 =", K.eval(quantized_tanh(2,0)(c)).astype(np.float16))
print("qt_210 =", K.eval(quantized_tanh(2,1)(c)).astype(np.float16))
print("qt_201 =", K.eval(quantized_tanh(2,0,1)(c)).astype(np.float16))
print("qt_211 =", K.eval(quantized_tanh(2,1,1)(c)).astype(np.float16))
set_internal_sigmoid("real"); print("with real sigmoid")
print("qr_101 =", K.eval(quantized_relu(1,0,1)(c)).astype(np.float16))
print("qr_111 =", K.eval(quantized_relu(1,1,1)(c)).astype(np.float16))
print("qr_201 =", K.eval(quantized_relu(2,0,1)(c)).astype(np.float16))
print("qr_211 =", K.eval(quantized_relu(2,1,1)(c)).astype(np.float16))
print("qt_200 =", K.eval(quantized_tanh(2,0)(c)).astype(np.float16))
print("qt_210 =", K.eval(quantized_tanh(2,1)(c)).astype(np.float16))
print("qt_201 =", K.eval(quantized_tanh(2,0,1)(c)).astype(np.float16))
print("qt_211 =", K.eval(quantized_tanh(2,1,1)(c)).astype(np.float16))
set_internal_sigmoid("hard")
print(" c =", K.eval(c).astype(np.float16))
print("q2_31 =", K.eval(quantized_po2(3,1)(c)).astype(np.float16))
print("q2_32 =", K.eval(quantized_po2(3,2)(c)).astype(np.float16))
print("qr2_21 =", K.eval(quantized_relu_po2(2,1)(c)).astype(np.float16))
print("qr2_22 =", K.eval(quantized_relu_po2(2,2)(c)).astype(np.float16))
print("qr2_44 =", K.eval(quantized_relu_po2(4,1)(c)).astype(np.float16))
with warnings.catch_warnings(record=True) as w:
warnings.simplefilter("always")
print("q2_32_2 =", K.eval(quantized_relu_po2(32,2)(c)).astype(np.float16))
assert len(w) == 1
assert issubclass(w[-1].category, UserWarning)
print(str(w[-1].message))
