def test_qpooling_in_model_quantize():
input_size = (16, 16, 3)
pool_size = (2, 2)
x = Input(input_size)
xin = x
x = AveragePooling2D(pool_size=pool_size, name="pooling")(x)
x = GlobalAveragePooling2D(name="global_pooling")(x)
model = Model(inputs=xin, outputs=x)
quantize_config = {
"QAveragePooling2D": {
"average_quantizer": "binary",
"activation_quantizer": "binary"
},
"QGlobalAveragePooling2D": {
"average_quantizer": "quantized_bits(4, 0, 1)",
"activation_quantizer": "ternary"
}
}
qmodel = model_quantize(model, quantize_config, 4)
print_qstats(qmodel)
assert_equal(str(qmodel.layers[1].average_quantizer_internal), "binary()")
assert_equal(str(qmodel.layers[1].activation), "binary()")
assert_equal(str(qmodel.layers[2].average_quantizer_internal),
"quantized_bits(4,0,1)")
assert_equal(str(qmodel.layers[2].activation), "ternary()")
def QDenseModel(weights_f, load_weights=False):
"""Construct QDenseModel."""
x = x_in = Input((RESHAPED, ), name="input")
x = QActivation("quantized_relu(4)", name="act_i")(x)
x = QDense(N_HIDDEN,
kernel_quantizer=ternary(),
bias_quantizer=quantized_bits(4, 0, 1),
name="dense0")(x)
x = QActivation("quantized_relu(2)", name="act0")(x)
x = QDense(NB_CLASSES,
kernel_quantizer=quantized_bits(4, 0, 1),
bias_quantizer=quantized_bits(4, 0, 1),
name="dense2")(x)
x = Activation("softmax", name="softmax")(x)
model = Model(inputs=[x_in], outputs=[x])
model.summary()
model.compile(loss="categorical_crossentropy",
optimizer=OPTIMIZER,
metrics=["accuracy"])
if load_weights and weights_f:
model.load_weights(weights_f)
print_qstats(model)
return model
def test_q_average_pooling(pooling, input_size, pool_size, strides, padding,
data_format, average_quantizer,
activation_quantizer, y):
"""q_average_pooling test utility."""
np.random.seed(33)
x = Input(input_size)
xin = x
if pooling == 'QAveragePooling2D':
x = QAveragePooling2D(pool_size=pool_size,
strides=strides,
padding=padding,
data_format=data_format,
average_quantizer=average_quantizer,
activation=activation_quantizer,
name='qpooling')(x)
else:
x = QGlobalAveragePooling2D(data_format=data_format,
average_quantizer=average_quantizer,
activation=activation_quantizer,
name='qpooling')(x)
model = Model(inputs=xin, outputs=x)
# Prints qstats to make sure it works with Conv1D layer
print_qstats(model)
size = (2, ) + input_size
inputs = np.random.rand(size[0], size[1], size[2], size[3])
if data_format == 'channels_first':
assert_raises(tf.errors.InvalidArgumentError, model.predict, inputs)
else:
p = model.predict(inputs).astype(np.float16)
assert_allclose(p, y, rtol=1e-4)
# Reloads the model to ensure saving/loading works
json_string = model.to_json()
clear_session()
reload_model = quantized_model_from_json(json_string)
p = reload_model.predict(inputs).astype(np.float16)
assert_allclose(p, y, rtol=1e-4)
# Saves the model as an h5 file using Keras's model.save()
fd, fname = tempfile.mkstemp(".h5")
model.save(fname)
del model # Delete the existing model
# Returns a compiled model identical to the previous one
loaded_model = load_qmodel(fname)
# Cleans the created h5 file after loading the model
os.close(fd)
os.remove(fname)
# Applys quantizer to weights
model_save_quantized_weights(loaded_model)
p = loaded_model.predict(inputs).astype(np.float16)
assert_allclose(p, y, rtol=1e-4)
def UseNetwork(weights_f, load_weights=False):
"""Use DenseModel.
Args:
weights_f: weight file location.
load_weights: load weights when it is True.
"""
model = QDenseModel(weights_f, load_weights)
batch_size = BATCH_SIZE
(x_train_, y_train_), (x_test_, y_test_) = mnist.load_data()
x_train_ = x_train_.reshape(60000, RESHAPED)
x_test_ = x_test_.reshape(10000, RESHAPED)
x_train_ = x_train_.astype("float32")
x_test_ = x_test_.astype("float32")
x_train_ /= 255
x_test_ /= 255
print(x_train_.shape[0], "train samples")
print(x_test_.shape[0], "test samples")
y_train_ = to_categorical(y_train_, NB_CLASSES)
y_test_ = to_categorical(y_test_, NB_CLASSES)
if not load_weights:
model.fit(x_train_,
y_train_,
batch_size=batch_size,
epochs=NB_EPOCH,
verbose=VERBOSE,
validation_split=VALIDATION_SPLIT)
if weights_f:
model.save_weights(weights_f)
score = model.evaluate(x_test_, y_test_, verbose=VERBOSE)
print_qstats(model)
print("Test score:", score[0])
print("Test accuracy:", score[1])
def test_qconv1d():
np.random.seed(33)
x = Input((4, 4,))
y = QConv1D(
2, 1,
kernel_quantizer=quantized_bits(6, 2, 1),
bias_quantizer=quantized_bits(4, 0, 1),
name='qconv1d')(
x)
model = Model(inputs=x, outputs=y)
#Extract model operations
model_ops = extract_model_operations(model)
# Assertion about the number of operations for this Conv1D layer
assert model_ops['qconv1d']["number_of_operations"] == 32
# Print qstats to make sure it works with Conv1D layer
print_qstats(model)
# reload the model to ensure saving/loading works
json_string = model.to_json()
clear_session()
model = quantized_model_from_json(json_string)
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 = 10 * 10
shape = weights.shape
assert input_size > 0, 'input size for {} {}'.format(layer.name, i)
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)
# Save the model as an h5 file using Keras's model.save()
fd, fname = tempfile.mkstemp('.h5')
model.save(fname)
del model # Delete the existing model
# Returns a compiled model identical to the previous one
model = load_qmodel(fname)
#Clean the created h5 file after loading the model
os.close(fd)
os.remove(fname)
# apply quantizer to weights
model_save_quantized_weights(model)
inputs = np.random.rand(2, 4, 4)
p = model.predict(inputs).astype(np.float16)
'''
y = np.array([[[0.1309, -1.229], [-0.4165, -2.639], [-0.08105, -2.299],
[1.981, -2.195]],
[[-0.3174, -3.94], [-0.3352, -2.316], [0.105, -0.833],
[0.2115, -2.89]]]).astype(np.float16)
'''
y = np.array([[[-2.441, 3.816], [-3.807, -1.426], [-2.684, -1.317],
[-1.659, 0.9834]],
[[-4.99, 1.139], [-2.559, -1.216], [-2.285, 1.905],
[-2.652, -0.467]]]).astype(np.float16)
assert np.all(p == y)
def test_qconv1d(layer_cls):
np.random.seed(33)
if layer_cls == "QConv1D":
x = Input((
4,
4,
))
y = QConv1D(2,
1,
kernel_quantizer=quantized_bits(6, 2, 1, alpha=1.0),
bias_quantizer=quantized_bits(4, 0, 1),
name='qconv1d')(x)
model = Model(inputs=x, outputs=y)
else:
x = Input((
4,
4,
))
y = QSeparableConv1D(2,
2,
depthwise_quantizer=quantized_bits(6,
2,
1,
alpha=1.0),
pointwise_quantizer=quantized_bits(4,
0,
1,
alpha=1.0),
bias_quantizer=quantized_bits(4, 0, 1),
name='qconv1d')(x)
model = Model(inputs=x, outputs=y)
# Extract model operations
model_ops = extract_model_operations(model)
# Check the input layer model operation was found correctly
assert model_ops['qconv1d']['type'][0] != 'null'
# Assertion about the number of operations for this (Separable)Conv1D layer
if layer_cls == "QConv1D":
assert model_ops['qconv1d']['number_of_operations'] == 32
else:
assert model_ops['qconv1d']['number_of_operations'] == 30
# Print qstats to make sure it works with Conv1D layer
print_qstats(model)
# reload the model to ensure saving/loading works
# json_string = model.to_json()
# clear_session()
# model = quantized_model_from_json(json_string)
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 = 10 * 10
shape = weights.shape
assert input_size > 0, 'input size for {} {}'.format(layer.name, i)
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)
# Save the model as an h5 file using Keras's model.save()
fd, fname = tempfile.mkstemp('.h5')
model.save(fname)
del model # Delete the existing model
# Return a compiled model identical to the previous one
model = load_qmodel(fname)
# Clean the created h5 file after loading the model
os.close(fd)
os.remove(fname)
# apply quantizer to weights
model_save_quantized_weights(model)
inputs = np.random.rand(2, 4, 4)
p = model.predict(inputs).astype(np.float16)
if layer_cls == "QConv1D":
y = np.array([[[-2.441, 3.816], [-3.807, -1.426], [-2.684, -1.317],
[-1.659, 0.9834]],
[[-4.99, 1.139], [-2.559, -1.216], [-2.285, 1.905],
[-2.652, -0.467]]]).astype(np.float16)
else:
y = np.array([[[-2.275, -3.178], [-0.4358, -3.262], [1.987, 0.3987]],
[[-0.01251, -0.376], [0.3928, -1.328],
[-1.243, -2.43]]]).astype(np.float16)
assert_allclose(p, y, rtol=1e-4)