def make_simple_model(self) -> Graph:
graph = Graph()
# two inputs
x = Input(
'input',
[1, 5, 5, 3],
Float32(),
)
w = Constant(
'weight',
Float32(),
np.zeros([1, 2, 2, 3]),
dimension_format='NHWC',
)
# Conv
conv = Conv('conv', [1, 4, 4, 1],
Float32(), {
'X': x,
'W': w
},
kernel_shape=[2, 2])
# One output
y = Output('output', [1, 4, 4, 1], Float32(), {'input': conv})
# add ops to the graph
graph.add_op_and_inputs(y)
return graph
def create_expected_graph(data: np.ndarray) -> Graph:
graph = Graph()
# input
x = Input('placeholder', [1, 5, 5, 3], Float32())
# constant and internal nodes
w = Constant('weight', Float32(), data)
q = BinaryMeanScalingQuantizer('qtz1', [1, 2, 2, 3], Float32(),
{'input': w})
# Conv
conv = Conv('conv', [1, 4, 4, 3],
Float32(), {
'X': x,
'W': q
},
kernel_shape=[2, 2])
# One output
rs = Reshape('reshape', [1, 48], Float32(), {'data': conv})
y = Output(
'output',
[1, 48],
Float32(),
{'input': rs},
)
# add ops to the graph
graph.add_op_and_inputs(y)
return graph
def create_sample_graph(data1: np.ndarray, data2: np.ndarray) -> Graph:
graph = Graph()
# input
x = Input('placeholder', [1, 5, 5, 3], Float32())
# Conv1
w1 = Constant('weight1', Float32(), data1)
conv1 = Conv('conv1', [1, 4, 4, 3], Float32(), {'X': x, 'W': w1}, kernel_shape=[2, 2])
# activation quantizer
s1 = Constant('aq_const1', Float32(), np.array(1))
s2 = Constant('aq_const2', Float32(), np.array(2))
aq = LinearMidTreadHalfQuantizer('aqtz1', [1, 4, 4, 3], Float32(), {'X': conv1, 'Y': s1, 'Z': s2})
# Conv2
w2 = Constant('weight2', Float32(), data2)
kq = BinaryMeanScalingQuantizer('kqtz1', [1, 2, 2, 3], Float32(), {'input': w2})
conv2 = Conv('conv2', [1, 3, 3, 3], Float32(), {'X': aq, 'W': kq}, kernel_shape=[2, 2])
conv2.a_quantizer = [aq]
conv2.quantizer = kq
# One output
y = Output('output', [1, 3, 3, 3], Float32(), {'input': conv2})
# add ops to the graph
graph.add_op_and_inputs(y)
return graph
def create_sample_graph() -> Graph:
graph = Graph()
x = Input('placeholder', [2], Float32())
s1 = Constant('potato_1', Float32(), np.array([1, 2]))
s2 = Constant('potato_2', Float32(), np.array([1, 3]))
add1 = Add('potatoes', [2], Float32(), {'A': s1, 'B': s2})
add2 = Add('more_potatoes', [2], Float32(), {'A': x, 'B': add1})
# One output
y = Output('output', [2], Float32(), {'input': add2})
# add ops to the graph
graph.add_op_and_inputs(y)
return graph
def test_conv_consistency(self) -> None:
"""Test code for Conv."""
x = Input(
'const1',
[1, 3, 3, 3],
Float32(),
)
w = Constant('weight', Float32(), np.zeros([1, 2, 2, 3]))
input_ops = {'X': cast(Operator, x), 'W': cast(Operator, w)}
add = Conv('conv_under_test', [1, 3, 3, 3],
Float32(),
input_ops,
pads=[1, 1, 2, 2],
strides=[2, 2])
print("Consistency test for conv operator passed!")
def create_sample_graph(data1: np.ndarray, data2: np.ndarray) -> Graph:
graph = Graph()
# input
x = Input('placeholder', [1, 5, 5, 3], Float32())
# Conv1
w1 = Constant('weight1', Float32(), data1)
conv1 = Conv('conv1', [1, 4, 4, 3], Float32(), {'X': x, 'W': w1}, kernel_shape=[2, 2])
# activation quantizer
s1 = Constant('aq_const1', Int32(), np.array([2], dtype=np.int32))
s2 = Constant('aq_const2', Float32(), np.array([2.0], dtype=np.float32))
aq1 = LinearMidTreadHalfQuantizer('aqtz1', [1, 4, 4, 3], Float32(), {'X': conv1, 'Y': s1, 'Z': s2})
# Conv2
w2 = Constant('weight2', Float32(), data2)
kq = BinaryMeanScalingQuantizer('kqtz1', [1, 2, 2, 3], Float32(), {'input': w2})
conv2 = Conv('conv2', [1, 3, 3, 3], Float32(), {'X': aq1, 'W': kq}, kernel_shape=[2, 2])
conv2.a_quantizer = [aq1]
conv2.quantizer = kq
conv2.is_quantized = True
sc = Constant('bn_scale', Float32(), np.random.rand(3))
be = Constant('bn_b', Float32(), np.random.rand(3))
mu = Constant('bn_mu', Float32(), np.random.rand(3))
va = Constant('bn_var', Float32(), np.random.rand(3))
bn = BatchNormalization('bn', [1, 3, 3, 3], Float32(), {'X': conv2,
'scale': sc,
'B': be,
'mean': mu,
'var': va})
# activation quantizer
s3 = Constant('aq_const3', Int32(), np.array([2], dtype=np.int32))
s4 = Constant('aq_const4', Float32(), np.array([2.0], dtype=np.float32))
aq2 = LinearMidTreadHalfQuantizer('aqtz2', [1, 3, 3, 3], Float32(), {'X': bn, 'Y': s3, 'Z': s4})
# One output
y = Output('output', [1, 3, 3, 3], Float32(), {'input': aq2})
# add ops to the graph
graph.add_op_and_inputs(y)
return graph
def create_sample_graph(data1: np.ndarray) -> Graph:
graph = Graph()
# input
x = Input('placeholder', [1, 5, 5, 3], Float32())
# Conv1
w1 = Constant('weight1', Float32(), data1)
conv1 = Conv('conv1', [1, 4, 4, 3], QUANTIZED_PACKED(), {'X': x, 'W': w1}, kernel_shape=[2, 2])
pool1 = SpaceToDepth('s2d', [1, 2, 2, 12], Float32(), {'input': conv1})
# One output
y = Output('output', [1, 2, 2, 12], Float32(), {'input': pool1})
# add ops to the graph
graph.add_op_and_inputs(y)
return graph
def create_sample_graph_2(data1: np.ndarray) -> Graph:
graph = Graph()
# input
x = Input('placeholder', [1, 5, 5, 3], Float32())
# Conv1
w1 = Constant('weight1', Float32(), data1)
conv1 = Conv('conv1', [1, 4, 4, 3], Float32(), {'X': x, 'W': w1}, kernel_shape=[2, 2])
s1 = Constant('const1', Float32(), np.zeros([1, 4, 4, 3]))
add1 = Add('add', [1, 4, 4, 3], Float32(), {'A': conv1, 'B': s1})
y = Output('output', [1, 4, 4, 3], Float32(), {'input': add1})
# add ops to the graph
graph.add_op_and_inputs(y)
return graph
def test_graph_conv(self) -> None:
"""Test code for making a simple graph with Conv."""
graph = Graph()
# two inputs
x = Input(
'input',
[1, 5, 5, 3],
Float32(),
)
w = Constant(
'weight',
Float32(),
np.zeros([1, 2, 2, 3])
)
# Conv
conv = Conv(
'conv',
[1, 4, 4, 3],
Float32(),
{'X': x, 'W': w}, # you can get these keys by 'Conv.input_names'
kernel_shape=[2, 2]
)
# One output
y = Output(
'output',
[1, 4, 4, 3],
Float32(),
{'input': conv} # you can get this key by 'Output.input_names'
)
# add ops to the graph
graph.add_op(x)
graph.add_op(w)
graph.add_op(conv)
graph.add_op(y)
self.assertTrue(graph.check_nodes(), "All inputs of operators must match their outputs.")
print("Graph test passed!")
def create_expected_graph(data: np.ndarray) -> Graph:
graph = Graph()
data = data.transpose([3, 2, 1, 0])
# input
x = Input('placeholder', [1, 5, 5, 3],
Float32(),
dimension_format='NHWC')
# constant and internal nodes
w = Constant('weight', Float32(), data, dimension_format='NHWC')
i1 = Identity('identity1', [1, 2, 2, 3],
Float32(), {'input': w},
dimension_format='NHWC')
q = BinaryMeanScalingQuantizer('qtz1', [1, 2, 2, 3],
Float32(), {'input': i1},
dimension_format='NHWC')
# Conv
conv = Conv('conv', [1, 4, 4, 3],
Float32(), {
'X': x,
'W': q
},
kernel_shape=[2, 2],
dimension_format='NHWC')
# One output
rs = Reshape('reshape', [1, 48], Float32(), {'data': conv})
y = Output(
'output',
[1, 48],
Float32(),
{'input': rs},
)
# add ops to the graph
graph.add_op_and_inputs(y)
return graph
def test_conv(self) -> None:
"""Test code for Conv."""
# get Conv's input names
i_names = Conv.input_names
self.assertTrue({'X', 'W'}.issubset(set(i_names)))
# set x to MaxPool m's input
x = Input(
'input',
[1, 3, 3, 3],
Float32(),
)
w = Constant('weight', Float32(), np.zeros([1, 2, 2, 5]))
inputs: Dict[str, Operator] = {i_names[0]: x, i_names[1]: w}
c = Conv("conv1", [1, 2, 2, 3], Float32(), inputs, kernel_shape=[2, 2])
self.assertEqual(c.batchsize, 1)
self.assertEqual(c.height, 2)
self.assertEqual(c.width, 2)
self.assertEqual(c.channel, 3)
self.assertEqual(c.kernel_height, 2)
self.assertEqual(c.kernel_width, 2)
print("Conv test passed!")
