def test_mat_mul(m, n, k, a, b):
"""Tests a MatMul replacement."""
a_constant_name = "a_constant"
b_constant_name = "b_constant"
mat_mul_name = "mat_mul"
float_graph_def = tf.GraphDef()
a_constant = quantize_graph.create_constant_node(a_constant_name,
value=a,
dtype=tf.float32,
shape=[m, k])
float_graph_def.node.extend([a_constant])
b_constant = quantize_graph.create_constant_node(b_constant_name,
value=b,
dtype=tf.float32,
shape=[k, n])
float_graph_def.node.extend([b_constant])
mat_mul_node = quantize_graph.create_node("MatMul", mat_mul_name,
[a_constant_name, b_constant_name])
quantize_graph.set_attr_dtype(mat_mul_node, "T", tf.float32)
quantize_graph.set_attr_bool(mat_mul_node, "transpose_a", False)
quantize_graph.set_attr_bool(mat_mul_node, "transpose_b", False)
float_graph_def.node.extend([mat_mul_node])
test_graph(float_graph_def, {}, [mat_mul_name])
def test_multiple_outputs(self):
input_constant_name = "input_constant"
split_constant_name = "split_constant"
split_name = "split"
concat_constant_name = "concat_constant"
concat_name = "concat"
float_graph_def = graph_pb2.GraphDef()
input_constant = quantize_graph.create_constant_node(
input_constant_name,
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=dtypes.float32,
shape=[2, 6])
float_graph_def.node.extend([input_constant])
split_constant = quantize_graph.create_constant_node(
split_constant_name, value=1, dtype=dtypes.int32, shape=[])
float_graph_def.node.extend([split_constant])
split_node = quantize_graph.create_node(
"Split", split_name, [split_constant_name, input_constant_name])
quantize_graph.set_attr_int(split_node, "num_split", 2)
quantize_graph.set_attr_dtype(split_node, "T", dtypes.float32)
float_graph_def.node.extend([split_node])
concat_constant = quantize_graph.create_constant_node(
concat_constant_name, value=1, dtype=dtypes.int32, shape=[])
float_graph_def.node.extend([concat_constant])
concat_node = quantize_graph.create_node(
"Concat", concat_name,
[concat_constant_name, split_name + ":0", split_name + ":1"])
quantize_graph.set_attr_int(concat_node, "N", 2)
quantize_graph.set_attr_dtype(concat_node, "T", dtypes.float32)
float_graph_def.node.extend([concat_node])
test_graph(float_graph_def, {}, [concat_name])
def test_bias_add_w_fallback_min_max_vars(self):
input_node = quantize_graph.create_constant_node(
"input", value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
dtype=tf.float32, shape=[1, 1, 2, 5])
offset_node = quantize_graph.create_constant_node(
"offset", value=[1, 2, 3, 4, 5], dtype=tf.float32, shape=[5])
bias_add_node = quantize_graph.create_node(
"BiasAdd", "bias_add", [input_node.name, offset_node.name])
quantize_graph.set_attr_dtype(bias_add_node, "T", tf.float32)
float_graph_def = tf.GraphDef()
float_graph_def.node.extend([input_node, offset_node, bias_add_node])
test_graph(float_graph_def, {}, [bias_add_node.name], log_graph=True)
# Verify there is only one Quantize, one Requantize op, and no
# RequantizationRange op.
eightbit_rewriter = quantize_graph.GraphRewriter(
float_graph_def, "eightbit", quantized_input_range=None,
fallback_quantization_range=[-.5, 15.5])
eightbit_graph_def = eightbit_rewriter.rewrite([bias_add_node.name])
ops = [node.op for node in eightbit_graph_def.node]
node_names = [node.name for node in eightbit_graph_def.node]
# No quantize since all inputs are const and can be quantized up-front.
self.assertEqual(0, ops.count("QuantizeV2") + ops.count("Quantize"))
# One dequantize at the end.
self.assertEqual(1, ops.count("Dequantize"))
# No RequantizationRange
self.assertEqual(0, ops.count("RequantizationRange"))
# The fallback constants are in the graph.
self.assertEqual(1, node_names.count("fallback_quantization_min_value"))
self.assertEqual(1, node_names.count("fallback_quantization_max_value"))
def test_concat(self):
shape_constant_name = "shape_constant"
a_constant_name = "a_constant"
b_constant_name = "b_constant"
concat_name = "concat"
float_graph_def = tf.GraphDef()
shape_constant = quantize_graph.create_constant_node(shape_constant_name,
value=0,
dtype=tf.int32,
shape=[])
float_graph_def.node.extend([shape_constant])
a_constant = quantize_graph.create_constant_node(a_constant_name,
value=[1, 2, 3, 4, 5, 6, 7,
8, 9, 10, 11, 12],
dtype=tf.float32,
shape=[2, 2, 3])
float_graph_def.node.extend([a_constant])
b_constant = quantize_graph.create_constant_node(b_constant_name,
value=[13, 14, 15, 16, 17,
18, 19, 20, 21, 22,
23, 24],
dtype=tf.float32,
shape=[2, 2, 3])
float_graph_def.node.extend([b_constant])
concat_node = quantize_graph.create_node("Concat", concat_name,
[shape_constant_name,
a_constant_name, b_constant_name])
quantize_graph.set_attr_int(concat_node, "N", 2)
quantize_graph.set_attr_dtype(concat_node, "T", tf.float32)
float_graph_def.node.extend([concat_node])
test_graph(float_graph_def, {}, [concat_name])
def test_conv(depth, image_width, image_height, image_batch_count, filter_size,
filter_count, stride, padding, input_values, filter_values):
"""Tests a Conv replacement."""
input_constant_name = "input_constant"
filter_constant_name = "filter_constant"
conv_name = "conv"
float_graph_def = graph_pb2.GraphDef()
input_constant = quantize_graph.create_constant_node(
input_constant_name,
value=input_values,
dtype=dtypes.float32,
shape=[image_batch_count, image_height, image_width, depth])
float_graph_def.node.extend([input_constant])
filter_constant = quantize_graph.create_constant_node(
filter_constant_name,
value=filter_values,
dtype=dtypes.float32,
shape=[filter_size, filter_size, depth, filter_count])
float_graph_def.node.extend([filter_constant])
conv_node = quantize_graph.create_node(
"Conv2D", conv_name, [input_constant_name, filter_constant_name])
quantize_graph.set_attr_dtype(conv_node, "T", dtypes.float32)
quantize_graph.set_attr_int_list(conv_node, "strides", [1, stride, stride, 1])
quantize_graph.set_attr_string(conv_node, "padding", padding)
float_graph_def.node.extend([conv_node])
test_graph(float_graph_def, {}, [conv_name])
def test_concat(self):
shape_constant_name = "shape_constant"
a_constant_name = "a_constant"
b_constant_name = "b_constant"
concat_name = "concat"
float_graph_def = tf.GraphDef()
shape_constant = quantize_graph.create_constant_node(
shape_constant_name, value=0, dtype=tf.int32, shape=[])
float_graph_def.node.extend([shape_constant])
a_constant = quantize_graph.create_constant_node(
a_constant_name,
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=tf.float32,
shape=[2, 2, 3])
float_graph_def.node.extend([a_constant])
b_constant = quantize_graph.create_constant_node(
b_constant_name,
value=[13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24],
dtype=tf.float32,
shape=[2, 2, 3])
float_graph_def.node.extend([b_constant])
concat_node = quantize_graph.create_node(
"Concat", concat_name,
[shape_constant_name, a_constant_name, b_constant_name])
quantize_graph.set_attr_int(concat_node, "N", 2)
quantize_graph.set_attr_dtype(concat_node, "T", tf.float32)
float_graph_def.node.extend([concat_node])
test_graph(float_graph_def, {}, [concat_name])
def test_mat_mul(m, n, k, a, b):
"""Tests a MatMul replacement."""
a_constant_name = "a_constant"
b_constant_name = "b_constant"
mat_mul_name = "mat_mul"
float_graph_def = tf.GraphDef()
a_constant = quantize_graph.create_constant_node(a_constant_name,
value=a,
dtype=tf.float32,
shape=[m, k])
float_graph_def.node.extend([a_constant])
b_constant = quantize_graph.create_constant_node(b_constant_name,
value=b,
dtype=tf.float32,
shape=[k, n])
float_graph_def.node.extend([b_constant])
mat_mul_node = quantize_graph.create_node(
"MatMul", mat_mul_name, [a_constant_name, b_constant_name])
quantize_graph.set_attr_dtype(mat_mul_node, "T", tf.float32)
quantize_graph.set_attr_bool(mat_mul_node, "transpose_a", False)
quantize_graph.set_attr_bool(mat_mul_node, "transpose_b", False)
float_graph_def.node.extend([mat_mul_node])
test_graph(float_graph_def, {}, [mat_mul_name])
def test_relu_w_fake_quant_w_min_max_vars(self):
input_node = quantize_graph.create_constant_node(
"input",
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=dtypes.float32,
shape=[1, 2, 6, 1])
relu_node = quantize_graph.create_node("Relu", "relu", [input_node.name])
quantize_graph.set_attr_dtype(relu_node, "T", dtypes.float32)
min_node = quantize_graph.create_constant_node(
"min_bias_add", value=0, dtype=dtypes.float32, shape=[])
max_node = quantize_graph.create_constant_node(
"max_bias_add", value=12, dtype=dtypes.float32, shape=[])
fake_quant_node = quantize_graph.create_node(
"FakeQuantWithMinMaxVars", "fake_quant",
[relu_node.name, min_node.name, max_node.name])
float_graph_def = graph_pb2.GraphDef()
float_graph_def.node.extend(
[input_node, relu_node, min_node, max_node, fake_quant_node])
test_graph(float_graph_def, {}, [fake_quant_node.name], log_graph=True)
# Verify there is only one Quantize and one Requantize op.
eightbit_rewriter = quantize_graph.GraphRewriter(
float_graph_def, "eightbit", quantized_input_range=None)
eightbit_graph_def = eightbit_rewriter.rewrite([fake_quant_node.name])
ops = [node.op for node in eightbit_graph_def.node]
# No quantize since all inputs are const and can be quantized up-front.
self.assertEqual(0, ops.count("QuantizeV2") + ops.count("Quantize"))
# One dequantize at the end.
self.assertEqual(1, ops.count("Dequantize"))
def test_keep_control_edges(self):
no_op_name = "no_op"
a_constant_name = "a_constant"
b_constant_name = "b_constant"
a_check_name = "a_check"
b_check_name = "b_check"
a_identity_name = "a_identity"
b_identity_name = "b_identity"
add_name = "add"
graph_def = graph_pb2.GraphDef()
no_op = quantize_graph.create_node("NoOp", no_op_name, [])
graph_def.node.extend([no_op])
a_constant = quantize_graph.create_constant_node(
a_constant_name, value=1, dtype=dtypes.float32, shape=[])
graph_def.node.extend([a_constant])
a_check_node = quantize_graph.create_node("CheckNumerics", a_check_name,
[a_constant_name])
graph_def.node.extend([a_check_node])
a_identity_node = quantize_graph.create_node(
"Identity", a_identity_name,
[a_constant_name, "^" + a_check_name, "^" + no_op_name])
graph_def.node.extend([a_identity_node])
b_constant = quantize_graph.create_constant_node(
b_constant_name, value=1, dtype=dtypes.float32, shape=[])
graph_def.node.extend([b_constant])
b_check_node = quantize_graph.create_node("CheckNumerics", b_check_name,
[b_constant_name])
graph_def.node.extend([b_check_node])
b_identity_node = quantize_graph.create_node(
"Identity", b_identity_name, [b_constant_name, "^" + b_check_name])
graph_def.node.extend([b_identity_node])
add_node = quantize_graph.create_node("Add", add_name,
[a_identity_name, b_identity_name])
quantize_graph.set_attr_dtype(add_node, "T", dtypes.float32)
graph_def.node.extend([add_node])
expected_output = graph_pb2.GraphDef()
no_op = quantize_graph.create_node("NoOp", no_op_name, [])
expected_output.node.extend([no_op])
a_constant = quantize_graph.create_constant_node(
a_constant_name, value=1, dtype=dtypes.float32, shape=[])
expected_output.node.extend([a_constant])
a_identity_node = quantize_graph.create_node(
"Identity", a_identity_name, [a_constant_name, "^" + no_op_name])
expected_output.node.extend([a_identity_node])
b_constant = quantize_graph.create_constant_node(
b_constant_name, value=1, dtype=dtypes.float32, shape=[])
expected_output.node.extend([b_constant])
add_node = quantize_graph.create_node("Add", add_name,
[a_identity_name, b_constant_name])
quantize_graph.set_attr_dtype(add_node, "T", dtypes.float32)
expected_output.node.extend([add_node])
expected_output.versions.CopyFrom(graph_def.versions)
expected_output.library.CopyFrom(graph_def.library)
output = graph_util.remove_training_nodes(graph_def)
stripped_output = graph_util.extract_sub_graph(output, [add_name])
self.assertProtoEquals(expected_output, stripped_output)
def test_reshape(self):
"""Tests that MatMul->Reshape->MatMul avoids extra quantize/dequantize."""
def make_matmul(name, a, b):
n = quantize_graph.create_node("MatMul", name, [a.name, b.name])
quantize_graph.set_attr_dtype(n, "T", tf.float32)
quantize_graph.set_attr_bool(n, "transpose_a", False)
quantize_graph.set_attr_bool(n, "transpose_b", False)
return n
# matmul_1 = input*weight_1
input_node = quantize_graph.create_constant_node("input",
value=[0, 1, 2, 3],
dtype=tf.float32,
shape=[4, 1])
weight_1_node = quantize_graph.create_constant_node(
"weight_1",
value=[.5, .6, .7, .8, .9],
dtype=tf.float32,
shape=[1, 5])
matmul_1_node = make_matmul("matmul_1", input_node, weight_1_node)
# Reshape 4x5 to 10x2.
new_shape_node = quantize_graph.create_constant_node("new_shape_node",
value=[10, 2],
dtype=tf.int32,
shape=[2])
reshape_node = quantize_graph.create_node(
"Reshape", "reshape", [matmul_1_node.name, new_shape_node.name])
quantize_graph.set_attr_dtype(reshape_node, "T", tf.float32)
# matmul_2_node = reshape*weight_2
weight_2_node = quantize_graph.create_constant_node("weight_2",
value=[1.5, 2.5],
dtype=tf.float32,
shape=[2, 1])
matmul_2_node = make_matmul("matmul_2", reshape_node, weight_2_node)
g = tf.GraphDef()
g.node.extend([
input_node, weight_1_node, matmul_1_node, new_shape_node,
reshape_node, weight_2_node, matmul_2_node
])
# Test the graph
test_graph(g, {}, ["matmul_2"])
# Verify there is only one Quantize and one Requantize op.
eightbit_rewriter = quantize_graph.GraphRewriter(g, "eightbit")
eightbit_graph_def = eightbit_rewriter.rewrite(["matmul_2"])
tf.logging.info("S:\n%s", str(eightbit_graph_def))
ops = [node.op for node in eightbit_graph_def.node]
# No quantize since all inputs are const and can be quantized up-front.
self.assertEqual(0, ops.count("QuantizeV2") + ops.count("Quantize"))
# One dequantize at the end.
self.assertEqual(1, ops.count("Dequantize"))
def test_bias_add_w_fake_quant_w_min_max_vars(self):
input_node = quantize_graph.create_constant_node(
"input",
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
dtype=tf.float32,
shape=[1, 1, 2, 5])
offset_node = quantize_graph.create_constant_node(
"offset", value=[1, 2, 3, 4, 5], dtype=tf.float32, shape=[5])
bias_add_node = quantize_graph.create_node(
"BiasAdd", "bias_add", [input_node.name, offset_node.name])
quantize_graph.set_attr_dtype(bias_add_node, "T", tf.float32)
min_node = quantize_graph.create_constant_node("min_bias_add",
value=-.5,
dtype=tf.float32,
shape=[])
max_node = quantize_graph.create_constant_node("max_bias_add",
value=15.5,
dtype=tf.float32,
shape=[])
fake_quant_node = quantize_graph.create_node(
"FakeQuantWithMinMaxVars", "fake_quant",
[bias_add_node.name, min_node.name, max_node.name])
float_graph_def = tf.GraphDef()
float_graph_def.node.extend([
input_node, offset_node, bias_add_node, min_node, max_node,
fake_quant_node
])
test_graph(float_graph_def, {}, [fake_quant_node.name], log_graph=True)
# Verify there is only one Quantize and one Requantize op.
# Pass in fallback_quantization_range, although it will have no effect
# because the FakeQuantWithMinMaxVars are used instead.
eightbit_rewriter = quantize_graph.GraphRewriter(
float_graph_def,
"eightbit",
quantized_input_range=None,
fallback_quantization_range=[-100, 100])
eightbit_graph_def = eightbit_rewriter.rewrite([fake_quant_node.name])
ops = [node.op for node in eightbit_graph_def.node]
node_names = [node.name for node in eightbit_graph_def.node]
# No quantize since all inputs are const and can be quantized up-front.
self.assertEqual(0, ops.count("QuantizeV2") + ops.count("Quantize"))
# One dequantize at the end.
self.assertEqual(1, ops.count("Dequantize"))
# The fallback constants are not in the graph.
self.assertEqual(0,
node_names.count("fallback_quantization_min_value"))
self.assertEqual(0,
node_names.count("fallback_quantization_max_value"))
def test_reshape(self):
"""Tests that MatMul->Reshape->MatMul avoids extra quantize/dequantize."""
def make_matmul(name, a, b):
n = quantize_graph.create_node("MatMul", name, [a.name, b.name])
quantize_graph.set_attr_dtype(n, "T", dtypes.float32)
quantize_graph.set_attr_bool(n, "transpose_a", False)
quantize_graph.set_attr_bool(n, "transpose_b", False)
return n
# matmul_1 = input*weight_1
input_node = quantize_graph.create_constant_node(
"input", value=[0, 1, 2, 3], dtype=dtypes.float32, shape=[4, 1])
weight_1_node = quantize_graph.create_constant_node(
"weight_1",
value=[.5, .6, .7, .8, .9],
dtype=dtypes.float32,
shape=[1, 5])
matmul_1_node = make_matmul("matmul_1", input_node, weight_1_node)
# Reshape 4x5 to 10x2.
new_shape_node = quantize_graph.create_constant_node(
"new_shape_node", value=[10, 2], dtype=dtypes.int32, shape=[2])
reshape_node = quantize_graph.create_node(
"Reshape", "reshape", [matmul_1_node.name, new_shape_node.name])
quantize_graph.set_attr_dtype(reshape_node, "T", dtypes.float32)
# matmul_2_node = reshape*weight_2
weight_2_node = quantize_graph.create_constant_node(
"weight_2", value=[1.5, 2.5], dtype=dtypes.float32, shape=[2, 1])
matmul_2_node = make_matmul("matmul_2", reshape_node, weight_2_node)
g = graph_pb2.GraphDef()
g.node.extend([
input_node, weight_1_node, matmul_1_node, new_shape_node, reshape_node,
weight_2_node, matmul_2_node
])
# Test the graph
test_graph(g, {}, ["matmul_2"])
# Verify there is only one Quantize and one Requantize op.
eightbit_rewriter = quantize_graph.GraphRewriter(
g, "eightbit", quantized_input_range=None)
eightbit_graph_def = eightbit_rewriter.rewrite(["matmul_2"])
ops = [node.op for node in eightbit_graph_def.node]
# No quantize since all inputs are const and can be quantized up-front.
self.assertEqual(0, ops.count("QuantizeV2") + ops.count("Quantize"))
self.assertEqual(1, ops.count("QuantizedReshape"))
# One dequantize at the end.
self.assertEqual(1, ops.count("Dequantize"))
def test_non_float_reshape(self):
a = quantize_graph.create_constant_node(
"a", value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=tf.int32, shape=[2, 2, 3])
shape = quantize_graph.create_constant_node(
"shape", value=[12], dtype=tf.int32, shape=[1])
reshape = quantize_graph.create_node(
"Reshape", "reshape", [a.name, shape.name])
quantize_graph.set_attr_dtype(reshape, "T", tf.int32)
g = tf.GraphDef()
g.node.extend([a, shape, reshape])
test_graph(g, {}, [reshape.name])
def test_non_float_reshape(self):
a = quantize_graph.create_constant_node(
"a",
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=dtypes.int32,
shape=[2, 2, 3])
shape = quantize_graph.create_constant_node(
"shape", value=[12], dtype=dtypes.int32, shape=[1])
reshape = quantize_graph.create_node("Reshape", "reshape",
[a.name, shape.name])
quantize_graph.set_attr_dtype(reshape, "T", dtypes.int32)
g = graph_pb2.GraphDef()
g.node.extend([a, shape, reshape])
test_graph(g, {}, [reshape.name])
def test_bias_add_w_fake_quant_w_min_max_vars(self):
input_node = quantize_graph.create_constant_node(
"input",
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
dtype=dtypes.float32,
shape=[1, 1, 2, 5])
offset_node = quantize_graph.create_constant_node(
"offset", value=[1, 2, 3, 4, 5], dtype=dtypes.float32, shape=[5])
bias_add_node = quantize_graph.create_node(
"BiasAdd", "bias_add", [input_node.name, offset_node.name])
quantize_graph.set_attr_dtype(bias_add_node, "T", dtypes.float32)
min_node = quantize_graph.create_constant_node(
"min_bias_add", value=-.5, dtype=dtypes.float32, shape=[])
max_node = quantize_graph.create_constant_node(
"max_bias_add", value=15.5, dtype=dtypes.float32, shape=[])
fake_quant_node = quantize_graph.create_node(
"FakeQuantWithMinMaxVars", "fake_quant",
[bias_add_node.name, min_node.name, max_node.name])
float_graph_def = graph_pb2.GraphDef()
float_graph_def.node.extend([
input_node, offset_node, bias_add_node, min_node, max_node,
fake_quant_node
])
test_graph(float_graph_def, {}, [fake_quant_node.name], log_graph=True)
# Verify there is only one Quantize and one Requantize op.
# Pass in fallback_quantization_range, although it will have no effect
# because the FakeQuantWithMinMaxVars are used instead.
eightbit_rewriter = quantize_graph.GraphRewriter(
float_graph_def,
"eightbit",
quantized_input_range=None,
fallback_quantization_range=[-100, 100])
eightbit_graph_def = eightbit_rewriter.rewrite([fake_quant_node.name])
ops = [node.op for node in eightbit_graph_def.node]
node_names = [node.name for node in eightbit_graph_def.node]
# No quantize since all inputs are const and can be quantized up-front.
self.assertEqual(0, ops.count("QuantizeV2") + ops.count("Quantize"))
# One dequantize at the end.
self.assertEqual(1, ops.count("Dequantize"))
# The fallback constants are not in the graph.
self.assertEqual(0, node_names.count("fallback_quantization_min_value"))
self.assertEqual(0, node_names.count("fallback_quantization_max_value"))
def test_quantized_input_range_bias_add(self):
input_shape = [1, 1, 2, 6]
input_n = quantize_graph.create_node("Placeholder", "input", [])
quantize_graph.set_attr_dtype(input_n, "dtype", dtypes.float32)
quantize_graph.set_attr_shape(input_n, "shape", input_shape)
offset_n = quantize_graph.create_constant_node(
"offset",
value=[1, 2, 3, 4, 5, 6],
dtype=dtypes.float32,
shape=[6])
bias_add_n = quantize_graph.create_node("BiasAdd", "bias_add",
[input_n.name, offset_n.name])
quantize_graph.set_attr_dtype(bias_add_n, "T", dtypes.float32)
float_graph_def = graph_pb2.GraphDef()
float_graph_def.node.extend([input_n, offset_n, bias_add_n])
input_map = {
input_n.name + ":0":
np.reshape([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12], input_shape)
}
self._RunTestsForQuantizedInputRange(float_graph_def, input_map,
[bias_add_n.name], [-1, 20.])
self._RunTestsForQuantizedInputRange(float_graph_def, input_map,
[bias_add_n.name], [0, 12.])
def test_negative_const_problem(self):
shape_constant_name = "shape_constant"
shape_constant = quantize_graph.create_constant_node(
shape_constant_name, value=-0.8, dtype=dtypes.float32, shape=[1])
quantization_result = quantize_graph.quantize_weight_eightbit(
shape_constant, b"MIN_COMBINED")
self.assertEqual(4, len(quantization_result))
def test_batch_norm(self):
input_constant_name = "input_constant"
mean_constant_name = "mean_constant"
variance_constant_name = "variance_constant"
beta_constant_name = "beta_constant"
gamma_constant_name = "gamma_constant"
batch_norm_name = "batch_norm"
float_graph_def = graph_pb2.GraphDef()
input_constant = quantize_graph.create_constant_node(
input_constant_name,
value=[1, 4, 2, 5, 3, 6, -1, -4, -2, -5, -3, -6],
dtype=dtypes.float32,
shape=[1, 1, 6, 2])
float_graph_def.node.extend([input_constant])
mean_constant = quantize_graph.create_constant_node(
mean_constant_name,
value=[10, 20],
dtype=dtypes.float32,
shape=[2])
float_graph_def.node.extend([mean_constant])
variance_constant = quantize_graph.create_constant_node(
variance_constant_name,
value=[0.25, 0.5],
dtype=dtypes.float32,
shape=[2])
float_graph_def.node.extend([variance_constant])
beta_constant = quantize_graph.create_constant_node(
beta_constant_name,
value=[0.1, 0.6],
dtype=dtypes.float32,
shape=[2])
float_graph_def.node.extend([beta_constant])
gamma_constant = quantize_graph.create_constant_node(
gamma_constant_name, value=[0, 0], dtype=dtypes.float32, shape=[2])
float_graph_def.node.extend([gamma_constant])
batch_norm_node = quantize_graph.create_node(
"BatchNormWithGlobalNormalization", batch_norm_name, [
input_constant_name, mean_constant_name,
variance_constant_name, beta_constant_name, gamma_constant_name
])
quantize_graph.set_attr_dtype(batch_norm_node, "T", dtypes.float32)
quantize_graph.set_attr_bool(batch_norm_node,
"scale_after_normalization", False)
quantize_graph.set_attr_float(batch_norm_node, "variance_epsilon",
0.001)
float_graph_def.node.extend([batch_norm_node])
test_graph(float_graph_def, {}, [batch_norm_name])
def test_non_float_concat(self):
concat_dim = quantize_graph.create_constant_node(
"concat_dim", value=0, dtype=tf.int32, shape=[])
a = quantize_graph.create_constant_node(
"a", value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=tf.int32, shape=[2, 2, 3])
b = quantize_graph.create_constant_node(
"b", value=[13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24],
dtype=tf.int32, shape=[2, 2, 3])
concat = quantize_graph.create_node(
"Concat", "concat", [concat_dim.name, a.name, b.name])
quantize_graph.set_attr_int(concat, "N", 2)
quantize_graph.set_attr_dtype(concat, "T", tf.int32)
g = tf.GraphDef()
g.node.extend([concat_dim, a, b, concat])
test_graph(g, {}, [concat.name])
def test_batch_norm(self):
input_constant_name = "input_constant"
mean_constant_name = "mean_constant"
variance_constant_name = "variance_constant"
beta_constant_name = "beta_constant"
gamma_constant_name = "gamma_constant"
batch_norm_name = "batch_norm"
float_graph_def = tf.GraphDef()
input_constant = quantize_graph.create_constant_node(input_constant_name,
value=[1, 4, 2, 5, 3,
6, -1, -4, -2,
-5, -3, -6],
dtype=tf.float32,
shape=[1, 1, 6, 2])
float_graph_def.node.extend([input_constant])
mean_constant = quantize_graph.create_constant_node(mean_constant_name,
value=[10, 20],
dtype=tf.float32,
shape=[2])
float_graph_def.node.extend([mean_constant])
variance_constant = quantize_graph.create_constant_node(
variance_constant_name, value=[0.25, 0.5], dtype=tf.float32, shape=[2])
float_graph_def.node.extend([variance_constant])
beta_constant = quantize_graph.create_constant_node(beta_constant_name,
value=[0.1, 0.6],
dtype=tf.float32,
shape=[2])
float_graph_def.node.extend([beta_constant])
gamma_constant = quantize_graph.create_constant_node(gamma_constant_name,
value=[0, 0],
dtype=tf.float32,
shape=[2])
float_graph_def.node.extend([gamma_constant])
batch_norm_node = quantize_graph.create_node(
"BatchNormWithGlobalNormalization", batch_norm_name,
[input_constant_name, mean_constant_name, variance_constant_name,
beta_constant_name, gamma_constant_name])
quantize_graph.set_attr_dtype(batch_norm_node, "T", tf.float32)
quantize_graph.set_attr_bool(batch_norm_node, "scale_after_normalization",
False)
quantize_graph.set_attr_float(batch_norm_node, "variance_epsilon", 0.001)
float_graph_def.node.extend([batch_norm_node])
test_graph(float_graph_def, {}, [batch_norm_name])
def test_bias_add_w_fallback_min_max_vars(self):
input_node = quantize_graph.create_constant_node(
"input",
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10],
dtype=tf.float32,
shape=[1, 1, 2, 5])
offset_node = quantize_graph.create_constant_node(
"offset", value=[1, 2, 3, 4, 5], dtype=tf.float32, shape=[5])
bias_add_node = quantize_graph.create_node(
"BiasAdd", "bias_add", [input_node.name, offset_node.name])
quantize_graph.set_attr_dtype(bias_add_node, "T", tf.float32)
float_graph_def = tf.GraphDef()
float_graph_def.node.extend([input_node, offset_node, bias_add_node])
test_graph(float_graph_def, {}, [bias_add_node.name], log_graph=True)
# Verify there is only one Quantize, one Requantize op, and no
# RequantizationRange op.
eightbit_rewriter = quantize_graph.GraphRewriter(
float_graph_def,
"eightbit",
quantized_input_range=None,
fallback_quantization_range=[-.5, 15.5])
eightbit_graph_def = eightbit_rewriter.rewrite([bias_add_node.name])
ops = [node.op for node in eightbit_graph_def.node]
node_names = [node.name for node in eightbit_graph_def.node]
# No quantize since all inputs are const and can be quantized up-front.
self.assertEqual(0, ops.count("QuantizeV2") + ops.count("Quantize"))
# One dequantize at the end.
self.assertEqual(1, ops.count("Dequantize"))
# No RequantizationRange
self.assertEqual(0, ops.count("RequantizationRange"))
# The fallback constants are in the graph.
self.assertEqual(1,
node_names.count("fallback_quantization_min_value"))
self.assertEqual(1,
node_names.count("fallback_quantization_max_value"))
def test_concat(self):
shape_constant_name = "shape_constant"
a_constant_name = "a_constant"
b_constant_name = "b_constant"
concat_name = "concat"
float_graph_def = tf.GraphDef()
shape_constant = quantize_graph.create_constant_node(shape_constant_name,
value=0,
dtype=tf.int32,
shape=[])
float_graph_def.node.extend([shape_constant])
a_constant = quantize_graph.create_constant_node(a_constant_name,
value=[1, 2, 3, 4, 5, 6, 7,
8, 9, 10, 11, 12],
dtype=tf.float32,
shape=[2, 2, 3])
float_graph_def.node.extend([a_constant])
b_constant = quantize_graph.create_constant_node(b_constant_name,
value=[13, 14, 15, 16, 17,
18, 19, 20, 21, 22,
23, 24],
dtype=tf.float32,
shape=[2, 2, 3])
float_graph_def.node.extend([b_constant])
concat_node = quantize_graph.create_node("Concat", concat_name,
[shape_constant_name,
a_constant_name, b_constant_name])
quantize_graph.set_attr_int(concat_node, "N", 2)
quantize_graph.set_attr_dtype(concat_node, "T", tf.float32)
float_graph_def.node.extend([concat_node])
test_graph(float_graph_def, {}, [concat_name])
# Verify the concat is quantized.
eightbit_rewriter = quantize_graph.GraphRewriter(
float_graph_def, "eightbit", quantized_input_range=None)
eightbit_graph_def = eightbit_rewriter.rewrite([concat_name])
ops = [node.op for node in eightbit_graph_def.node]
self.assertEqual(1, ops.count("QuantizedConcat"))
def test_non_float_concat(self):
concat_dim = quantize_graph.create_constant_node(
"concat_dim", value=0, dtype=dtypes.int32, shape=[])
a = quantize_graph.create_constant_node(
"a",
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=dtypes.int32,
shape=[2, 2, 3])
b = quantize_graph.create_constant_node(
"b",
value=[13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24],
dtype=dtypes.int32,
shape=[2, 2, 3])
concat = quantize_graph.create_node("Concat", "concat",
[concat_dim.name, a.name, b.name])
quantize_graph.set_attr_int(concat, "N", 2)
quantize_graph.set_attr_dtype(concat, "T", dtypes.int32)
g = graph_pb2.GraphDef()
g.node.extend([concat_dim, a, b, concat])
test_graph(g, {}, [concat.name])
def test_bias_add(self):
input_constant_name = "input_constant"
offset_constant_name = "offset_constant"
bias_add_name = "bias_add"
float_graph_def = graph_pb2.GraphDef()
input_constant = quantize_graph.create_constant_node(
input_constant_name,
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=dtypes.float32,
shape=[1, 1, 2, 6])
float_graph_def.node.extend([input_constant])
offset_constant = quantize_graph.create_constant_node(
offset_constant_name,
value=[1, 2, 3, 4, 5, 6],
dtype=dtypes.float32,
shape=[6])
float_graph_def.node.extend([offset_constant])
bias_add_node = quantize_graph.create_node(
"BiasAdd", bias_add_name, [input_constant_name, offset_constant_name])
quantize_graph.set_attr_dtype(bias_add_node, "T", dtypes.float32)
float_graph_def.node.extend([bias_add_node])
test_graph(float_graph_def, {}, [bias_add_name])
def test_concat(self):
shape_constant_name = "shape_constant"
a_constant_name = "a_constant"
b_constant_name = "b_constant"
concat_name = "concat"
float_graph_def = graph_pb2.GraphDef()
shape_constant = quantize_graph.create_constant_node(
shape_constant_name, value=0, dtype=dtypes.int32, shape=[])
float_graph_def.node.extend([shape_constant])
a_constant = quantize_graph.create_constant_node(
a_constant_name,
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=dtypes.float32,
shape=[2, 2, 3])
float_graph_def.node.extend([a_constant])
b_constant = quantize_graph.create_constant_node(
b_constant_name,
value=[13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24],
dtype=dtypes.float32,
shape=[2, 2, 3])
float_graph_def.node.extend([b_constant])
concat_node = quantize_graph.create_node(
"Concat", concat_name,
[shape_constant_name, a_constant_name, b_constant_name])
quantize_graph.set_attr_int(concat_node, "N", 2)
quantize_graph.set_attr_dtype(concat_node, "T", dtypes.float32)
float_graph_def.node.extend([concat_node])
test_graph(float_graph_def, {}, [concat_name])
# Verify the concat is quantized.
eightbit_rewriter = quantize_graph.GraphRewriter(
float_graph_def, "eightbit", quantized_input_range=None)
eightbit_graph_def = eightbit_rewriter.rewrite([concat_name])
ops = [node.op for node in eightbit_graph_def.node]
self.assertEqual(1, ops.count("QuantizedConcat"))
def test_relu6(self):
input_constant_name = "input_constant"
relu6_name = "relu6"
float_graph_def = graph_pb2.GraphDef()
input_constant = quantize_graph.create_constant_node(
input_constant_name,
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=dtypes.float32,
shape=[1, 2, 6, 1])
float_graph_def.node.extend([input_constant])
relu6_node = quantize_graph.create_node("Relu6", relu6_name,
[input_constant_name])
quantize_graph.set_attr_dtype(relu6_node, "T", dtypes.float32)
float_graph_def.node.extend([relu6_node])
test_graph(float_graph_def, {}, [relu6_name])
def test_max_pool(self):
input_constant_name = "input_constant"
max_pool_name = "max_pool"
float_graph_def = graph_pb2.GraphDef()
input_constant = quantize_graph.create_constant_node(
input_constant_name,
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=dtypes.float32,
shape=[1, 2, 6, 1])
float_graph_def.node.extend([input_constant])
max_pool_node = quantize_graph.create_node("MaxPool", max_pool_name,
[input_constant_name])
quantize_graph.set_attr_int_list(max_pool_node, "ksize", [1, 2, 2, 1])
quantize_graph.set_attr_int_list(max_pool_node, "strides", [1, 1, 1, 1])
quantize_graph.set_attr_string(max_pool_node, "padding", b"SAME")
float_graph_def.node.extend([max_pool_node])
test_graph(float_graph_def, {}, [max_pool_name])
def test_avg_pool(self):
input_constant_name = "input_constant"
avg_pool_name = "avg_pool"
float_graph_def = tf.GraphDef()
input_constant = quantize_graph.create_constant_node(
input_constant_name,
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=tf.float32,
shape=[1, 2, 6, 1])
float_graph_def.node.extend([input_constant])
avg_pool_node = quantize_graph.create_node("AvgPool", avg_pool_name,
[input_constant_name])
quantize_graph.set_attr_dtype(avg_pool_node, "T", tf.float32)
quantize_graph.set_attr_int_list(avg_pool_node, "ksize", [1, 2, 2, 1])
quantize_graph.set_attr_int_list(avg_pool_node, "strides",
[1, 1, 1, 1])
quantize_graph.set_attr_string(avg_pool_node, "padding", b"SAME")
float_graph_def.node.extend([avg_pool_node])
test_graph(float_graph_def, {}, [avg_pool_name])
def test_avg_pool(self):
input_constant_name = "input_constant"
avg_pool_name = "avg_pool"
float_graph_def = tf.GraphDef()
input_constant = quantize_graph.create_constant_node(input_constant_name,
value=[1, 2, 3, 4, 5,
6, 7, 8, 9, 10,
11, 12],
dtype=tf.float32,
shape=[1, 2, 6, 1])
float_graph_def.node.extend([input_constant])
avg_pool_node = quantize_graph.create_node("AvgPool", avg_pool_name,
[input_constant_name])
quantize_graph.set_attr_dtype(avg_pool_node, "T", tf.float32)
quantize_graph.set_attr_int_list(avg_pool_node, "ksize", [1, 2, 2, 1])
quantize_graph.set_attr_int_list(avg_pool_node, "strides", [1, 1, 1, 1])
quantize_graph.set_attr_string(avg_pool_node, "padding", b"SAME")
float_graph_def.node.extend([avg_pool_node])
test_graph(float_graph_def, {}, [avg_pool_name])
def test_quantized_input_range_bias_add(self):
input_shape = [1, 1, 2, 6]
input_n = quantize_graph.create_node("PlaceholderV2", "input", [])
quantize_graph.set_attr_dtype(input_n, "dtype", dtypes.float32)
quantize_graph.set_attr_shape(input_n, "shape", input_shape)
offset_n = quantize_graph.create_constant_node(
"offset", value=[1, 2, 3, 4, 5, 6], dtype=dtypes.float32, shape=[6])
bias_add_n = quantize_graph.create_node("BiasAdd", "bias_add",
[input_n.name, offset_n.name])
quantize_graph.set_attr_dtype(bias_add_n, "T", dtypes.float32)
float_graph_def = graph_pb2.GraphDef()
float_graph_def.node.extend([input_n, offset_n, bias_add_n])
input_map = {
input_n.name + ":0":
np.reshape([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12], input_shape)
}
self._RunTestsForQuantizedInputRange(float_graph_def, input_map,
[bias_add_n.name], [-1, 20.])
self._RunTestsForQuantizedInputRange(float_graph_def, input_map,
[bias_add_n.name], [0, 12.])
def test_identity(self):
input_constant_name = "input_constant"
identity_name = "identity"
float_graph_def = graph_pb2.GraphDef()
input_constant = quantize_graph.create_constant_node(
input_constant_name,
value=[1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12],
dtype=dtypes.float32,
shape=[2, 6])
float_graph_def.node.extend([input_constant])
identity_node = quantize_graph.create_node("Identity", identity_name,
[input_constant_name])
quantize_graph.set_attr_dtype(identity_node, "T", dtypes.float32)
float_graph_def.node.extend([identity_node])
mul_name = "mul"
mul_node = quantize_graph.create_node("Mul", mul_name,
[identity_name, identity_name])
quantize_graph.set_attr_dtype(mul_node, "T", dtypes.float32)
float_graph_def.node.extend([mul_node])
test_graph(float_graph_def, {}, [mul_name])
def test_remove_redundant_quantization(self):
a_constant_name = "a_constant"
a_constant_min_name = "a_constant_min"
a_constant_max_name = "a_constant_max"
a_dequantize_name = "a_dequantize"
a_quantize_name = "a_quantize"
b_constant_name = "b_constant"
b_constant_min_name = "b_constant_min"
b_constant_max_name = "b_constant_max"
b_dequantize_name = "b_dequantize"
b_quantize_name = "b_quantize"
mat_mul_name = "mat_mul"
graph_def = graph_pb2.GraphDef()
a_constant = quantize_graph.create_constant_node(
a_constant_name, value=(0,), dtype=dtypes.quint8, shape=[])
graph_def.node.extend([a_constant])
a_constant_min = quantize_graph.create_constant_node(
a_constant_min_name, value=2, dtype=dtypes.float32, shape=[])
graph_def.node.extend([a_constant_min])
a_constant_max = quantize_graph.create_constant_node(
a_constant_max_name, value=2, dtype=dtypes.float32, shape=[])
graph_def.node.extend([a_constant_max])
a_dequantize_node = quantize_graph.create_node(
"Dequantize", a_dequantize_name,
[a_constant_name, a_constant_min_name, a_constant_max_name])
quantize_graph.set_attr_dtype(a_dequantize_node, "T", dtypes.uint8)
graph_def.node.extend([a_dequantize_node])
a_quantize_node = quantize_graph.create_node(
"QuantizeV2", a_quantize_name,
[a_dequantize_name, a_dequantize_name + ":1", a_dequantize_name + ":2"])
quantize_graph.set_attr_dtype(a_quantize_node, "T", dtypes.uint8)
graph_def.node.extend([a_quantize_node])
b_constant = quantize_graph.create_constant_node(
b_constant_name, value=(0,), dtype=dtypes.quint8, shape=[])
graph_def.node.extend([b_constant])
b_constant_min = quantize_graph.create_constant_node(
b_constant_min_name, value=3, dtype=dtypes.float32, shape=[])
graph_def.node.extend([b_constant_min])
b_constant_max = quantize_graph.create_constant_node(
b_constant_max_name, value=3, dtype=dtypes.float32, shape=[])
graph_def.node.extend([b_constant_max])
b_dequantize_node = quantize_graph.create_node(
"Dequantize", b_dequantize_name,
[b_constant_name, b_constant_min_name, b_constant_max_name])
quantize_graph.set_attr_dtype(b_dequantize_node, "T", dtypes.uint8)
graph_def.node.extend([b_dequantize_node])
b_quantize_node = quantize_graph.create_node(
"QuantizeV2", b_quantize_name,
[b_dequantize_name, b_dequantize_name + ":1", b_dequantize_name + ":2"])
quantize_graph.set_attr_dtype(b_quantize_node, "T", dtypes.uint8)
graph_def.node.extend([b_quantize_node])
mat_mul_node = quantize_graph.create_node("QuantizedMatMul", mat_mul_name, [
a_quantize_name, b_quantize_name, a_quantize_name + ":1",
a_quantize_name + ":2", b_quantize_name + ":1", b_quantize_name + ":2"
])
quantize_graph.set_attr_dtype(mat_mul_node, "T1", dtypes.uint8)
quantize_graph.set_attr_dtype(mat_mul_node, "T2", dtypes.int32)
graph_def.node.extend([mat_mul_node])
expected_output = graph_pb2.GraphDef()
a_constant = quantize_graph.create_constant_node(
a_constant_name, value=(0,), dtype=dtypes.quint8, shape=[])
expected_output.node.extend([a_constant])
a_constant_min = quantize_graph.create_constant_node(
a_constant_min_name, value=2, dtype=dtypes.float32, shape=[])
expected_output.node.extend([a_constant_min])
a_constant_max = quantize_graph.create_constant_node(
a_constant_max_name, value=2, dtype=dtypes.float32, shape=[])
expected_output.node.extend([a_constant_max])
b_constant = quantize_graph.create_constant_node(
b_constant_name, value=(0,), dtype=dtypes.quint8, shape=[])
expected_output.node.extend([b_constant])
b_constant_min = quantize_graph.create_constant_node(
b_constant_min_name, value=3, dtype=dtypes.float32, shape=[])
expected_output.node.extend([b_constant_min])
b_constant_max = quantize_graph.create_constant_node(
b_constant_max_name, value=3, dtype=dtypes.float32, shape=[])
expected_output.node.extend([b_constant_max])
mat_mul_node = quantize_graph.create_node("QuantizedMatMul", mat_mul_name, [
a_constant_name, b_constant_name, a_constant_min_name,
a_constant_max_name, b_constant_min_name, b_constant_max_name
])
quantize_graph.set_attr_dtype(mat_mul_node, "T1", dtypes.uint8)
quantize_graph.set_attr_dtype(mat_mul_node, "T2", dtypes.int32)
expected_output.node.extend([mat_mul_node])
expected_output.versions.CopyFrom(graph_def.versions)
expected_output.library.CopyFrom(graph_def.library)
rewriter = quantize_graph.GraphRewriter(
graph_def, [mat_mul_name], quantized_input_range=None)
output = rewriter.remove_redundant_quantization(graph_def)
stripped_output = graph_util.extract_sub_graph(output, [mat_mul_name])
self.assertProtoEquals(expected_output, stripped_output)
