def test_quantized_input_range_errors(self):
with self.assertRaises(ValueError):
# Invalid mode.
quantize_graph.GraphRewriter(graph_pb2.GraphDef(),
"weights_rounded", [0, 1])
with self.assertRaises(ValueError):
# Invalid range.
quantize_graph.GraphRewriter(graph_pb2.GraphDef(), "eightbit",
[0, -1])
def _RunTestsForQuantizedInputRange(self, float_graph_def, input_map,
output_names, input_range):
if sys.version_info[0] == 3:
# uint8->quint8 conversion for numpy is not working currently.
return
quantized_input_map = {}
for k, v in input_map.items():
arr = [
int(
round((n - input_range[0]) * 255 /
(input_range[1] - input_range[0]))) for n in v.flat
]
arr = np.array(arr, np.uint8)
arr = arr.reshape(v.shape)
arr = arr.astype(dtypes.quint8.as_numpy_dtype)
quantized_input_map[k] = arr
output_tensors = [output_name + ":0" for output_name in output_names]
float_results = run_graph_def(float_graph_def, input_map,
output_tensors)
# Quantize treating the input as quantized in range <input_range>.
rewriter = quantize_graph.GraphRewriter(float_graph_def, "eightbit",
input_range)
graph_def = rewriter.rewrite(output_names)
results = run_graph_def(graph_def, quantized_input_map, output_tensors)
for expected, result in zip(float_results, results):
assert are_tensors_near(expected, result, .5)
ops = [node.op for node in graph_def.node]
self.assertEqual(0, ops.count("QuantizeV2") + ops.count("Quantize"))
self.assertEqual(len(output_names), ops.count("Dequantize"))
# Quantize without treating input as quantized.
rewriter = quantize_graph.GraphRewriter(float_graph_def,
"eightbit",
quantized_input_range=None)
graph_def = rewriter.rewrite(output_names)
results = run_graph_def(graph_def, input_map, output_tensors)
for expected, result in zip(float_results, results):
assert are_tensors_near(expected, result, .5)
ops = [node.op for node in graph_def.node]
self.assertEqual(len(input_map),
ops.count("QuantizeV2") + ops.count("Quantize"))
self.assertEqual(len(output_names), 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=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
])
graph_test(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
graph_test(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 graph_test(float_graph_def, input_map, output_names, log_graph=False):
"""Runs the float graph through the rewriter and tests the results."""
float_results = run_graph_def(
float_graph_def, input_map,
[output_name + ":0" for output_name in output_names])
# TODO(petewarden): round test is currently failing because there is no
# RoundToSteps op available.
# round_rewriter = quantize_graph.GraphRewriter(float_graph_def, "round")
# round_graph_def = round_rewriter.rewrite(output_name)
# round_results = run_graph_def(round_graph_def, input_map,
# [output_name + ":0"])
# assert are_tensors_near(expected, round_results[0], 1.0)
#
# TODO(petewarden): Add test for "quantize" mode.
eightbit_rewriter = quantize_graph.GraphRewriter(
float_graph_def, "eightbit", quantized_input_range=None)
eightbit_graph_def = eightbit_rewriter.rewrite(output_names)
eightbit_results = run_graph_def(
eightbit_graph_def, input_map,
[output_name + ":0" for output_name in output_names])
for expected, result in zip(float_results, eightbit_results):
assert are_tensors_near(expected, result, 1.0)
if log_graph:
tf_logging.info("8bit:\n%s", str(eightbit_graph_def))
# Test the weights_rounded mode. This uses the default bit_depth.
weights_rounded_rewriter = quantize_graph.GraphRewriter(
float_graph_def, "weights_rounded", quantized_input_range=None)
weights_rounded_graph_def = weights_rounded_rewriter.rewrite(output_names)
weights_rounded_results = run_graph_def(
weights_rounded_graph_def, input_map,
[output_name + ":0" for output_name in output_names])
for expected, result in zip(float_results, weights_rounded_results):
assert are_tensors_near(expected, result, 1.0)
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=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)
float_graph_def = graph_pb2.GraphDef()
float_graph_def.node.extend([input_node, offset_node, bias_add_node])
graph_test(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_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])
graph_test(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_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])
graph_test(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_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)