def test_invalid_inputs(self):
inputs = constant_op.constant(value=[[1.0], [2.0], [4.0]],
dtype=dtypes.float32)
with self.assertRaisesRegex((ValueError, errors.InvalidArgumentError),
"must be rank 1"):
self.evaluate(
array_ops.fake_quant_with_min_max_vars_per_channel(
inputs=inputs, min=[[0.0]], max=[1.0]))
with self.assertRaisesRegex((ValueError, errors.InvalidArgumentError),
"Dimensions must be equal|incorrect size"):
self.evaluate(
array_ops.fake_quant_with_min_max_vars_per_channel(
inputs=inputs, min=[0.0, 0.1], max=[1.0]))
with self.assertRaisesRegex((ValueError, errors.InvalidArgumentError),
"must be rank 1"):
self.evaluate(
array_ops.fake_quant_with_min_max_vars_per_channel(
inputs=inputs, min=[1.0], max=[[1.0]]))
with self.assertRaisesRegex((ValueError, errors.InvalidArgumentError),
"Dimensions must be equal|incorrect size"):
self.evaluate(
array_ops.fake_quant_with_min_max_vars_per_channel(
inputs=inputs, min=[0.0], max=[1.0, 1.1]))
def _FakeQuantWithMinMaxVars(inputs, min_var, max_var, per_channel, num_bits,
narrow_range):
"""Adds a fake quantization operation.
Depending on value of per_channel, this operation may do global quantization
or per channel quantization. min_var and max_var should have corresponding
shapes: [1] when per_channel == False and [d] when per_channel == True.
Args:
inputs: a tensor containing values to be quantized.
min_var: a variable containing quantization range lower end(s).
max_var: a variable containing quantization range upper end(s).
per_channel: a boolean specifying whether to use per-channel quantization.
num_bits: Number of bits to use for quantization, must be between 2 and 8.
narrow_range: Whether to use the narrow quantization range
[1; 2^num_bits - 1] or wide range [0; 2^num_bits - 1].
Returns:
a tensor containing quantized values.
"""
if per_channel:
assert len(min_var.get_shape()) == 1
assert len(max_var.get_shape()) == 1
return array_ops.fake_quant_with_min_max_vars_per_channel(
inputs, min_var, max_var, num_bits=num_bits, narrow_range=narrow_range)
else:
assert min_var.get_shape() == [] # pylint: disable=g-explicit-bool-comparison
assert max_var.get_shape() == [] # pylint: disable=g-explicit-bool-comparison
return array_ops.fake_quant_with_min_max_vars(
inputs, min_var, max_var, num_bits=num_bits, narrow_range=narrow_range)
def _FakeQuantWithMinMaxVars(inputs, min_var, max_var, per_channel, num_bits,
narrow_range):
"""Adds a fake quantization operation.
Depending on value of per_channel, this operation may do global quantization
or per channel quantization. min_var and max_var should have corresponding
shapes: [1] when per_channel == False and [d] when per_channel == True.
Args:
inputs: a tensor containing values to be quantized.
min_var: a variable containing quantization range lower end(s).
max_var: a variable containing quantization range upper end(s).
per_channel: a boolean specifying whether to use per-channel quantization.
num_bits: Number of bits to use for quantization, must be between 2 and 8.
narrow_range: Whether to use the narrow quantization range
[1; 2^num_bits - 1] or wide range [0; 2^num_bits - 1].
Returns:
a tensor containing quantized values.
"""
if per_channel:
assert len(min_var.get_shape()) == 1
assert len(max_var.get_shape()) == 1
return array_ops.fake_quant_with_min_max_vars_per_channel(
inputs, min_var, max_var, num_bits=num_bits, narrow_range=narrow_range)
else:
assert min_var.get_shape() == [] # pylint: disable=g-explicit-bool-comparison
assert max_var.get_shape() == [] # pylint: disable=g-explicit-bool-comparison
return array_ops.fake_quant_with_min_max_vars(
inputs, min_var, max_var, num_bits=num_bits, narrow_range=narrow_range)
def _TestOp(self, input_mins, input_maxs, num_bits, narrow_range,
expected_nudged_input_mins, expected_nudged_input_maxs,
expected_steps):
num_channels = len(input_mins)
inputs_list = []
expected_list = []
for i in range(num_channels):
expected_nudged_input_min = expected_nudged_input_mins[i]
expected_nudged_input_max = expected_nudged_input_maxs[i]
expected_step = expected_steps[i]
inputs_list.append([
expected_nudged_input_min - expected_step,
expected_nudged_input_min - 0.01, expected_nudged_input_min,
expected_nudged_input_min + 0.01,
expected_nudged_input_min + expected_step - 0.01,
expected_nudged_input_min + expected_step,
expected_nudged_input_min + expected_step + 0.01,
expected_nudged_input_max - 0.01, expected_nudged_input_max,
expected_nudged_input_max + 0.01,
expected_nudged_input_max + expected_step
])
expected_list.append([
expected_nudged_input_min, expected_nudged_input_min,
expected_nudged_input_min, expected_nudged_input_min,
expected_nudged_input_min + expected_step,
expected_nudged_input_min + expected_step,
expected_nudged_input_min + expected_step,
expected_nudged_input_max, expected_nudged_input_max,
expected_nudged_input_max, expected_nudged_input_max
])
inputs = np.transpose(np.array(inputs_list, dtype=np.float32))
expected = np.transpose(np.array(expected_list, dtype=np.float32))
with self.session() as session:
with self.test_scope():
input_placeholder = array_ops.placeholder(dtypes.float32,
inputs.shape,
name="inputs")
min_placeholder = array_ops.placeholder(dtypes.float32,
(num_channels),
name="min")
max_placeholder = array_ops.placeholder(dtypes.float32,
(num_channels),
name="max")
outputs = array_ops.fake_quant_with_min_max_vars_per_channel(
input_placeholder,
min_placeholder,
max_placeholder,
num_bits=num_bits,
narrow_range=narrow_range)
result = session.run(
outputs, {
input_placeholder: inputs,
min_placeholder: input_mins,
max_placeholder: input_maxs
})
self.assertAllCloseAccordingToType(result,
expected,
rtol=1e-3,
atol=1e-5,
bfloat16_rtol=0.03)
