def __init__(self,
dtype=torch.quint8,
qscheme=torch.per_tensor_affine,
quant_min=None,
quant_max=None,
output_obs=None,
factory_kwargs=None) -> None:
super(_InputEqualizationObserver, self).__init__()
if qscheme not in {
torch.per_tensor_affine, torch.per_tensor_symmetric
}:
raise TypeError("Input qscheme must be per-tensor")
self.input_obs = PerChannelMinMaxObserver(
ch_axis=1,
dtype=dtype,
qscheme=qscheme,
quant_min=quant_min,
quant_max=quant_max,
factory_kwargs=factory_kwargs)
if output_obs is None:
self.output_obs = MinMaxObserver(dtype=dtype,
qscheme=qscheme,
quant_min=quant_min,
quant_max=quant_max,
factory_kwargs=factory_kwargs)
else:
self.output_obs = output_obs
self.equalization_scale = torch.empty(0)
def __init__(self,
dtype=torch.qint8,
qscheme=torch.per_tensor_affine,
quant_min=None,
quant_max=None,
factory_kwargs=None) -> None:
super(_WeightEqualizationObserver, self).__init__()
self.weight_col_obs = PerChannelMinMaxObserver(
ch_axis=1,
dtype=dtype,
qscheme=qscheme,
quant_min=quant_min,
quant_max=quant_max,
factory_kwargs=factory_kwargs)
self.weight_row_obs = PerChannelMinMaxObserver(
ch_axis=0,
dtype=dtype,
qscheme=qscheme,
quant_min=quant_min,
quant_max=quant_max,
factory_kwargs=factory_kwargs)
self.equalization_scale = torch.empty(0)
def test_input_weight_eq_observer(self, ndim, input_qdtype, input_qscheme, weight_qdtype, weight_qscheme):
sizes = []
for _ in range((ndim - 1) * 2):
sizes.append(np.random.randint(2, 10))
channel = np.random.randint(1, 10)
if ndim == 2:
x = np.random.random(size=(sizes[0], channel))
w = np.random.random(size=(sizes[1], channel))
elif ndim == 3:
x = np.random.random(size=(sizes[0], channel, sizes[1]))
w = np.random.random(size=(sizes[2], channel, sizes[3]))
elif ndim == 4:
x = np.random.random(size=(sizes[0], channel, sizes[1], sizes[2]))
w = np.random.random(size=(sizes[3], channel, sizes[4], sizes[5]))
elif ndim == 5:
x = np.random.random(size=(sizes[0], channel, sizes[1], sizes[2], sizes[3]))
w = np.random.random(size=(sizes[4], channel, sizes[5], sizes[6], sizes[7]))
x = (x * 10).round(decimals=2).astype(np.float32)
w = (w * 10).round(decimals=2).astype(np.float32)
input_eq_obs = _InputEqualizationObserver(dtype=input_qdtype, qscheme=input_qscheme)
weight_eq_obs = _WeightEqualizationObserver(dtype=weight_qdtype, qscheme=weight_qscheme)
ret_x = input_eq_obs(torch.tensor(x))
ret_w = weight_eq_obs(torch.tensor(w))
self.assertEqual((ret_x, ret_w), (x, w))
# Check the min/max input columns are correct
ref_min_inputs, ref_max_inputs = self.channel_minmax(x)
min_inputs, max_inputs = input_eq_obs.get_input_minmax()
self.assertEqual(min_inputs, torch.tensor(ref_min_inputs, dtype=torch.float32))
self.assertEqual(max_inputs, torch.tensor(ref_max_inputs, dtype=torch.float32))
# Check the min/max weight columns are correct
ref_min_weights_col, ref_max_weights_col = self.channel_minmax(w)
min_weights_col, max_weights_col = weight_eq_obs.get_weight_col_minmax()
self.assertEqual(min_weights_col, torch.tensor(ref_min_weights_col, dtype=torch.float32))
self.assertEqual(max_weights_col, torch.tensor(ref_max_weights_col, dtype=torch.float32))
# Check the equalization scale is correct
equalization_scale = calculate_equalization_scale(input_eq_obs, weight_eq_obs)
ref_equalization_scale = np.sqrt((ref_max_weights_col - ref_min_weights_col) /
(ref_max_inputs - ref_min_inputs))
self.assertEqual(equalization_scale, torch.tensor(ref_equalization_scale, dtype=torch.float32))
input_eq_obs.set_equalization_scale(equalization_scale)
weight_eq_obs.set_equalization_scale(equalization_scale)
# Check the input scale/zero-point values
min_input_scaled, max_input_scaled = input_eq_obs.calculate_scaled_minmax()
input_quant_obs = MinMaxObserver(dtype=input_qdtype, qscheme=input_qscheme)
input_quant_obs.min_val = min_input_scaled
input_quant_obs.max_val = max_input_scaled
input_qparams = input_quant_obs.calculate_qparams()
ref_min_input_scaled = np.min(ref_min_inputs * ref_equalization_scale)
ref_min_input_scaled = min(0, ref_min_input_scaled)
ref_max_input_scaled = np.max(ref_max_inputs * ref_equalization_scale)
ref_max_input_scaled = max(0, ref_max_input_scaled)
if input_qscheme == torch.per_tensor_symmetric:
ref_scale = 2 * max(abs(ref_min_input_scaled), ref_max_input_scaled) / 255
ref_zero_point = 0 if input_qdtype is torch.qint8 else 128
else:
ref_scale = (ref_max_input_scaled - ref_min_input_scaled) / 255
quant_min = -128 if input_qdtype is torch.qint8 else 0
quant_max = 127 if input_qdtype is torch.qint8 else 255
ref_zero_point = quant_min - np.round(ref_min_input_scaled / ref_scale)
np.clip(ref_zero_point, quant_min, quant_max)
self.assertEqual(input_qparams[0].item(), ref_scale, atol=1e-5, rtol=0)
self.assertEqual(input_qparams[1].item(), ref_zero_point)
# During input-weight equalization, we will scale the weights so that
# the following weight quantized observer will have the correct scaled qparams
# Check the weight scale/zero-point values of the quantized observer
weight_quant_obs = PerChannelMinMaxObserver(ch_axis=1, dtype=weight_qdtype, qscheme=weight_qscheme)
# Scale the weights for input-weight equalization
new_shape = [1] * w.ndim
new_shape[1] = w.shape[1]
ref_w_scaled = w * np.reciprocal(ref_equalization_scale.reshape(tuple(new_shape)))
w = torch.tensor(w)
new_shape[1] = w.size(1)
w_scaled = torch.mul(w, torch.reciprocal(equalization_scale.view(new_shape)))
self.assertEqual(w_scaled, ref_w_scaled)
# Call forward on the weight quantization observer
weight_quant_obs(w_scaled)
# Check the min/max weight rows are correct
ref_min_weights_scaled, ref_max_weights_scaled = self.channel_minmax(ref_w_scaled)
self.assertEqual(weight_quant_obs.min_val, torch.tensor(ref_min_weights_scaled, dtype=torch.float32))
self.assertEqual(weight_quant_obs.max_val, torch.tensor(ref_max_weights_scaled, dtype=torch.float32))
weight_qparams = weight_quant_obs.calculate_qparams()
if weight_qscheme == torch.per_channel_symmetric:
ref_min_weights_scaled = np.minimum(np.zeros(ref_min_weights_scaled.shape), ref_min_weights_scaled)
ref_max_weights_scaled = np.maximum(np.zeros(ref_max_weights_scaled.shape), ref_max_weights_scaled)
ref_scales = 2 * np.maximum(np.abs(ref_min_weights_scaled), ref_max_weights_scaled) / 255
ref_zero_points = np.zeros_like(
ref_scales) if weight_qdtype is torch.qint8 else np.ones_like(ref_scales) * 128
elif weight_qscheme == torch.per_channel_affine_float_qparams:
ref_scales = (ref_max_weights_scaled - ref_min_weights_scaled) / 255
ref_scales = np.where(ref_scales > 1e-7, ref_scales, np.ones_like(ref_scales))
ref_zero_points = -1 * ref_min_weights_scaled / ref_scales
else:
ref_min_weights_scaled = np.minimum(np.zeros_like(ref_min_weights_scaled), ref_min_weights_scaled)
ref_max_weights_scaled = np.maximum(np.zeros_like(ref_max_weights_scaled), ref_max_weights_scaled)
ref_scales = (ref_max_weights_scaled - ref_min_weights_scaled) / 255
ref_zero_points = -128 if weight_qdtype is torch.qint8 else 0
ref_zero_points = ref_zero_points - np.round(ref_min_weights_scaled / ref_scales)
self.assertTrue(torch.allclose(weight_qparams[0], torch.tensor(
ref_scales, dtype=weight_qparams[0].dtype), atol=0.0001))
self.assertTrue(torch.allclose(weight_qparams[1], torch.tensor(
ref_zero_points, dtype=weight_qparams[1].dtype), atol=1))
def test_input_weight_eq_observer(self, input_qdtype, input_qscheme,
weight_qdtype, weight_qscheme):
""" Tests that the Input- and Weight- EqualizationObservers perform as expected
"""
input_eq_obs = _InputEqualizationObserver(dtype=input_qdtype,
qscheme=input_qscheme)
weight_eq_obs = _WeightEqualizationObserver(dtype=weight_qdtype,
qscheme=weight_qscheme)
width = np.random.randint(1, 10)
x_height = np.random.randint(2, 10)
w_height = np.random.randint(2, 10)
x = (np.random.random(size=(x_height, width)) *
10).round(decimals=2).astype(np.float32)
w = (np.random.random(size=(w_height, width)) *
10).round(decimals=2).astype(np.float32)
ret_x = input_eq_obs(torch.tensor(x))
ret_w = weight_eq_obs(torch.tensor(w))
self.assertEqual((ret_x, ret_w), (x, w))
# Check the min/max input columns are correct
ref_min_inputs = x.min(axis=0)
ref_max_inputs = x.max(axis=0)
self.assertEqual(input_eq_obs.get_input_minmax(),
(ref_min_inputs, ref_max_inputs))
# Check the min/max weight columns are correct
ref_min_weights_col = w.min(axis=0)
ref_max_weights_col = w.max(axis=0)
self.assertEqual(weight_eq_obs.get_weight_col_minmax(),
(ref_min_weights_col, ref_max_weights_col))
# Check the equalization scale is correct
equalization_scale = calculate_equalization_scale(
input_eq_obs, weight_eq_obs)
ref_equalization_scale = np.sqrt(
(ref_max_weights_col - ref_min_weights_col) /
(ref_max_inputs - ref_min_inputs))
self.assertEqual(equalization_scale, ref_equalization_scale)
input_eq_obs.set_equalization_scale(equalization_scale)
weight_eq_obs.set_equalization_scale(equalization_scale)
# Check the input scale/zero-point values
min_input_scaled, max_input_scaled = input_eq_obs.calculate_scaled_minmax(
)
input_quant_obs = MinMaxObserver(dtype=input_qdtype,
qscheme=input_qscheme)
input_quant_obs.min_val = min_input_scaled
input_quant_obs.max_val = max_input_scaled
input_qparams = input_quant_obs.calculate_qparams()
ref_min_input_scaled = np.min(ref_min_inputs * ref_equalization_scale)
ref_min_input_scaled = min(0, ref_min_input_scaled)
ref_max_input_scaled = np.max(ref_max_inputs * ref_equalization_scale)
ref_max_input_scaled = max(0, ref_max_input_scaled)
if input_qscheme == torch.per_tensor_symmetric:
ref_scale = 2 * max(abs(ref_min_input_scaled),
ref_max_input_scaled) / 255
ref_zero_point = 0 if input_qdtype is torch.qint8 else 128
else:
ref_scale = (ref_max_input_scaled - ref_min_input_scaled) / 255
ref_zero_point = -128 if input_qdtype is torch.qint8 else 0
self.assertEqual(input_qparams[0].item(), ref_scale, atol=1e-5, rtol=0)
self.assertEqual(input_qparams[1].item(), ref_zero_point)
# During input-weight equalization, we will scale the weights so that
# the following weight quantized observer will have the correct scaled qparams
# Check the weight scale/zero-point values of the quantized observer
weight_quant_obs = PerChannelMinMaxObserver(dtype=weight_qdtype,
qscheme=weight_qscheme)
# Scale the weights for input-weight equalization
ref_w_scaled = w * np.reciprocal(ref_equalization_scale)
w_scaled = torch.mul(torch.tensor(w),
torch.reciprocal(equalization_scale))
self.assertEqual(ref_w_scaled, w_scaled)
# Call forward on the weight quantization observer
weight_quant_obs(w_scaled)
# Check the min/max weight rows are correct
ref_min_weights_scaled = ref_w_scaled.min(axis=1)
ref_max_weights_scaled = ref_w_scaled.max(axis=1)
self.assertEqual(weight_quant_obs.min_vals, ref_min_weights_scaled)
self.assertEqual(weight_quant_obs.max_vals, ref_max_weights_scaled)
weight_qparams = weight_quant_obs.calculate_qparams()
if weight_qscheme == torch.per_channel_symmetric:
ref_min_weights_scaled = np.minimum(
np.zeros(ref_min_weights_scaled.shape), ref_min_weights_scaled)
ref_max_weights_scaled = np.maximum(
np.zeros(ref_max_weights_scaled.shape), ref_max_weights_scaled)
ref_scales = 2 * np.maximum(np.abs(ref_min_weights_scaled),
ref_max_weights_scaled) / 255
ref_zero_points = np.zeros_like(
ref_scales) if weight_qdtype is torch.qint8 else np.ones_like(
ref_scales) * 128
elif weight_qscheme == torch.per_channel_affine_float_qparams:
ref_scales = (ref_max_weights_scaled -
ref_min_weights_scaled) / 255
ref_scales = np.where(ref_scales > 1e-7, ref_scales,
np.ones_like(ref_scales))
ref_zero_points = -1 * ref_min_weights_scaled / ref_scales
else:
ref_min_weights_scaled = np.minimum(
np.zeros_like(ref_min_weights_scaled), ref_min_weights_scaled)
ref_max_weights_scaled = np.maximum(
np.zeros_like(ref_max_weights_scaled), ref_max_weights_scaled)
ref_scales = (ref_max_weights_scaled -
ref_min_weights_scaled) / 255
ref_zero_points = -128 if weight_qdtype is torch.qint8 else 0
ref_zero_points = ref_zero_points - np.round(
ref_min_weights_scaled / ref_scales)
self.assertTrue(
torch.allclose(weight_qparams[0],
torch.tensor(ref_scales,
dtype=weight_qparams[0].dtype),
atol=0.0001))
self.assertTrue(
torch.allclose(weight_qparams[1],
torch.tensor(ref_zero_points,
dtype=weight_qparams[1].dtype),
atol=1))
class _InputEqualizationObserver(nn.Module):
r"""Observer for tracking the running min/max values of input columns, and
computing the quantization parameters for the overall min/max input values.
Args:
dtype: Quantized data type
qscheme: Quantization scheme
quant_min: Minimum quantization value. If unspecified, it will
follow the 8-bit setup.
quant_max: Maximum quantization value. If unspecified, it will
follow the 8-bit setup.
output_obs: For the user to specify what kind of output observer they
would like to use
The running minimum/maximum :math:`x_\text{min/max}` are computed in the
same way as :class:`~torch.quantization.observer.PerChannelMinMaxObserver`,
with the difference that the running min/max values are stored per column.
The qparams are calculated by multiplying the min/max input column values
with the equalization scale, reducing to find the global min/max input
values, and then calculating in the same way as in
:class:`~torch.quantization.observer.MinMaxObserver`
.. note:: If the running minimum equals to the running maximum, the scales
and zero_points are set to 1.0 and 0.
"""
def __init__(self,
dtype=torch.quint8,
qscheme=torch.per_tensor_affine,
quant_min=None,
quant_max=None,
output_obs=None,
factory_kwargs=None) -> None:
super(_InputEqualizationObserver, self).__init__()
if qscheme not in {
torch.per_tensor_affine, torch.per_tensor_symmetric
}:
raise TypeError("Input qscheme must be per-tensor")
self.input_obs = PerChannelMinMaxObserver(
ch_axis=1,
dtype=dtype,
qscheme=qscheme,
quant_min=quant_min,
quant_max=quant_max,
factory_kwargs=factory_kwargs)
if output_obs is None:
self.output_obs = MinMaxObserver(dtype=dtype,
qscheme=qscheme,
quant_min=quant_min,
quant_max=quant_max,
factory_kwargs=factory_kwargs)
else:
self.output_obs = output_obs
self.equalization_scale = torch.empty(0)
def forward(self, x_orig):
# TODO: Allow for convoluational layers
if not (x_orig.ndim == 2):
raise ValueError(
"InputEqualizationObserver only supports Linear layers")
return self.input_obs(x_orig)
def get_input_minmax(self):
return (self.input_obs.min_vals, self.input_obs.max_vals)
def set_equalization_scale(self, equalization_scale):
self.equalization_scale = equalization_scale
def calculate_qparams(self):
r"""
Returns the scale/zero_point for the input and weight rows
"""
if self.equalization_scale.nelement() == 0:
warnings.warn(
"Must call calculate_scale before calling calculate_qparams.\
Returning default scale and zero point. ")
return torch.tensor([1.0]), torch.tensor([0]), torch.tensor(
[1.0]), torch.tensor([0])
# Calculate qparams for the scaled min/max inputs
# Scale the input by the equalization scale located at the same column
# index
(min_inputs, max_inputs) = self.get_input_minmax()
min_input_scaled = torch.min(
torch.mul(min_inputs, self.equalization_scale))
max_input_scaled = torch.max(
torch.mul(max_inputs, self.equalization_scale))
(scale_input, zero_point_input) = self.input_obs._calculate_qparams(
min_input_scaled, max_input_scaled)
return scale_input, zero_point_input
class _WeightEqualizationObserver(nn.Module):
r"""Observer for tracking the running min/max values of weight columns and
rows, and computing the quantization parameters for the weight rows.
Args:
dtype: Quantized data type
qscheme: Quantization scheme
quant_min: Minimum quantization value. If unspecified, it will
follow the 8-bit setup.
quant_max: Maximum quantization value. If unspecified, it will
follow the 8-bit setup.
This observer is made up of 2 PerChannelMinMaxObservers
- weight_col_obs: Used to record the running minimum and maximum of
columns of incoming weight tensors
- weight_row_obs: Used to record the running minimum and maximum of
rows of incoming weight tensors
The running minimum/maximum :math:`w_\text{min/max}` are computed in the
same way as :class:`~torch.quantization.observer.PerChannelMinMaxObserver`.
The qparams are calculated by multiplying the min/max weight row values
with the inverse of the equalization scale, and then calculating in the same
way as in :class:`~torch.quantization.observer.PerChannelMinMaxObserver`
.. note:: If the running minimum equals to the running maximum, the scales
and zero_points are set to 1.0 and 0.
"""
def __init__(self,
dtype=torch.qint8,
qscheme=torch.per_tensor_affine,
quant_min=None,
quant_max=None,
factory_kwargs=None) -> None:
super(_WeightEqualizationObserver, self).__init__()
self.weight_col_obs = PerChannelMinMaxObserver(
ch_axis=1,
dtype=dtype,
qscheme=qscheme,
quant_min=quant_min,
quant_max=quant_max,
factory_kwargs=factory_kwargs)
self.weight_row_obs = PerChannelMinMaxObserver(
ch_axis=0,
dtype=dtype,
qscheme=qscheme,
quant_min=quant_min,
quant_max=quant_max,
factory_kwargs=factory_kwargs)
self.equalization_scale = torch.empty(0)
def forward(self, w_orig):
# TODO: Allow for convoluational layers
if not (w_orig.ndim == 2):
raise ValueError(
"WeightEqualizationObserver only supports Linear layers")
return self._forward(w_orig)
def _forward(self, w_orig):
r"""
Calculates the min/max values of each weight column and weight row.
"""
w_orig = self.weight_col_obs(w_orig)
w_orig = self.weight_row_obs(w_orig)
# Calculate the column indices of the min/max weight in each row
num_row, _ = w_orig.shape
min_weights_ind = []
max_weights_ind = []
for i in range(num_row):
min_weights_ind.append(
torch.nonzero(
w_orig[i] == self.weight_row_obs.min_vals[i])[0][0])
max_weights_ind.append(
torch.nonzero(
w_orig[i] == self.weight_row_obs.max_vals[i])[0][0])
self.min_weights_ind = torch.tensor(min_weights_ind)
self.max_weights_ind = torch.tensor(max_weights_ind)
return w_orig
def get_weight_col_minmax(self):
return (self.weight_col_obs.min_vals, self.weight_col_obs.max_vals)
def get_weight_row_minmax(self):
return (self.weight_row_obs.min_vals, self.weight_row_obs.max_vals)
def set_equalization_scale(self, equalization_scale):
self.equalization_scale = equalization_scale
def calculate_qparams(self):
r"""
Returns the scale/zero_point for the input and weight rows
"""
if self.equalization_scale.nelement() == 0:
warnings.warn(
"Must call calculate_scale before calling calculate_qparams.\
Returning default scale and zero point. ")
return torch.tensor([1.0]), torch.tensor([0]), torch.tensor(
[1.0]), torch.tensor([0])
if self.min_weights_ind is None or self.max_weights_ind is None:
warnings.warn(
"Must find the column indicies of the minimum of each row in the \
weights in order to calculate the qparams calculate the \
qparams. Returning default scale and zero point. ")
return torch.tensor([1.0]), torch.tensor([0]), torch.tensor(
[1.0]), torch.tensor([0])
# Calculate the qparams for weights by using the rows
# Scale the weight rows by the reciprocal of the equalization scale
# located at the same column index
(min_weights, max_weights) = self.get_weight_row_minmax()
min_weights_scaled = torch.mul(
min_weights,
torch.reciprocal(self.equalization_scale[self.min_weights_ind]))
max_weights_scaled = torch.mul(
max_weights,
torch.reciprocal(self.equalization_scale[self.max_weights_ind]))
(scale_weight,
zero_point_weight) = self.weight_row_obs._calculate_qparams(
min_weights_scaled, max_weights_scaled)
return scale_weight, zero_point_weight