def test_minmax_observer(self, qdtype, qscheme, reduce_range):
# reduce_range cannot be true for symmetric quantization with uint8
if qdtype == torch.quint8 and qscheme == torch.per_tensor_symmetric:
reduce_range = False
myobs = MinMaxObserver(dtype=qdtype, qscheme=qscheme, reduce_range=reduce_range)
x = torch.tensor([1.0, 2.0, 2.0, 3.0, 4.0, 5.0, 6.0])
y = torch.tensor([4.0, 5.0, 5.0, 6.0, 7.0, 8.0])
result = myobs(x)
result = myobs(y)
self.assertEqual(result, y)
self.assertEqual(myobs.min_val, 1.0)
self.assertEqual(myobs.max_val, 8.0)
qparams = myobs.calculate_qparams()
if reduce_range:
if qscheme == torch.per_tensor_symmetric:
ref_scale = 0.062745 * 255 / 127
ref_zero_point = 0 if qdtype is torch.qint8 else 128
else:
ref_scale = 0.0313725 * 255 / 127
ref_zero_point = -64 if qdtype is torch.qint8 else 0
else:
if qscheme == torch.per_tensor_symmetric:
ref_scale = 0.062745
ref_zero_point = 0 if qdtype is torch.qint8 else 128
else:
ref_scale = 0.0313725
ref_zero_point = -128 if qdtype is torch.qint8 else 0
self.assertEqual(qparams[1].item(), ref_zero_point)
self.assertAlmostEqual(qparams[0].item(), ref_scale, delta=1e-5)
def test_observer_scriptable(self):
obs_list = [
MinMaxObserver(),
MovingAverageMinMaxObserver(),
MinMaxDynamicQuantObserver()
]
for obs in obs_list:
scripted = torch.jit.script(obs)
x = torch.rand(3, 4)
obs(x)
scripted(x)
self.assertEqual(obs.calculate_qparams(),
scripted.calculate_qparams())
buf = io.BytesIO()
torch.jit.save(scripted, buf)
buf.seek(0)
loaded = torch.jit.load(buf)
self.assertEqual(obs.calculate_qparams(),
loaded.calculate_qparams())
# Check TensorListObserver
from torch.quantization.observer import _MinMaxTensorListObserver
obs = _MinMaxTensorListObserver()
scripted = torch.jit.script(obs)
x = [torch.rand(3, 4), torch.rand(4, 5)]
obs(x)
scripted(x)
self.assertEqual(obs.calculate_qparams(), scripted.calculate_qparams())
def test_minmax_observer(self, qdtype, qscheme):
myobs = MinMaxObserver(dtype=qdtype, qscheme=qscheme)
x = torch.tensor([1.0, 2.0, 2.0, 3.0, 4.0, 5.0, 6.0])
y = torch.tensor([4.0, 5.0, 5.0, 6.0, 7.0, 8.0])
result = myobs(x)
result = myobs(y)
self.assertEqual(result, y)
self.assertEqual(myobs.min_val, 1.0)
self.assertEqual(myobs.max_val, 8.0)
qparams = myobs.calculate_qparams()
if qscheme == torch.per_tensor_symmetric:
ref_scale = 0.062745
ref_zero_point = 0 if qdtype is torch.qint8 else 128
else:
ref_scale = 0.0313725
ref_zero_point = -128 if qdtype is torch.qint8 else 0
self.assertEqual(qparams[1].item(), ref_zero_point)
self.assertAlmostEqual(qparams[0].item(), ref_scale, delta=1e-5)
def test_per_tensor_observers(self, qdtype, qscheme, reduce_range):
# reduce_range cannot be true for symmetric quantization with uint8
if qdtype == torch.quint8 and qscheme == torch.per_tensor_symmetric:
reduce_range = False
ObserverList = [MinMaxObserver(dtype=qdtype, qscheme=qscheme, reduce_range=reduce_range),
MovingAverageMinMaxObserver(averaging_constant=0.5,
dtype=qdtype,
qscheme=qscheme,
reduce_range=reduce_range)]
for myobs in ObserverList:
# Calculate Qparams should return with a warning for observers with no data
qparams = myobs.calculate_qparams()
if type(myobs) == MinMaxObserver:
x = torch.tensor([1.0, 2.0, 2.0, 3.0, 4.0, 5.0, 6.0])
y = torch.tensor([4.0, 5.0, 5.0, 6.0, 7.0, 8.0])
else:
# Moving average of min/max for x and y matches that of
# extreme values for x/y used for minmax observer
x = torch.tensor([0.0, 2.0, 2.0, 3.0, 4.0, 5.0, 6.0])
y = torch.tensor([2.0, 5.0, 5.0, 6.0, 7.0, 10.0])
result = myobs(x)
result = myobs(y)
self.assertEqual(result, y)
self.assertEqual(myobs.min_val, 1.0)
self.assertEqual(myobs.max_val, 8.0)
qparams = myobs.calculate_qparams()
if reduce_range:
if qscheme == torch.per_tensor_symmetric:
ref_scale = 0.062745 * 255 / 127
ref_zero_point = 0 if qdtype is torch.qint8 else 128
else:
ref_scale = 0.0313725 * 255 / 127
ref_zero_point = -64 if qdtype is torch.qint8 else 0
else:
if qscheme == torch.per_tensor_symmetric:
ref_scale = 0.062745
ref_zero_point = 0 if qdtype is torch.qint8 else 128
else:
ref_scale = 0.0313725
ref_zero_point = -128 if qdtype is torch.qint8 else 0
self.assertEqual(qparams[1].item(), ref_zero_point)
self.assertAlmostEqual(qparams[0].item(), ref_scale, delta=1e-5)
state_dict = myobs.state_dict()
b = io.BytesIO()
torch.save(state_dict, b)
b.seek(0)
loaded_dict = torch.load(b)
for key in state_dict:
self.assertEqual(state_dict[key], loaded_dict[key])
loaded_obs = MinMaxObserver(dtype=qdtype, qscheme=qscheme, reduce_range=reduce_range)
loaded_obs.load_state_dict(loaded_dict)
loaded_qparams = loaded_obs.calculate_qparams()
self.assertEqual(myobs.min_val, loaded_obs.min_val)
self.assertEqual(myobs.max_val, loaded_obs.max_val)
self.assertEqual(myobs.calculate_qparams(), loaded_obs.calculate_qparams())
def test_tensor_list_observer(self):
from torch.quantization.observer import _MinMaxTensorListObserver
x = [torch.tensor([1.0, 2.5, 3.5]),
torch.tensor([2.0, 4.5, 3.5]),
torch.tensor([4.0, 2.5, 3.5]), ]
obs = _MinMaxTensorListObserver()
obs(x)
qparams = obs.calculate_qparams()
ref_min_val = []
ref_max_val = []
ref_qparams = []
for i in x:
obs_ref = MinMaxObserver()
obs_ref(i)
ref_min_val.append(obs_ref.min_val)
ref_max_val.append(obs_ref.max_val)
ref_qparams.append(obs_ref.calculate_qparams())
for i in range(len(x)):
self.assertEqual(obs.min_val[i], ref_min_val[i])
self.assertEqual(obs.max_val[i], ref_max_val[i])
self.assertEqual(qparams[0][i], ref_qparams[i][0])
self.assertEqual(qparams[1][i], ref_qparams[i][1])
def test_observer_scriptable(self):
obs_list = [MinMaxObserver(), MovingAverageMinMaxObserver(), MinMaxDynamicQuantObserver()]
for obs in obs_list:
scripted = torch.jit.script(obs)
x = torch.rand(3, 4)
obs(x)
scripted(x)
self.assertEqual(obs.calculate_qparams(), scripted.calculate_qparams())
buf = io.BytesIO()
torch.jit.save(scripted, buf)
buf.seek(0)
loaded = torch.jit.load(buf)
self.assertEqual(obs.calculate_qparams(), loaded.calculate_qparams())
def test_observer_qparams_respects_device_affinity(self):
"""
Ensure that the scale and zero_point returned by the observer
are on the same device as the input tensor.
"""
observerList = [MinMaxObserver(),
MovingAverageMinMaxObserver(),
PerChannelMinMaxObserver(),
MovingAveragePerChannelMinMaxObserver()]
for obs in observerList:
device = torch.device('cuda:1')
x = torch.randn(1, 2, device=device)
obs.to(device)
result = obs(x)
scale, zero_point = obs.calculate_qparams()
self.assertEqual(x.device, scale.device)
self.assertEqual(x.device, zero_point.device)
def test_minmax_observer(self, qdtype, qscheme, reduce_range):
# reduce_range cannot be true for symmetric quantization with uint8
if qdtype == torch.quint8 and qscheme == torch.per_tensor_symmetric:
reduce_range = False
myobs = MinMaxObserver(dtype=qdtype, qscheme=qscheme, reduce_range=reduce_range)
x = torch.tensor([1.0, 2.0, 2.0, 3.0, 4.0, 5.0, 6.0])
y = torch.tensor([4.0, 5.0, 5.0, 6.0, 7.0, 8.0])
result = myobs(x)
result = myobs(y)
self.assertEqual(result, y)
self.assertEqual(myobs.min_val, 1.0)
self.assertEqual(myobs.max_val, 8.0)
qparams = myobs.calculate_qparams()
if reduce_range:
if qscheme == torch.per_tensor_symmetric:
ref_scale = 0.062745 * 255 / 127
ref_zero_point = 0 if qdtype is torch.qint8 else 128
else:
ref_scale = 0.0313725 * 255 / 127
ref_zero_point = -64 if qdtype is torch.qint8 else 0
else:
if qscheme == torch.per_tensor_symmetric:
ref_scale = 0.062745
ref_zero_point = 0 if qdtype is torch.qint8 else 128
else:
ref_scale = 0.0313725
ref_zero_point = -128 if qdtype is torch.qint8 else 0
self.assertEqual(qparams[1].item(), ref_zero_point)
self.assertAlmostEqual(qparams[0].item(), ref_scale, delta=1e-5)
# Test for serializability
state_dict = myobs.state_dict()
b = io.BytesIO()
torch.save(state_dict, b)
b.seek(0)
loaded_dict = torch.load(b)
for key in state_dict:
self.assertEqual(state_dict[key], loaded_dict[key])
loaded_obs = MinMaxObserver(dtype=qdtype, qscheme=qscheme, reduce_range=reduce_range)
loaded_obs.load_state_dict(loaded_dict)
loaded_qparams = loaded_obs.calculate_qparams()
self.assertEqual(myobs.min_val, loaded_obs.min_val)
self.assertEqual(myobs.max_val, loaded_obs.max_val)
self.assertEqual(myobs.calculate_qparams(), loaded_obs.calculate_qparams())
def test_per_tensor_observer(self):
obs = MinMaxObserver()
self._test_obs(obs, input_size=[5, 5], generate=False)
get_default_qconfig,
)
# TensorFlow Lite Quantization Specs
# https://www.tensorflow.org/lite/performance/quantization_spec?hl=en
# For activations: int8 asymmetric per-tensor [-128, 127] range
# For weights: int8 symmetric per-tensor [-127, 127] range
_TFLITE_QCONFIG = QConfig(
activation=MovingAverageMinMaxObserver.with_args(
dtype=torch.qint8,
quant_min=-128,
quant_max=127,
qscheme=torch.per_tensor_affine,
),
weight=MinMaxObserver.with_args(dtype=torch.qint8,
quant_min=-127,
quant_max=127,
qscheme=torch.per_tensor_symmetric),
)
_TFLITE_QAT_QCONFIG = QConfig(
activation=FakeQuantize.with_args(
observer=MovingAverageMinMaxObserver,
dtype=torch.qint8,
quant_min=-128,
quant_max=127,
qscheme=torch.per_tensor_affine,
),
weight=FakeQuantize.with_args(observer=MinMaxObserver,
dtype=torch.qint8,
quant_min=-127,
quant_max=127,
qscheme=torch.per_tensor_symmetric),
