def test_forward_per_channel_half_precision_numerics(self):
scale = torch.randn(5).abs()
zero = torch.randn(5).to(dtype=torch.int)
axis = 1
mini = 0
maxi = 255
for i in range(20):
X1 = torch.randn(4, 5).to(torch.float16)
Y1 = torch.fake_quantize_per_channel_affine(X1, scale, zero, axis, mini, maxi)
Y1r = _fake_quantize_per_channel_affine_reference(X1, scale, zero, axis, mini, maxi)
self.assertTrue(torch.allclose(Y1, Y1r, rtol=tolerance, atol=tolerance))
# to force overflow
X2 = torch.randn(4, 5).to(torch.float16)
X2[0, 0] = 2**15 + .01
Y2 = torch.fake_quantize_per_channel_affine(X2, scale, zero, axis, mini, maxi)
Y2r = _fake_quantize_per_channel_affine_reference(X2, scale, zero, axis, mini, maxi)
self.assertTrue(torch.allclose(Y2, Y2r, rtol=tolerance, atol=tolerance))
scale = torch.zeros(5) + 10
# to force underflow
X3 = torch.randn(4, 5).to(torch.float16)
X3[0, 0] = 2**-24
Y3 = torch.fake_quantize_per_channel_affine(X3, scale, zero, axis, mini, maxi)
Y3r = _fake_quantize_per_channel_affine_reference(X3, scale, zero, axis, mini, maxi)
self.assertTrue(torch.allclose(Y3, Y3r, rtol=tolerance, atol=tolerance))
def fake_quantize_per_channel(input, scale_inv, zero_point, axis, quant_min, quant_max, method, inplace):
if method == -1:
if (int(torch.__version__.split('.')[1]) > 9) and (int(torch.__version__.split('.')[0]) > 0):
zero_point = zero_point.to(torch.int32)
else:
zero_point = zero_point.to(torch.long)
return torch.fake_quantize_per_channel_affine(input, 1.0 / scale_inv, zero_point, axis, quant_min, quant_max)
else:
device_id = 1 if input.device == torch.device("cpu") else 0
input_split = torch.split(input, 1, dim=axis)
input_cat = []
if support_onnx_export():
for i in range(len(input_split)):
input_cat.append(torch.ops.vai.fix_neuron(input_split[i], quant_min, quant_max,
scale_inv[i], zero_point[i], method,
device_id, inplace))
output = torch.cat(input_cat, axis)
return output
else:
for i in range(len(input_split)):
nndct_kernels.FixNeuronV2(input_split[i], input_split[i], quant_min,
quant_max, scale_inv, zero_point,
method, device_id)
input_cat.append(input_split[i])
output = torch.cat(input_cat, axis)
return output
def _fb_fake_quant(self, inputs, amax):
"""Native pytorch fake quantization."""
logging.log_first_n(
logging.WARNING,
"Use Pytorch's native experimental fake quantization.", 1)
bound = (1 << (self._num_bits - 1 + int(self._unsigned))) - 1
# To be consistent with ONNX, full range is used. e.g. range is [-128, 127] in int8
if amax.numel() == 1:
outputs = torch.fake_quantize_per_tensor_affine(
inputs,
amax.item() / bound, 0,
-bound - 1 if not self._unsigned else 0, bound)
else:
amax_sequeeze = amax.squeeze().detach()
if len(amax_sequeeze.shape) != 1:
raise TypeError(
"Pytorch's native quantization doesn't support multiple axes"
)
quant_dim = list(amax.shape).index(list(amax_sequeeze.shape)[0])
scale = amax_sequeeze / bound
outputs = torch.fake_quantize_per_channel_affine(
inputs, scale.data,
torch.zeros_like(scale, dtype=torch.long).data, quant_dim,
-bound - 1 if not self._unsigned else 0, bound)
return outputs
def _test_backward_per_channel_cachemask_impl(self, device):
torch_types = (torch.qint8, torch.quint8)
float_types = (torch.float32, torch.float16, torch.float64)
for torch_type, float_type in itertools.product(
torch_types, float_types):
X = torch.randn(1, 2, 4, 4, dtype=float_type).to(device)
# pick the scale + zp so that some values get clipped
axis = 1
obs = torch.quantization.PerChannelMinMaxObserver(
axis, torch_type).to(device)
obs(X * 0.75)
scale, zero_point = obs.calculate_qparams()
# TODO(future PR): fix the wrong dtype in obs.calculate_qparams and remove the cast
zero_point = zero_point.to(torch.int64)
quant_min, quant_max = obs._calculate_qmin_qmax()
X.requires_grad_()
Y_prime = torch.fake_quantize_per_channel_affine(
X, scale, zero_point, axis, quant_min, quant_max)
dout = torch.rand_like(X, dtype=float_type).to(device)
dX = _fake_quantize_per_channel_affine_grad_reference(
dout, X, scale, zero_point, axis, quant_min, quant_max)
Y_prime.backward(dout)
np.testing.assert_allclose(dX.cpu().detach().numpy(),
X.grad.cpu().detach().numpy(),
rtol=tolerance,
atol=tolerance)
assert (X.grad.dtype == float_type)
def forward(self, X):
self.activation_post_process(X.detach())
_scale, _zero_point = self.calculate_qparams()
_scale = _scale.to(self.scale.device)
_zero_point = _zero_point.to(self.zero_point.device)
if self.static_enabled[0] == 1:
self.scale.data.copy_(_scale)
self.zero_point.data.copy_(_zero_point)
if self.fake_quant_enabled[0] == 1:
if self.learning_enabled[0] == 1:
self.zero_point.clamp(self.quant_min, self.quant_max)
if self.qscheme in (torch.per_channel_symmetric,
torch.per_channel_affine):
X = _LearnableFakeQuantizePerChannelOp.apply(
X, self.scale, self.zero_point, self.ch_axis,
self.quant_min, self.quant_max, self.grad_factor)
else:
X = _LearnableFakeQuantizePerTensorOp.apply(
X, self.scale, self.zero_point, self.quant_min,
self.quant_max, self.grad_factor)
else:
if self.qscheme == torch.per_channel_symmetric or \
self.qscheme == torch.per_channel_affine:
X = torch.fake_quantize_per_channel_affine(
X, self.scale, self.zero_point, self.ch_axis,
self.quant_min, self.quant_max)
else:
X = torch.fake_quantize_per_tensor_affine(
X, float(self.scale.item()),
int(self.zero_point.item()), self.quant_min,
self.quant_max)
return X
def forward(ctx, X, scale, zero_point, ch_axis, q_min, q_max, grad_factor):
ctx.save_for_backward(X, scale, zero_point)
scale_vec = scale.detach().type(torch.float32)
zp_vec = ((zero_point.detach() + 0.5).clamp(q_min, q_max)).type(torch.int64)
X_fq = torch.fake_quantize_per_channel_affine(
X, scale_vec, zp_vec, ch_axis, q_min, q_max)
ctx.other = q_min, q_max, X_fq, ch_axis, grad_factor
return X_fq
def forward(self, X):
if self.static_enabled[0] == 1:
self.activation_post_process(X.detach())
_scale, _zero_point = self.calculate_qparams()
_scale = _scale.to(self.scale.device)
_zero_point = _zero_point.to(self.zero_point.device)
if self.init and self.wt:
assert self.static_enabled[0] == 1
self.scale = Parameter(torch.FloatTensor([1.] * len(_scale)))
self.zero_point = Parameter(
torch.FloatTensor([0.] * len(_zero_point)))
self.to(_zero_point.device)
self.scale.requires_grad = False
self.zero_point.requires_grad = False
self.init = False
if self.static_enabled[0] == 1:
self.scale.data.copy_(_scale)
self.zero_point.data.copy_(_zero_point)
if self.fake_quant_enabled[0] == 1:
if self.learning_enabled[0] == 1:
if self.use_grad_scaling:
if self.wt:
grad_factor = 1.0 / (X.numel() * self.quant_max)**0.5
else:
grad_factor = 1.0 / (X[0].numel() *
self.quant_max)**0.5
else:
grad_factor = 1.0
if self.qscheme in (torch.per_channel_symmetric,
torch.per_channel_affine):
X = _LearnableFakeQuantizePerChannelOp.apply(
X, self.scale, self.zero_point, self.ch_axis,
self.quant_min, self.quant_max, grad_factor)
else:
X = _LearnableFakeQuantizePerTensorOp.apply(
X, self.scale, self.zero_point, self.quant_min,
self.quant_max, grad_factor)
else:
if self.qscheme == torch.per_channel_symmetric or \
self.qscheme == torch.per_channel_affine:
zero_point = torch.LongTensor([
i.round() for i in self.zero_point
]).to(self.zero_point.device)
X = torch.fake_quantize_per_channel_affine(
X, self.scale, zero_point, self.ch_axis,
self.quant_min, self.quant_max)
else:
X = torch.fake_quantize_per_tensor_affine(
X, float(self.scale.item()),
int(self.zero_point.item()), self.quant_min,
self.quant_max)
return X
def forward(self, X):
if self.observer_enabled:
self.activation_post_process(X.detach())
self.scale, self.zero_point = self.calculate_qparams()
if self.fake_quant_enabled:
if self.qscheme == torch.per_channel_symmetric or self.qscheme == torch.per_channel_affine:
X = torch.fake_quantize_per_channel_affine(X, self.scale, self.zero_point,
self.ch_axis, self.quant_min, self.quant_max)
else:
X = torch.fake_quantize_per_tensor_affine(X, float(self.scale),
int(self.zero_point), self.quant_min,
self.quant_max)
return X
def _test_numerical_consistency(self, test_type):
r"""Comparing numerical consistency between quantize/dequantize op and the fake quantize op across devices and dtypes
"""
torch.random.manual_seed(NP_RANDOM_SEED)
torch_types = [torch.qint8, torch.quint8]
float_types = [torch.float, torch.float16, torch.float64]
zero_types = [torch.long]
devices = [torch.device('cpu'),
torch.device('cuda')
] if torch.cuda.is_available() else [torch.device('cpu')]
axis = 1
for i in range(20):
for torch_type, float_type, device, zero_type in itertools.product(
torch_types, float_types, devices, zero_types):
X = torch.randn(3, 3, device=device).to(float_type)
scales = (10 * torch.randn(3, device=device)).abs()
scale = scales.mean().to(float).item()
zeros = (10 * torch.randn(3, device=device)).abs().to(
dtype=zero_type)
zero = zeros.max().view(1).item()
quant_min = torch.iinfo(torch_type).min
quant_max = torch.iinfo(torch_type).max
test_was_run = False
if test_type == "per_tensor":
test_was_run = True
Y = torch.dequantize(
torch.quantize_per_tensor(
X.to('cpu').to(torch.float), scale, zero,
torch_type)).to(device).to(float_type)
Y_prime = torch.fake_quantize_per_tensor_affine(
X, scale, zero, quant_min, quant_max)
self.assertEqual(
Y, Y_prime,
"Difference found between dequant+quant_per_tensor and fake_quantize_per_tensor"
)
if test_type == "per_channel":
test_was_run = True
Y = torch.dequantize(
torch.quantize_per_channel(
X.to('cpu').to(torch.float), scales.to('cpu'),
zeros.to('cpu'), axis,
torch_type)).to(device).to(float_type)
Y_prime = torch.fake_quantize_per_channel_affine(
X, scales, zeros, axis, quant_min, quant_max)
self.assertEqual(
Y, Y_prime,
"Difference found between dequant+quant_per_channel and fake_quantize_per_channel"
)
self.assertTrue(test_was_run)
def fake_quant(self, X):
if self.qscheme == torch.per_channel_symmetric:
zero_point = torch.LongTensor([i.round() for i in self.zero_point
]).to(self.zero_point.device)
X = torch.fake_quantize_per_channel_affine(X, self.scale,
zero_point,
self.ch_axis,
self.quant_min,
self.quant_max)
else:
X = torch.fake_quantize_per_tensor_affine(
X, float(self.scale.item()), int(self.zero_point.item()),
self.quant_min, self.quant_max)
return X
def test_forward_per_channel(self, device, X):
r"""Tests the forward path of the FakeQuantizePerTensorAffine op.
"""
np.random.seed(NP_RANDOM_SEED)
X, (scale, zero_point, axis, torch_type) = X
quant_min = torch.iinfo(torch_type).min
quant_max = torch.iinfo(torch_type).max
X = to_tensor(X, device)
scale = to_tensor(scale, device)
zero_point = torch.tensor(zero_point).to(dtype=torch.int64, device=device)
Y = _fake_quantize_per_channel_affine_reference(X.cpu(), scale.cpu(), zero_point.cpu(), axis, quant_min, quant_max)
Y_prime = torch.fake_quantize_per_channel_affine(
X, scale, zero_point, axis, quant_min, quant_max)
np.testing.assert_allclose(Y, Y_prime.cpu(), rtol=tolerance, atol=tolerance)
def test_numerical_consistency_per_channel(self, device, X):
r"""Comparing numerical consistency between CPU quantize/dequantize op and the CPU fake quantize op
"""
np.random.seed(NP_RANDOM_SEED)
X, (scale, zero_point, axis, torch_type) = X
quant_min = torch.iinfo(torch_type).min
quant_max = torch.iinfo(torch_type).max
X = to_tensor(X, device)
scale = to_tensor(scale, device)
zero_point = torch.tensor(zero_point).to(dtype=torch.int64, device=device)
# quantize_linear and dequantize are only implemented in CPU
Y = torch.dequantize(torch.quantize_per_channel(X.cpu(), scale.cpu(), zero_point.cpu(), axis, torch_type))
Y_prime = torch.fake_quantize_per_channel_affine(
X, scale, zero_point, axis, quant_min, quant_max)
np.testing.assert_allclose(Y, Y_prime.cpu(), rtol=tolerance, atol=tolerance)
def test_backward_per_channel(self, device, X):
r"""Tests the backward method.
"""
np.random.seed(NP_RANDOM_SEED)
X, (scale, zero_point, axis, torch_type) = X
quant_min = torch.iinfo(torch_type).min
quant_max = torch.iinfo(torch_type).max
X = to_tensor(X, device)
scale = to_tensor(scale, device)
zero_point = torch.tensor(zero_point).to(dtype=torch.int64, device=device)
X.requires_grad_()
Y_prime = torch.fake_quantize_per_channel_affine(
X, scale, zero_point, axis, quant_min, quant_max)
dout = torch.rand(X.shape, dtype=torch.float).to(device)
dX = _fake_quantize_per_channel_affine_grad_reference(
dout, X, scale, zero_point, axis, quant_min, quant_max)
Y_prime.backward(dout)
np.testing.assert_allclose(dX.cpu().detach().numpy(), X.grad.cpu().detach().numpy(), rtol=tolerance, atol=tolerance)
def forward(self, X):
if self.observer_enabled[0] == 1:
self.activation_post_process(X.detach())
_scale, _zero_point = self.calculate_qparams()
_scale, _zero_point = _scale.to(self.scale.device), _zero_point.to(self.zero_point.device)
self.scale.resize_(_scale.shape)
self.scale.copy_(_scale)
self.zero_point.resize_(_zero_point.shape)
self.zero_point.copy_(_zero_point)
if self.fake_quant_enabled[0] == 1:
if self.qscheme == torch.per_channel_symmetric or self.qscheme == torch.per_channel_affine:
X = torch.fake_quantize_per_channel_affine(X, self.scale, self.zero_point,
self.ch_axis, self.quant_min, self.quant_max)
else:
X = torch.fake_quantize_per_tensor_affine(X, float(self.scale),
int(self.zero_point), self.quant_min,
self.quant_max)
return X
def _test_forward_per_channel_cachemask_impl(self, device):
torch_types = (torch.qint8, torch.quint8)
float_types = (torch.float32, torch.float16, torch.float64)
for torch_type, float_type in itertools.product(torch_types, float_types):
X = torch.randn(1, 2, 4, 4, dtype=float_type).to(device)
# pick the scale + zp so that some values get clipped
axis = 1
obs = torch.quantization.PerChannelMinMaxObserver(axis, torch_type).to(device)
obs(X * 0.75)
scale, zero_point = obs.calculate_qparams()
# TODO(future PR): fix the wrong dtype in obs.calculate_qparams and remove the cast
zero_point = zero_point.to(torch.int32)
quant_min, quant_max = obs._calculate_qmin_qmax()
Y = _fake_quantize_per_channel_affine_reference(
X.cpu(), scale.cpu(), zero_point.cpu(), axis, quant_min, quant_max)
Y_prime = torch.fake_quantize_per_channel_affine(
X, scale, zero_point, axis, quant_min, quant_max)
np.testing.assert_allclose(Y, Y_prime.cpu(), rtol=tolerance, atol=tolerance)
self.assertTrue(Y.dtype == float_type)
def fake_quantize(self, input, inplace):
if self._round_method == "half_even":
if NndctOption.nndct_tensorrt_quant_algo.value and self._symmetric_mode == "symmetric":
return fake_quantize_per_channel_tensorrt(input, self._float_max, self._quant_min, self._quant_max, self._axis)
else:
if (int(torch.__version__.split('.')[1]) > 9) and (int(torch.__version__.split('.')[0]) > 0):
self._zero_point = self._zero_point.to(torch.int32)
else:
self._zero_point = self._zero_point.to(torch.long)
return torch.fake_quantize_per_channel_affine(input, self._scale, self._zero_point,
self._axis, self._quant_min, self._quant_max)
else:
if self._round_method == "half_up":
method = 2
elif self._round_method == "half_down":
method = 6
elif self._round_method == "std_round":
method = 3
return fake_quantize_per_channel(input, 1.0/self._scale, self._zero_point,
self._axis, self._quant_min, self._quant_max, method, inplace)
def pointwise_ops(self):
a = torch.randn(4)
b = torch.randn(4)
t = torch.tensor([-1, -2, 3], dtype=torch.int8)
r = torch.tensor([0, 1, 10, 0], dtype=torch.int8)
t = torch.tensor([-1, -2, 3], dtype=torch.int8)
s = torch.tensor([4, 0, 1, 0], dtype=torch.int8)
f = torch.zeros(3)
g = torch.tensor([-1, 0, 1])
w = torch.tensor([0.3810, 1.2774, -0.2972, -0.3719, 0.4637])
return (
torch.abs(torch.tensor([-1, -2, 3])),
torch.absolute(torch.tensor([-1, -2, 3])),
torch.acos(a),
torch.arccos(a),
torch.acosh(a.uniform_(1.0, 2.0)),
torch.add(a, 20),
torch.add(a, torch.randn(4, 1), alpha=10),
torch.addcdiv(torch.randn(1, 3),
torch.randn(3, 1),
torch.randn(1, 3),
value=0.1),
torch.addcmul(torch.randn(1, 3),
torch.randn(3, 1),
torch.randn(1, 3),
value=0.1),
torch.angle(a),
torch.asin(a),
torch.arcsin(a),
torch.asinh(a),
torch.arcsinh(a),
torch.atan(a),
torch.arctan(a),
torch.atanh(a.uniform_(-1.0, 1.0)),
torch.arctanh(a.uniform_(-1.0, 1.0)),
torch.atan2(a, a),
torch.bitwise_not(t),
torch.bitwise_and(t, torch.tensor([1, 0, 3], dtype=torch.int8)),
torch.bitwise_or(t, torch.tensor([1, 0, 3], dtype=torch.int8)),
torch.bitwise_xor(t, torch.tensor([1, 0, 3], dtype=torch.int8)),
torch.ceil(a),
torch.clamp(a, min=-0.5, max=0.5),
torch.clamp(a, min=0.5),
torch.clamp(a, max=0.5),
torch.clip(a, min=-0.5, max=0.5),
torch.conj(a),
torch.copysign(a, 1),
torch.copysign(a, b),
torch.cos(a),
torch.cosh(a),
torch.deg2rad(
torch.tensor([[180.0, -180.0], [360.0, -360.0], [90.0,
-90.0]])),
torch.div(a, b),
torch.divide(a, b, rounding_mode="trunc"),
torch.divide(a, b, rounding_mode="floor"),
torch.digamma(torch.tensor([1.0, 0.5])),
torch.erf(torch.tensor([0.0, -1.0, 10.0])),
torch.erfc(torch.tensor([0.0, -1.0, 10.0])),
torch.erfinv(torch.tensor([0.0, 0.5, -1.0])),
torch.exp(torch.tensor([0.0, math.log(2.0)])),
torch.exp2(torch.tensor([0.0, math.log(2.0), 3.0, 4.0])),
torch.expm1(torch.tensor([0.0, math.log(2.0)])),
torch.fake_quantize_per_channel_affine(
torch.randn(2, 2, 2),
(torch.randn(2) + 1) * 0.05,
torch.zeros(2),
1,
0,
255,
),
torch.fake_quantize_per_tensor_affine(a, 0.1, 0, 0, 255),
torch.float_power(torch.randint(10, (4, )), 2),
torch.float_power(torch.arange(1, 5), torch.tensor([2, -3, 4,
-5])),
torch.floor(a),
# torch.floor_divide(torch.tensor([4.0, 3.0]), torch.tensor([2.0, 2.0])),
# torch.floor_divide(torch.tensor([4.0, 3.0]), 1.4),
torch.fmod(torch.tensor([-3, -2, -1, 1, 2, 3]), 2),
torch.fmod(torch.tensor([1, 2, 3, 4, 5]), 1.5),
torch.frac(torch.tensor([1.0, 2.5, -3.2])),
torch.randn(4, dtype=torch.cfloat).imag,
torch.ldexp(torch.tensor([1.0]), torch.tensor([1])),
torch.ldexp(torch.tensor([1.0]), torch.tensor([1, 2, 3, 4])),
torch.lerp(torch.arange(1.0, 5.0),
torch.empty(4).fill_(10), 0.5),
torch.lerp(
torch.arange(1.0, 5.0),
torch.empty(4).fill_(10),
torch.full_like(torch.arange(1.0, 5.0), 0.5),
),
torch.lgamma(torch.arange(0.5, 2, 0.5)),
torch.log(torch.arange(5) + 10),
torch.log10(torch.rand(5)),
torch.log1p(torch.randn(5)),
torch.log2(torch.rand(5)),
torch.logaddexp(torch.tensor([-1.0]), torch.tensor([-1, -2, -3])),
torch.logaddexp(torch.tensor([-100.0, -200.0, -300.0]),
torch.tensor([-1, -2, -3])),
torch.logaddexp(torch.tensor([1.0, 2000.0, 30000.0]),
torch.tensor([-1, -2, -3])),
torch.logaddexp2(torch.tensor([-1.0]), torch.tensor([-1, -2, -3])),
torch.logaddexp2(torch.tensor([-100.0, -200.0, -300.0]),
torch.tensor([-1, -2, -3])),
torch.logaddexp2(torch.tensor([1.0, 2000.0, 30000.0]),
torch.tensor([-1, -2, -3])),
torch.logical_and(r, s),
torch.logical_and(r.double(), s.double()),
torch.logical_and(r.double(), s),
torch.logical_and(r, s, out=torch.empty(4, dtype=torch.bool)),
torch.logical_not(torch.tensor([0, 1, -10], dtype=torch.int8)),
torch.logical_not(
torch.tensor([0.0, 1.5, -10.0], dtype=torch.double)),
torch.logical_not(
torch.tensor([0.0, 1.0, -10.0], dtype=torch.double),
out=torch.empty(3, dtype=torch.int16),
),
torch.logical_or(r, s),
torch.logical_or(r.double(), s.double()),
torch.logical_or(r.double(), s),
torch.logical_or(r, s, out=torch.empty(4, dtype=torch.bool)),
torch.logical_xor(r, s),
torch.logical_xor(r.double(), s.double()),
torch.logical_xor(r.double(), s),
torch.logical_xor(r, s, out=torch.empty(4, dtype=torch.bool)),
torch.logit(torch.rand(5), eps=1e-6),
torch.hypot(torch.tensor([4.0]), torch.tensor([3.0, 4.0, 5.0])),
torch.i0(torch.arange(5, dtype=torch.float32)),
torch.igamma(a, b),
torch.igammac(a, b),
torch.mul(torch.randn(3), 100),
torch.multiply(torch.randn(4, 1), torch.randn(1, 4)),
torch.mvlgamma(torch.empty(2, 3).uniform_(1.0, 2.0), 2),
torch.tensor([float("nan"),
float("inf"), -float("inf"), 3.14]),
torch.nan_to_num(w),
torch.nan_to_num(w, nan=2.0),
torch.nan_to_num(w, nan=2.0, posinf=1.0),
torch.neg(torch.randn(5)),
# torch.nextafter(torch.tensor([1, 2]), torch.tensor([2, 1])) == torch.tensor([eps + 1, 2 - eps]),
torch.polygamma(1, torch.tensor([1.0, 0.5])),
torch.polygamma(2, torch.tensor([1.0, 0.5])),
torch.polygamma(3, torch.tensor([1.0, 0.5])),
torch.polygamma(4, torch.tensor([1.0, 0.5])),
torch.pow(a, 2),
torch.pow(torch.arange(1.0, 5.0), torch.arange(1.0, 5.0)),
torch.rad2deg(
torch.tensor([[3.142, -3.142], [6.283, -6.283],
[1.570, -1.570]])),
torch.randn(4, dtype=torch.cfloat).real,
torch.reciprocal(a),
torch.remainder(torch.tensor([-3.0, -2.0]), 2),
torch.remainder(torch.tensor([1, 2, 3, 4, 5]), 1.5),
torch.round(a),
torch.rsqrt(a),
torch.sigmoid(a),
torch.sign(torch.tensor([0.7, -1.2, 0.0, 2.3])),
torch.sgn(a),
torch.signbit(torch.tensor([0.7, -1.2, 0.0, 2.3])),
torch.sin(a),
torch.sinc(a),
torch.sinh(a),
torch.sqrt(a),
torch.square(a),
torch.sub(torch.tensor((1, 2)), torch.tensor((0, 1)), alpha=2),
torch.tan(a),
torch.tanh(a),
torch.trunc(a),
torch.xlogy(f, g),
torch.xlogy(f, g),
torch.xlogy(f, 4),
torch.xlogy(2, g),
)
torch.erfinv(torch.tensor([0, 0.5, -1.])) # exp torch.exp(torch.tensor([0, math.log(2.)])) # exp2 torch.exp2(torch.tensor([0, math.log2(2.), 3, 4])) # expm1 torch.expm1(torch.tensor([0, math.log(2.)])) # fake_quantize_per_channel_affine x = torch.randn(2, 2, 2) scales = (torch.randn(2) + 1) * 0.05 zero_points = torch.zeros(2).to(torch.long) torch.fake_quantize_per_channel_affine(x, scales, zero_points, 1, 0, 255) # fake_quantize_per_tensor_affine torch.fake_quantize_per_tensor_affine(a, 0.1, 0, 0, 255) # float_power torch.float_power(torch.randint(10, (4, )), 2) torch.float_power(torch.arange(1, 5), torch.tensor([2, -3, 4, -5])) # floor torch.floor(a) # floor_divide torch.floor_divide(torch.tensor([4., 3.]), torch.tensor([2., 2.])) torch.floor_divide(torch.tensor([4., 3.]), 1.4)
def fakeQuantizePerChannelOriginalKernel(input, scale, zero_point, axis: int,
quant_min: int, quant_max: int):
return torch.fake_quantize_per_channel_affine(input, scale, zero_point,
axis, quant_min, quant_max)
