def from_observed(cls, other):
converted = torch.quantization.convert(other,
mapping=None,
inplace=False,
remove_qconfig=True,
convert_custom_config_dict=None)
# Remove the parameters for the bias_k and bias_v to quantize them
# TODO: This is a potential source of accuracy drop.
# quantized cat takes the scale and zp of the first
# element, which might lose the precision in the bias_k
# and the bias_v (which are cat'ed with k/v being first).
if converted.bias_k is not None:
bias_k = converted._parameters.pop('bias_k')
sc, zp = torch._choose_qparams_per_tensor(bias_k,
reduce_range=False)
bias_k = torch.quantize_per_tensor(bias_k, sc, zp, torch.quint8)
setattr(converted, 'bias_k', bias_k) # noqa: B010
if converted.bias_v is not None:
bias_v = converted._parameters.pop('bias_v')
sc, zp = torch._choose_qparams_per_tensor(bias_k,
reduce_range=False)
bias_v = torch.quantize_per_tensor(bias_v, sc, zp, torch.quint8)
setattr(converted, 'bias_v', bias_v) # noqa: B010
return converted
def test_choose_qparams(self, X, reduce_range):
X, (scale, zero_point, torch_type) = X
X = torch.from_numpy(X)
X_scale, X_zp = _calculate_dynamic_qparams(X, torch.quint8, reduce_range=reduce_range)
qparams = torch._choose_qparams_per_tensor(X, reduce_range)
np.testing.assert_array_almost_equal(X_scale, qparams[0], decimal=3)
self.assertEqual(X_zp, qparams[1])
def test_per_tensor_dynamic_quant_observers(self, X, reduce_range):
X, (scale, zero_point, torch_type) = X
x = torch.from_numpy(X)
obs = MinMaxDynamicQuantObserver(dtype=torch.quint8, reduce_range=reduce_range)
result = obs(x)
qparams = obs.calculate_qparams()
ref = torch._choose_qparams_per_tensor(x, reduce_range)
self.assertEqual(ref[0], qparams[0])
self.assertEqual(ref[1], qparams[1])
