def test_get_quantized_range(num_bits, signed, restrict, expected_q_min,
expected_q_max):
q_min, q_max = qu.get_quantized_range(num_bits,
signed=signed,
signed_restrict_qrange=restrict)
assert q_min == expected_q_min
assert q_max == expected_q_max
def attach_quant_metadata(t, num_bits, quant_mode, stats=None, clip_mode=ClipMode.NONE, per_channel=False,
num_stds=None, scale_approx_mult_bits=None):
if stats is None:
scale, zp = _get_quant_params_from_tensor(t, num_bits, quant_mode, clip_mode, per_channel, num_stds,
scale_approx_mult_bits)
else:
scale, zp = _get_quant_params_from_stats_dict(stats, num_bits, quant_mode, clip_mode, num_stds,
scale_approx_mult_bits)
signed = quant_mode != LinearQuantMode.ASYMMETRIC_UNSIGNED
restrict = quant_mode == LinearQuantMode.SYMMETRIC_RESTRICTED
min_q_val, max_q_val = q_utils.get_quantized_range(num_bits, signed)
t.quant_metadata = TensorQuantMetadata(scale, zp, min_q_val, max_q_val)
return t
def _fake_quant_tensor(tensor, n_bits, mode, per_channel):
q_min, q_max = q_utils.get_quantized_range(
n_bits, mode != LinearQuantMode.ASYMMETRIC_UNSIGNED)
scale, zp = _get_quant_params_from_tensor(tensor,
n_bits,
mode,
per_channel=per_channel)
q_utils.linear_quantize_clamp(tensor,
scale,
zp,
q_min,
q_max,
inplace=True)
q_utils.linear_dequantize(tensor, scale, zp, inplace=True)