def piecewise_mult(params,
in_tensors,
qrec: MultQuantizationRecord,
details=None):
del details
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")
func = PIECEWISE_OPS[params.__class__]
op = func['op']
if func['is_mult']:
qrec.set_scale(in_idx=(0, 1), out_idx=0)
i1 = in_tensors[0].astype(np.int32)
i2 = in_tensors[1].astype(np.int32)
res = qrec.scale_mul_biases_q.apply_scales(op(i1, i2, np.int32))
else:
# larger scale should be scaled
qrec.set_add_scale()
if qrec.scaled_idx:
i1 = in_tensors[0].astype(np.int32)
i2 = qrec.scale_in_mul_biases_q.apply_scales(in_tensors[1])
else:
i1 = qrec.scale_in_mul_biases_q.apply_scales(in_tensors[0])
i2 = in_tensors[1].astype(np.int32)
res = qrec.scale_mul_biases_q.apply_scales(op(i1, i2, None))
return qrec.get_outputs(params, [qrec.out_qs[0].clip(res)],
ktype="symmetric")
def hsigmoid_mult(params,
in_tensors,
qrec: MultQuantizationRecord,
details=None):
del details
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")[0]
fac_1, upper_bound, lower_bound = hsigmoid_mult_gen_factors(params, qrec)
in_tensor = in_tensor.astype(np.int32)
in_tensor_relued = np.minimum(np.maximum(in_tensor + fac_1, lower_bound),
upper_bound)
in_tensor = qrec.scale_mul_biases_q.apply_scales(in_tensor_relued)
return qrec.get_outputs(params, [in_tensor], ktype="symmetric")
def average_execute_mult(cls, params,
in_tensors,
qrec: MultQuantizationRecord):
# Prepare the quantization levels
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")[0]
out_dims = params.out_dims[0]
qrec.set_scale(in_idx=0, out_idx=0)
sum_by_chan = np.sum(in_tensor, dtype=np.int32, axis=tuple(
params.axis), keepdims=params.keep_dims)
sz = reduce(lambda x, y: x * y, [i for idx,
i in enumerate(in_tensor.shape) if idx in params.axis])
res = at_norm(((sum_by_chan << 7) / sz).astype(np.int32), 7)
res = out_tensor = qrec.scale_mul_biases_q.apply_scales(res)
return qrec.get_outputs(params,
[out_tensor.reshape(out_dims.shape)],
ktype="symmetric")
def average_execute(cls, params,
in_tensors,
qrec: MultQuantizationRecord):
# Prepare the quantization levels
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")[0]
out_dims = params.out_dims[0]
sum_by_chan = np.sum(in_tensor, dtype=np.int32, axis=tuple(
params.axis), keepdims=params.keep_dims)
norm = (np.array([31], dtype=np.int32) - gap_clb(sum_by_chan.flatten())).astype(np.int32)
sz = reduce(lambda x, y: x * y, [i for idx,
i in enumerate(in_tensor.shape) if idx in params.axis])
inv_wh = ((1 << norm) // sz).reshape(sum_by_chan.shape)
out_tensor = at_norm((inv_wh * sum_by_chan), norm.reshape(sum_by_chan.shape))
return qrec.get_outputs(params,
[qrec.out_qs[0].clip(out_tensor).reshape(out_dims.shape)],
ktype="symmetric")
def av_global_pool_mult(params,
in_tensors,
qrec: MultQuantizationRecord,
details=None):
# Prepare the quantization levels
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")[0]
in_dims = params.in_dims[0]
out_dims = params.out_dims[0]
qrec.set_scale(in_idx=0, out_idx=0)
sum_by_chan = np.sum(in_tensor,
dtype=np.int32,
axis=(in_dims.get_order_idx('w'),
in_dims.get_order_idx('h')))
res = at_norm((sum_by_chan << 7) // (in_dims.h * in_dims.w), 7)
res = out_tensor = qrec.scale_mul_biases_q.apply_scales(res)
return qrec.get_outputs(params, [out_tensor.reshape(out_dims.shape)],
ktype="symmetric")