def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")
if params.transpose_in:
in_tensors = [(np.transpose(in_tensor, params.transpose_in[idx])
if params.transpose_in[idx] else in_tensor)
for idx, in_tensor in enumerate(in_tensors)]
if isinstance(params, Broadcastable) and params.is_broadcasted:
in_tensors = params.broadcast_inputs(in_tensors)
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)
out_tensor = 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)
out_tensor = qrec.scale_mul_biases_q.apply_scales(op(i1, i2, None))
if params.transpose_out:
out_tensor = np.transpose(out_tensor, params.transpose_out[0])
return qrec.get_outputs(params, [qrec.out_qs[0].clip(out_tensor)],
ktype="symmetric")
def execute(cls, params,
in_tensors,
qrec: QuantizationRecordBase,
**kwargs):
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")[0]
qrec.set_scale()
neg_in = at_norm(in_tensor * leak_mult_gen_factor_q7(params), 7)
in_tensor = in_tensor * (in_tensor > 0) + neg_in * (in_tensor < 0)
in_tensor = qrec.scale_mul_biases_q.apply_scales(in_tensor)
if qrec.out_qs[0] != qrec.in_qs[0]:
return qrec.get_outputs(params, [qrec.out_qs[0].reduce_from(in_tensor, qrec.in_qs[0])], ktype="symmetric")
return qrec.get_outputs(params, [in_tensor], ktype="symmetric")
def sum_execute(cls, params, in_tensors, qrec: QuantizationRecordBase):
in_tensor = qrec.prepare_inputs(params, in_tensors,
ktype="symmetric")[0]
if isinstance(qrec, MultQuantizationRecord):
qrec.set_scale(in_idx=0, out_idx=0)
res = np.sum(in_tensor,
axis=tuple(params.axis),
keepdims=params.keep_dims,
dtype=np.int32)
if isinstance(qrec, MultQuantizationRecord):
res = qrec.scale_mul_biases_q.apply_scales(res)
return qrec.get_outputs(params, [res], ktype="symmetric")
def max_execute(cls, params,
in_tensors,
qrec: QuantizationRecordBase,
details=None):
del details
# Prepare the quantization levels
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")[0]
if isinstance(qrec, MultQuantizationRecord):
qrec.set_scale(in_idx=0, out_idx=0)
return qrec.get_outputs(params, [np.max(in_tensor,
axis=tuple(params.axis),
keepdims=params.keep_dims)], ktype="symmetric")
def execute(cls, params,
in_tensors,
qrec: QuantizationRecordBase,
**kwargs):
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")[0]
qrec.set_scale()
relu_lb = qrec.in_qs[0].quantize(params.lower_bound)
in_tensor = np.maximum(in_tensor, relu_lb)
if params.upper_bound is not None and not NNForceRelu.FORCE_RELU:
relu_ub = qrec.in_qs[0].quantize(params.upper_bound)
in_tensor = np.minimum(in_tensor, relu_ub)
in_tensor = qrec.scale_mul_biases_q.apply_scales(in_tensor)
if qrec.out_qs[0] != qrec.in_qs[0]:
return qrec.get_outputs(params, [qrec.out_qs[0].reduce_from(in_tensor, qrec.in_qs[0])], ktype="symmetric")
return qrec.get_outputs(params, [in_tensor], ktype="symmetric")
def max_global_pool(params,
in_tensors,
qrec: QuantizationRecordBase,
details=None):
del details
# Prepare the quantization levels
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")[0]
if isinstance(qrec, MultQuantizationRecord):
qrec.set_scale(in_idx=0, out_idx=0)
in_dims = params.in_dims[0]
return qrec.get_outputs(params, [
np.max(in_tensor,
axis=(in_dims.get_order_idx('w'), in_dims.get_order_idx('h')),
keepdims=True)
],
ktype="symmetric")
def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
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")
