def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
qname = kwargs['qname']
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype=qname)
if qrec:
in_q = qrec.in_qs[0]
out_q = qrec.out_qs[0]
float_conversion = in_q.is_floating or out_q.is_floating
bit_conversion = in_q.bits != out_q.bits
if not float_conversion:
same_sign = in_q.signed == out_q.signed
if in_q.bits > out_q.bits:
bit_diff = in_q.bits - out_q.bits
same_scale = np.allclose(in_q.scale *
np.power(2, bit_diff),
out_q.scale,
atol=0.0001)
same_zeropoint = np.all(
in_q.zero_point >> bit_diff == out_q.zero_point)
elif out_q.bits > in_q.bits:
bit_diff = out_q.bits - in_q.bits
same_scale = np.allclose(out_q.scale *
np.power(2, bit_diff),
in_q.scale,
atol=0.0001)
same_zeropoint = np.all(
in_q.zero_point == out_q.zero_point >> bit_diff)
else:
same_scale = np.allclose(out_q.scale,
in_q.scale,
atol=0.0001)
same_zeropoint = np.all(
in_q.zero_point == out_q.zero_point)
if same_scale and same_sign and bit_conversion and same_zeropoint:
if in_q.bits > out_q.bits:
if in_q.signed:
out_tensor = out_q.clip(
at_norm(in_tensors[0].astype(np.int32),
in_q.bits - out_q.bits))
else:
out_tensor = out_q.clip(
at_norm(in_tensors[0].astype(np.uint32),
in_q.bits - out_q.bits))
else:
out_tensor = in_tensors[0].astype(
out_q.dtype) << (out_q.bits - in_q.bits)
return qrec.get_outputs(params, [out_tensor], ktype=qname)
# in all other conversions should be numerically equivalent to this (within 1 bit)
out_tensor = qrec.out_qs[0].quantize_from(in_tensors[0],
qrec.in_qs[0])
else:
out_tensor = in_tensors[0]
return qrec.get_outputs(params, [out_tensor], ktype=qname)
def execute(cls, params,
in_tensors,
qrec: QRec,
**kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="float")
if len(params.in_dims) == 3:
return qrec.get_outputs(params, [in_tensors[0] * in_tensors[1] * in_tensors[2]], ktype="float")
return qrec.get_outputs(params, [in_tensors[0] * in_tensors[1]], ktype="float")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
in_tensor = qrec.prepare_inputs(params, in_tensors,
ktype="symmetric")[0]
compute_in_out_scale(qrec)
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)
scale_mul_biases_q = qrec.cache['scale_mul_biases_q']
in_tensor = 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 execute(cls, params, in_tensors, qrec: QRec, **kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float")[0]
if params.upper_bound is None:
return qrec.get_outputs(
params, [np.maximum(in_tensor, params.lower_bound)],
ktype="float")
return qrec.get_outputs(params, [
np.minimum(np.maximum(in_tensor, params.lower_bound),
params.upper_bound)
],
ktype="float")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
in_tensors = [
in_tensor.astype(np.int32) for in_tensor in qrec.prepare_inputs(
params, in_tensors, ktype="symmetric")
]
if isinstance(params, MatMulTransposedParameters):
mat1, mat2 = in_tensors[0], np.transpose(in_tensors[1], (1, 0))
else:
mat1, mat2 = in_tensors[0], in_tensors[1]
mat2 = mat2.astype(np.int32) - qrec.in_qs[1].zero_point.astype(
np.int32)
if len(in_tensors) > 2:
biases = in_tensors[2]
if len(biases.shape) == 1:
if biases.shape[0] == mat1.shape[0]:
biases = np.expand_dims(biases, -1)
else:
biases = 0
out_tensor = np.matmul(mat1, mat2) + biases
mul_biases_q = qrec.cache['mul_biases_q']
scale_axis = None if len(mul_biases_q.scale) == 1 else 1
out_tensor = mul_biases_q.apply_scales(out_tensor, scale_axis)
return qrec.get_outputs(params, [out_tensor], ktype="symmetric")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
qname = kwargs['qname']
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype=qname)[0]
return qrec.get_outputs(
params, [in_tensor * np.ones(params.shape, dtype=in_tensor.dtype)],
ktype=qname)
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
qname = kwargs['qname']
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype=qname)[0]
if params.transpose:
in_tensor = np.transpose(in_tensor, params.transpose)
return qrec.get_outputs(params, [in_tensor], ktype=qname)
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float")[0]
return qrec.get_outputs(params, [1 / (1 + np.exp(-in_tensor))],
ktype="float")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
in_tensors = [
in_tensor.astype(np.int32) for in_tensor in qrec.prepare_inputs(
params, in_tensors, ktype="symmetric")
]
if isinstance(params, MatMulTransposedParameters):
mat1, mat2 = in_tensors[0], np.transpose(in_tensors[1], (1, 0))
else:
mat1, mat2 = in_tensors[0], in_tensors[1]
if len(in_tensors) > 2:
biases = in_tensors[2]
if len(biases.shape) == 1:
if biases.shape[0] == mat1.shape[0]:
biases = np.expand_dims(biases, -1)
else:
biases = 0
# expect biases in in_q1 + in_q2
q_calc = QType.Pow2(bits=32,
q=qrec.in_qs[0].q + qrec.in_qs[1].q,
signed=True)
out_tensor = np.matmul(mat1, mat2) + biases
out_tensor = qrec.out_qs[0].reduce_from(out_tensor, q_calc)
return qrec.get_outputs(params, [out_tensor], ktype="symmetric")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float")[0]
in_dim, out_dim = params.in_dims[0], params.out_dims[0]
in_tensor = in_tensor.transpose(
in_dim.transpose_to_order(("h", "w", "c")))
w_out = out_dim.w
h_out = out_dim.h
c_out = out_dim.c
w_in = in_dim.w
h_in = in_dim.h
wstep = (w_in - 1) / (w_out - 1)
hstep = (h_in - 1) / (h_out - 1)
out_tensor = np.empty((h_out, w_out, c_out))
for i in range(h_out):
h_rounded = int(round(hstep * i))
for j in range(w_out):
w_rounded = int(round(wstep * j))
out_tensor[i, j, :] = in_tensor[h_rounded, w_rounded, :]
out_tensor = out_tensor.transpose(
out_dim.transpose_from_order(("h", "w", "c")))
return qrec.get_outputs(params, [out_tensor], ktype="float")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
in_tensor = qrec.prepare_inputs(params, in_tensors,
ktype="symmetric")[0]
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:
relu_ub = qrec.in_qs[0].quantize(params.upper_bound)
in_tensor = np.minimum(in_tensor, relu_ub)
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 execute(cls, params,
in_tensors,
qrec: QRec,
**kwargs):
if qrec is None:
qrec = AllFloatQRec()
details = kwargs.get('details')
if details is not None:
current_control = SymbolStats()
Symbol.set_default_control(current_control)
results = {}
else:
results = None
current_control = None
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="float")
in_vars = {params.input_symbols[i]: in_tensor
for i, in_tensor in enumerate(in_tensors)}
func_col = qrec.cache.get('qfunc_col')
if func_col is None:
func_col = params.func_col
out_vars = func_col(**in_vars,
calculate_ranges=current_control is not None,
track_results=results)
out_tensors = [out_vars[out_sym_name]
for out_sym_name in params.output_symbols]
if current_control:
details.update(current_control.stats)
details['results'] = results
return qrec.get_outputs(params, out_tensors, ktype="float")
def execute(cls, params,
in_tensors,
qrec: QRec,
**kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="float")
if isinstance(params, MatMulTransposedParameters):
mat1, mat2 = in_tensors[0], np.transpose(in_tensors[1], (1, 0))
else:
mat1, mat2 = in_tensors[0], in_tensors[1]
if len(in_tensors) > 2:
biases = in_tensors[2]
if len(biases.shape) == 1:
if biases.shape[0] == mat1.shape[0]:
biases = np.expand_dims(biases, -1)
else:
biases = 0
out_dtype = qrec.out_qs[0].dtype if qrec.ktype.startswith(
'float') else np.float32
output_tensor = np.matmul(mat1, mat2).astype(
out_dtype) + np.atleast_1d(biases).astype(out_dtype)
return qrec.get_outputs(params, [output_tensor], ktype="float")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float")[0]
in_dtype = qrec.in_qs[0].dtype if qrec.ktype.startswith(
'float') else np.float32
if params.upper_bound is None:
return qrec.get_outputs(
params, [np.maximum(in_tensor, in_dtype(params.lower_bound))],
ktype="float")
return qrec.get_outputs(params, [
np.minimum(np.maximum(in_tensor, in_dtype(params.lower_bound)),
in_dtype(params.upper_bound))
],
ktype="float")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float")[0]
in_dims, out_dims = cls.calc_transposed_dims(params)
in_tensor = in_tensor.transpose(in_dims[0].transpose_to_order(
("h", "w", "c")))
w_out = out_dims[0].w
h_out = out_dims[0].h
c_out = out_dims[0].c
w_in = in_dims[0].w
h_in = in_dims[0].h
wstep = (w_in - 1) / w_out
hstep = (h_in - 1) / h_out
out_tensor = np.empty((h_out, w_out, c_out))
for i in range(h_out):
y_l, y_h = math.floor(hstep * i), math.ceil(hstep * i)
hc = (hstep * i) - y_l
for j in range(w_out):
x_l, x_h = math.floor(wstep * j), math.ceil(wstep * j)
wc = (wstep * j) - x_l
P1 = in_tensor[y_l, x_l, :]
P2 = in_tensor[y_l, x_h, :]
P3 = in_tensor[y_h, x_l, :]
P4 = in_tensor[y_h, x_h, :]
out_tensor[i, j, :] = P1 * (1 - wc) * (1 - hc) \
+ P2 * wc * (1 - hc) \
+ P3 * (1 - wc) * hc \
+ P4 * wc * hc
out_tensor = out_tensor.transpose(out_dims[0].transpose_from_order(
("h", "w", "c")))
return qrec.get_outputs(params, [out_tensor], ktype="float")
def execute(cls, params,
in_tensors,
qrec: QRec,
**kwargs):
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")
func = PIECEWISE_OPS[params.__class__]
op = func['op']
if func['is_mult']:
i1 = in_tensors[0].astype(np.int32)
i2 = in_tensors[1].astype(np.int32)
res = op(i1, i2, np.int32)
q_calc = QType.Pow2(
bits=32, q=qrec.in_qs[0].q+qrec.in_qs[1].q, signed=True)
res = qrec.out_qs[0].reduce_from(res, q_calc)
else:
off_in = abs(qrec.in_qs[0].q - qrec.in_qs[1].q)
if qrec.in_qs[0].q > qrec.in_qs[1].q:
i1 = at_norm(in_tensors[0].astype(np.int32), off_in)
i2 = in_tensors[1].astype(np.int32)
else:
i1 = in_tensors[0].astype(np.int32)
i2 = at_norm(in_tensors[1].astype(np.int32), off_in)
res = op(i1, i2, None)
q_calc = QType.Pow2(bits=32, q=min(qrec.in_qs[0].q, qrec.in_qs[1].q), signed=True)
res = qrec.out_qs[0].reduce_from(res, q_calc)
return qrec.get_outputs(params, [res], ktype="symmetric")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
qname = kwargs['qname']
params = typing_cast(SplitParameters, params)
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype=qname)[0]
out_tensors = params.numpy_split(in_tensor)
return qrec.get_outputs(params, out_tensors, ktype=qname)
def execute(cls, params,
in_tensors,
qrec: QRec,
**kwargs):
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")
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']:
compute_in_out_scale(qrec, in_idx=(0, 1), out_idx=0)
scale_mul_biases_q = qrec.cache['scale_mul_biases_q']
i1 = in_tensors[0].astype(np.int32)
i2 = in_tensors[1].astype(np.int32)
out_tensor = scale_mul_biases_q.apply_scales(op(i1, i2, np.int32))
else:
# larger scale should be scaled
set_add_in_scale(qrec)
scale_mul_biases_q = qrec.cache['scale_mul_biases_q']
if qrec.cache['scaled_idx']:
i1 = in_tensors[0].astype(np.int32)
i2 = qrec.cache['scale_in_mul_biases_q'].apply_scales(in_tensors[1])
else:
i1 = qrec.cache['scale_in_mul_biases_q'].apply_scales(in_tensors[0])
i2 = in_tensors[1].astype(np.int32)
out_tensor = scale_mul_biases_q.apply_scales(op(i1, i2, None))
return qrec.get_outputs(params, [qrec.out_qs[0].clip(out_tensor)], ktype="symmetric")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
in_tensor = qrec.prepare_inputs(params, in_tensors,
ktype="symmetric")[0]
compute_in_out_scale(qrec)
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)
scale_mul_biases_q = qrec.cache['scale_mul_biases_q']
in_tensor = 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 execute(cls, params, in_tensors, qrec: QRec, **kwargs):
boxes = in_tensors[0][
0] if not params.center_point_box else convert_cnts2cors(
in_tensors[0][0])
scores = in_tensors[1][0]
n_boxes = len(scores[0])
n_classes = len(scores)
scores_q = qrec.in_qs[1]
indexes = np.zeros((params.max_output_boxes_per_class * n_classes, 3))
idxs_count = 0
for class_id in range(n_classes):
bbox_buff = []
counter = 0
for box_id in range(n_boxes):
class_score = scores[class_id, box_id]
if class_score > scores_q.quanize(params.nms_score_threshold):
bbox_buff.append({
"index": box_id,
"score": class_score,
"box": boxes[box_id],
"alive": True
})
counter += 1
# Bubble sort to sort the scores
changed = True
while changed:
changed = False
for i in range(counter - 1):
if bbox_buff[i]["score"] < bbox_buff[i + 1]["score"]:
temp = bbox_buff[i]
bbox_buff[i] = bbox_buff[i + 1]
bbox_buff[i + 1] = temp
changed = True
# NMS
for idx in range(counter):
for idx_int in range(idx + 1, counter):
if not bbox_buff[idx_int]["alive"]:
continue
intersection = rect_intersect_area(
bbox_buff[idx]["box"], bbox_buff[idx_int]["box"])
union = rect_union_area(bbox_buff[idx]["box"],
bbox_buff[idx_int]["box"])
if intersection >= (scores_q.quanize(
params.nms_iou_threshold) * union):
bbox_buff[idx_int]["alive"] = False
class_idxs_count_start = idxs_count
for bb in bbox_buff:
if (idxs_count - class_idxs_count_start
) > params.max_output_boxes_per_class:
break
if bb["alive"]:
indexes[idxs_count] = np.array([0, class_id, bb['index']])
idxs_count += 1
return qrec.get_outputs(params, [np.array(indexes)], ktype="float")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float")[0]
return qrec.get_outputs(params,
[np.minimum(np.maximum(in_tensor, -1.0), 1.0)],
ktype="float")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float")[0]
output = in_tensor * (in_tensor > 0) + in_tensor * \
params.leak_factor * (in_tensor < 0)
return qrec.get_outputs(params, [output], ktype="float")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
in_tensor = qrec.prepare_inputs(params, in_tensors,
ktype="symmetric")[0]
upper_bound = qrec.in_qs[0].quantize(np.array([1.]))
in_tensor = np.minimum(np.maximum(in_tensor, -upper_bound),
upper_bound)
return qrec.get_outputs(params, [in_tensor], ktype="symmetric")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
if qrec is None:
qrec = AllFloatQRec()
old_err = np.seterr(over='raise')
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float")[0]
in_tensor = softmax_func(in_tensor, axis=params.axis)
np.seterr(**old_err)
return qrec.get_outputs(params, [in_tensor], ktype="float")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")
LOG.debug("matscale input %s",
",".join([t.dtype.name for t in in_tensors]))
if len(params.in_dims) == 3:
output_tensor = cls.matscale3(in_tensors, qrec)
else:
output_tensor = cls.matscale2(in_tensors, qrec)
return qrec.get_outputs(params, [output_tensor], ktype="symmetric")
def sigmoid(params, in_tensors, qrec: QRec, details=None):
del details
if qrec.ktype == 'scaled':
raise NotImplementedError()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")[0]
dqinput = qrec.in_qs[0].dequantize(in_tensor)
return qrec.get_outputs(
params, [qrec.out_qs[0].quantize(1 / (1 + np.exp(-dqinput)))],
ktype="symmetric")
def sum_execute(cls, params, in_tensors, qrec: QRec, **kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float")[0]
return qrec.get_outputs(params, [
np.sum(
in_tensor, axis=tuple(params.axis), keepdims=params.keep_dims)
],
ktype="float")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
old_err = np.seterr(over='raise')
in_tensor = qrec.prepare_inputs(params, in_tensors,
ktype="symmetric")[0]
# TODO - Implement properly quantized version
in_tensor = qrec.in_qs[0].dequantize(in_tensor)
in_tensor = qrec.out_qs[0].quantize(softmax_func(in_tensor))
np.seterr(**old_err)
return qrec.get_outputs(params, [in_tensor], ktype="symmetric")
def execute(cls, params, in_tensors, qrec: QRec, **kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float")[0]
in_dtype = qrec.in_qs[0].dtype if qrec.ktype.startswith(
'float') else np.float32
if in_dtype == np.float32:
return qrec.get_outputs(params,
[np.tanh(in_tensor).astype(in_dtype)],
ktype="float")
else:
if qrec.cache.get('kernel_type') == "lut":
return qrec.get_outputs(params,
[tanh_lut_float(in_tensor, in_dtype)],
ktype="float")
return qrec.get_outputs(
params, [np_fasttanh(in_tensor, dtype=in_dtype, doalt=True)],
ktype="float")
def execute(cls, params,
in_tensors,
qrec: QRec,
**kwargs):
if qrec is None:
qrec = AllFloatQRec()
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="float")
output = cls.FUNC(*in_tensors)
return qrec.get_outputs(params, [output], ktype="float")