def matscale(params, in_tensors, qrec: QuantizationRecordBase, details=None):
del details
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="float32")
if len(params.in_dims) == 3:
return qrec.get_outputs(
params, [in_tensors[0] * in_tensors[1] * in_tensors[2]],
ktype="float32")
return qrec.get_outputs(params, [in_tensors[0] * in_tensors[1]],
ktype="float32")
def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="float32")
if len(params.in_dims) == 3:
return qrec.get_outputs(
params, [in_tensors[0] * in_tensors[1] * in_tensors[2]],
ktype="float32")
return qrec.get_outputs(params, [in_tensors[0] * in_tensors[1]],
ktype="float32")
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 execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float32")[0]
if params.upper_bound is None:
return qrec.get_outputs(
params, [np.maximum(in_tensor, params.lower_bound)],
ktype="float32")
return qrec.get_outputs(params, [
np.minimum(np.maximum(in_tensor, params.lower_bound),
params.upper_bound)
],
ktype="float32")
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 resize_nearest_neighbor(params,
in_tensors,
qrec: QuantizationRecordBase,
details=None):
del details
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")[0]
in_tensor = in_tensor.transpose(params.in_dims[0].transpose_to_order(
("h", "w", "c")))
w_out = params.out_dims[0].w
h_out = params.out_dims[0].h
c_out = params.out_dims[0].c
w_in = params.in_dims[0].w
h_in = params.in_dims[0].h
wstep = ((w_in - 1) << 16) // (w_out - 1)
hstep = ((h_in - 1) << 16) // (h_out - 1)
out_tensor = np.empty((h_out, w_out, c_out))
for i in range(h_out):
h_rounded = ((hstep * i) + (1 << (16 - 1))) >> 16
for j in range(w_out):
w_rounded = ((wstep * j) + (1 << (16 - 1))) >> 16
out_tensor[i, j, :] = in_tensor[h_rounded, w_rounded, :]
out_tensor = out_tensor.transpose(params.out_dims[0].transpose_from_order(
("h", "w", "c")))
return qrec.get_outputs(params, [out_tensor], ktype="symmetric")
def av_global_pool(params,
in_tensors,
qrec: QuantizationRecordBase,
details=None):
if isinstance(qrec, MultQuantizationRecord):
return av_global_pool_mult(params, in_tensors, qrec, details=details)
# 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]
sum_by_chan = np.sum(in_tensor,
dtype=np.int32,
axis=(in_dims.get_order_idx('w'),
in_dims.get_order_idx('h')))
norm = (np.array([31], dtype=np.int32) - gap_clb(sum_by_chan)).astype(
np.int32)
inv_wh = (1 << norm) // (in_dims.h * in_dims.w)
out_tensor = at_norm((inv_wh * sum_by_chan), norm)
return qrec.get_outputs(
params, [qrec.out_qs[0].clip(out_tensor).reshape(out_dims.shape)],
ktype="symmetric")
def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float32")[0]
in_tensor = in_tensor.transpose(params.in_dims[0].transpose_to_order(
("h", "w", "c")))
w_out = params.out_dims[0].w
h_out = params.out_dims[0].h
c_out = params.out_dims[0].c
w_in = params.in_dims[0].w
h_in = params.in_dims[0].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(
params.out_dims[0].transpose_from_order(("h", "w", "c")))
return qrec.get_outputs(params, [out_tensor], ktype="float32")
def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float32")[0]
in_tensor = in_tensor.transpose(params.in_dims[0].transpose_to_order(
("h", "w", "c")))
w_out = params.out_dims[0].w
h_out = params.out_dims[0].h
c_out = params.out_dims[0].c
w_in = params.in_dims[0].w
h_in = params.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(
params.out_dims[0].transpose_from_order(("h", "w", "c")))
return qrec.get_outputs(params, [out_tensor], ktype="float32")
def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
if qrec is None:
qrec = Float32QuantizationRecord()
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="float32")
in_vars = {
params.input_symbols[i]: in_tensor
for i, in_tensor in enumerate(in_tensors)
}
out_vars = params.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="float32")
def expression(params, in_tensors, qrec: QuantizationRecordBase, details=None):
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="float32")
return qrec.get_outputs(params,
params.execute(in_tensors, details=details),
ktype="float32")
def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
qname = kwargs['qname']
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype=qname)[0]
out_tensors = [in_tensor]
return qrec.get_outputs(params, out_tensors, ktype=qname)
def execute(cls, params,
in_tensors,
qrec: QuantizationRecordBase,
**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 cast(params, in_tensors, qrec: QuantizationRecordBase, details=None):
del details
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float32")[0]
out_tensors = [in_tensor]
return qrec.get_outputs(params, out_tensors, ktype="float32")
def softmax(params, in_tensors, qrec: QuantizationRecordBase, details=None):
del details
if qrec is None:
qrec = Float32QuantizationRecord()
np.seterr(over='raise')
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float32")[0]
return qrec.get_outputs(params, [softmax_func(in_tensor)], ktype="float32")
def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
qname = kwargs['qname']
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype=qname)[0]
return qrec.get_outputs(params, [np.flip(in_tensor, axis=params.axis)],
ktype=qname)
def sigmoid(params, in_tensors, qrec: QuantizationRecordBase, details=None):
del details
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float32")[0]
return qrec.get_outputs(params, [1 / (1 + np.exp(-in_tensor))],
ktype="float32")
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):
qname = kwargs['qname']
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype=qname)[0]
out_tensor = np.take(in_tensor, params.indices, axis=params.axis)
return qrec.get_outputs(params, [out_tensor], ktype=qname)
def transpose(params, in_tensors, qrec: QuantizationRecordBase, details=None):
del details
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float32")[0]
if params.transpose_in:
in_tensor = np.transpose(in_tensor, params.transpose_in[0])
return qrec.get_outputs(params, [in_tensor], ktype="float32")
def strided_slice(params,
in_tensors,
qrec: QuantizationRecordBase,
details=None):
del details
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="symmetric")[0]
out_tensors = [params.numpy_slice(in_tensor)]
return qrec.get_outputs(params, out_tensors, ktype="symmetric")
def leaky(params, in_tensors, qrec: QuantizationRecordBase, details=None):
del details
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float32")[0]
output = in_tensor * (in_tensor >
0) + in_tensor * params.leak_factor * (in_tensor < 0)
return qrec.get_outputs(params, [output], ktype="float32")
def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
qname = kwargs['qname']
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype=qname)[0]
if params.transpose_in:
in_tensor = np.transpose(in_tensor, params.transpose_in[0])
return qrec.get_outputs(params, [in_tensor], ktype=qname)
def hsigmoid(params, in_tensors, qrec: QuantizationRecordBase, details=None):
del details
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float32")[0]
return qrec.get_outputs(
params, [np.minimum(np.maximum(in_tensor + params.offset, 0), 6) / 6],
ktype="float32")
def execute_piecewise(cls, params, in_tensors,
qrec: QuantizationRecordBase, op, **kwargs):
del kwargs
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="float32")
return qrec.get_outputs(params, [op(in_tensors[0], in_tensors[1])],
ktype="float32")
def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
qname = kwargs['qname']
params = typing_cast(StridedSliceParameters, params)
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype=qname)[0]
out_tensors = [params.numpy_slice(in_tensor)]
return qrec.get_outputs(params, out_tensors, ktype=qname)
def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float32")[0]
return qrec.get_outputs(params, [1 / (1 + np.exp(-in_tensor))],
ktype="float32")
def relu(params: ReluActivationParameters,
in_tensors,
qrec: QuantizationRecordBase,
details=None):
del details
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensor = qrec.prepare_inputs(params, in_tensors, ktype="float32")[0]
if params.upper_bound is None:
return qrec.get_outputs(params,
[np.maximum(in_tensor, params.lower_bound)],
ktype="float32")
return qrec.get_outputs(params, [
np.minimum(np.maximum(in_tensor, params.lower_bound),
params.upper_bound)
],
ktype="float32")
def execute(cls, params, in_tensors, qrec: QuantizationRecordBase,
**kwargs):
if qrec is None:
qrec = Float32QuantizationRecord()
in_tensors = qrec.prepare_inputs(params, in_tensors, ktype="float32")
return qrec.get_outputs(params,
params.execute(in_tensors),
ktype="float32")
def constant_input(params,
in_tensors,
qrec: QuantizationRecordBase,
details=None):
del in_tensors, details
if qrec is None:
qrec = Float32QuantizationRecord()
return qrec.get_outputs(params, [params.value], ktype="float32")
