def leaky_relu(g, x, negative_slope, inplace, op_scale, op_zero_point):
x, _, _, _ = symbolic_helper.dequantize_helper(g, x)
output = opset9.leaky_relu(g, x, negative_slope, inplace)
return symbolic_helper.quantize_helper(g, output, op_scale,
op_zero_point)
def mul(g, x, y, op_scale, op_zero_point):
x, _, _, _ = sym_help.dequantize_helper(g, x)
y, _, _, _ = sym_help.dequantize_helper(g, y)
output = mul(g, x, y)
return sym_help.quantize_helper(g, output, op_scale, op_zero_point)
def conv2d(
g,
q_input,
q_weight,
bias,
stride,
padding,
dilation,
groups,
op_scale,
op_zero_point,
):
input, input_scale, _, _ = symbolic_helper.dequantize_helper(
g, q_input)
weight, weight_scale, _, _ = symbolic_helper.dequantize_helper(
g, q_weight)
q_bias = symbolic_helper.requantize_bias_helper(
g, bias, input_scale, weight_scale)
bias, _, _, _ = symbolic_helper.dequantize_helper(g, q_bias)
output = opset9.conv2d(g, input, weight, bias, stride, padding,
dilation, groups)
return symbolic_helper.quantize_helper(g, output, op_scale,
op_zero_point)
def quantize_per_tensor(g, input, scale, zero_point, dtype):
dtype = symbolic_helper._get_const(dtype, "i", "dtype")
zero_point = g.op("Cast",
zero_point,
to_i=symbolic_helper.scalar_type_to_onnx[dtype])
scale = g.op("Cast", scale, to_i=_C_onnx.TensorProtoDataType.FLOAT)
return symbolic_helper.quantize_helper(g, input, scale, zero_point)
def hardswish(g, x, op_scale, op_zero_point):
x, _, _, _ = symbolic_helper.dequantize_helper(g, x)
output = opset9.hardswish(g, x)
return symbolic_helper.quantize_helper(g, output, op_scale,
op_zero_point)
def sigmoid(g, x, op_scale, op_zero_point):
x, _, _, _ = symbolic_helper.dequantize_helper(g, x)
output = opset9.sigmoid(g, x)
return symbolic_helper.quantize_helper(g, output, op_scale,
op_zero_point)
def quantize_per_tensor(g, input, scale, zero_point, dtype):
dtype = symbolic_helper._get_const(dtype, "i", "dtype")
# TODO(justinchuby): Extract all the cast ops into a helper function.
zero_point = g.op("Cast",
zero_point,
to_i=_type_utils.JitScalarType(dtype).onnx_type())
scale = g.op("Cast", scale, to_i=_C_onnx.TensorProtoDataType.FLOAT)
return symbolic_helper.quantize_helper(g, input, scale, zero_point)
def mul(g, x, y, op_scale, op_zero_point):
x, _, _, _ = symbolic_helper.dequantize_helper(g, x)
y, _, _, _ = symbolic_helper.dequantize_helper(g, y)
output = opset9.mul(g, x, y)
return symbolic_helper.quantize_helper(g, output, op_scale,
op_zero_point)
def group_norm(g, x, num_groups, weight, bias, eps, op_scale,
op_zero_point):
x, _, _, _ = symbolic_helper.dequantize_helper(g, x)
output = opset9.group_norm(g, x, num_groups, weight, bias, eps, False)
return symbolic_helper.quantize_helper(g, output, op_scale,
op_zero_point)
def add_relu(g, x, y, op_scale, op_zero_point):
x, _, _, _ = symbolic_helper.dequantize_helper(g, x)
y, _, _, _ = symbolic_helper.dequantize_helper(g, y)
output = opset9.add(g, x, y)
output = opset9.relu(g, output)
return symbolic_helper.quantize_helper(g, output, op_scale,
op_zero_point)
def linear(g, q_input, q_weight, bias, op_scale, op_zero_point):
input, input_scale, _ = sym_help.dequantize_helper(g, q_input)
weight, weight_scale, _ = sym_help.dequantize_helper(g, q_weight)
q_bias = sym_help.requantize_bias_helper(g, bias, input_scale, weight_scale)
bias, _, _ = sym_help.dequantize_helper(g, q_bias)
output = linear(g, input, weight, bias)
return sym_help.quantize_helper(g, output, op_scale, op_zero_point)
def layer_norm(g, x, normalized_shape, weight, bias, eps, op_scale,
op_zero_point):
x, _, _, _ = symbolic_helper.dequantize_helper(g, x)
output = opset9.layer_norm(g, x, normalized_shape, weight, bias, eps,
False)
return symbolic_helper.quantize_helper(g, output, op_scale,
op_zero_point)
def conv2d_relu(g, q_input, q_weight, bias, stride, padding, dilation, groups, op_scale, op_zero_point):
input, input_scale, _ = sym_help.dequantize_helper(g, q_input)
weight, weight_scale, _ = sym_help.dequantize_helper(g, q_weight)
q_bias = sym_help.requantize_bias_helper(g, bias, input_scale, weight_scale)
bias, _, _ = sym_help.dequantize_helper(g, q_bias)
output = conv2d(g, input, weight, bias, stride, padding, dilation, groups)
output = relu(g, output)
return sym_help.quantize_helper(g, output, op_scale, op_zero_point)
def linear(g, q_input, q_weight, bias, op_scale, op_zero_point):
input, input_scale, _, _ = symbolic_helper.dequantize_helper(g, q_input)
weight, weight_scale, _, axis = symbolic_helper.dequantize_helper(g, q_weight)
q_bias = symbolic_helper.requantize_bias_helper(
g, bias, input_scale, weight_scale, axis
)
bias, _, _, _ = symbolic_helper.dequantize_helper(g, q_bias)
output = opset9.linear(g, input, weight, bias)
return symbolic_helper.quantize_helper(g, output, op_scale, op_zero_point)
def cat(
g,
q_inputs: _C.Value,
dim: int,
op_scale: _C.Value,
op_zero_point: _C.Value,
) -> _C.Value:
unpacked_inputs = symbolic_helper._unpack_list(q_inputs)
dequantized = [
symbolic_helper.dequantize_helper(g, input)[0]
for input in unpacked_inputs
]
concatenated = g.op("Concat", *dequantized, axis_i=dim)
return symbolic_helper.quantize_helper(g, concatenated, op_scale,
op_zero_point)
def instance_norm(
g,
q_input,
weight,
bias,
eps,
op_scale,
op_zero_point,
):
input, _, _, _ = symbolic_helper.dequantize_helper(g, q_input)
output = opset9.instance_norm(g, input, weight, bias, None, None,
False, 0, eps, False)
return symbolic_helper.quantize_helper(g, output, op_scale,
op_zero_point)
def hardswish(g, x, op_scale, op_zero_point):
x, _, _ = sym_help.dequantize_helper(g, x)
output = hardswish(g, x)
return sym_help.quantize_helper(g, output, op_scale, op_zero_point)
