def symbolic(g, input_, levels, input_low, input_high, output_low,
output_high):
return g.op(add_domain("FakeQuantize"),
input_,
input_low,
input_high,
output_low,
output_high,
levels_i=levels)
def symbolic(g, x, scale, threshold):
zero = g.constant(0, [1], 'float')
zero = g.op("Unsqueeze", zero, axes_i=[0, 2, 3])
threshold = g.op("Mul", threshold, scale)
scale = g.op("Unsqueeze", scale, axes_i=[0, 2, 3])
return g.op(add_domain("FakeQuantize"),
x,
threshold,
threshold,
zero,
scale,
levels_i=2)
def symbolic(g, x):
zero = g.constant(0, [1], 'float')
zero = g.op("Unsqueeze", zero, axes_i=[1, 2, 3])
scale = g.op("Abs", x)
scale = g.op("ReduceMean", scale, axes_i=[0, 1, 2, 3])
scale_neg = g.op("Neg", scale)
return g.op(add_domain("FakeQuantize"),
x,
zero,
zero,
scale_neg,
scale,
levels_i=2)
def symbolic(g, input_fm, img_tensor, priorbox_params):
return g.op(add_domain("PriorBox"),
input_fm,
img_tensor,
min_size_f=[priorbox_params.min_size],
max_size_f=[priorbox_params.max_size],
aspect_ratio_f=priorbox_params.aspect_ratio,
flip_i=priorbox_params.flip,
clip_i=priorbox_params.clip,
variance_f=priorbox_params.variance,
step_f=priorbox_params.step,
offset_f=priorbox_params.offset,
step_h_f=priorbox_params.step_h,
step_w_f=priorbox_params.step_w,
img_size_i=priorbox_params.img_size,
img_h_i=priorbox_params.img_h,
img_w_i=priorbox_params.img_w)
def symbolic(g, loc_data, conf_data, prior_data, detection_output_params):
return g.op(
add_domain("DetectionOutput"),
loc_data,
conf_data,
prior_data,
num_classes_i=detection_output_params.num_classes,
background_label_id_i=detection_output_params.background_label_id,
top_k_i=detection_output_params.top_k,
keep_top_k_i=detection_output_params.keep_top_k,
confidence_threshold_f=detection_output_params.
confidence_threshold,
nms_threshold_f=detection_output_params.nms_threshold,
eta_f=detection_output_params.eta,
share_location_i=detection_output_params.share_location,
code_type_s=detection_output_params.code_type,
variance_encoded_in_target_i=detection_output_params.
variance_encoded_in_target)
def symbolic(g, x, weight, l2NormParams):
return g.op(add_domain("Normalize"), x, weight, eps_f=l2NormParams.eps,
across_spatial_i=l2NormParams.across_spatial, channel_shared_i=l2NormParams.channel_shared)