def roi_align(g, input, rois, spatial_scale, pooled_height, pooled_width,
sampling_ratio, aligned):
batch_indices = _cast_Long(
g,
squeeze(
g,
select(
g, rois, 1,
g.op('Constant',
value_t=torch.tensor([0], dtype=torch.long))), 1),
False)
rois = select(
g, rois, 1,
g.op('Constant',
value_t=torch.tensor([1, 2, 3, 4], dtype=torch.long)))
if aligned:
warnings.warn(
"ONNX export of ROIAlign with aligned=True does not match PyTorch when using malformed boxes,"
" ONNX forces ROIs to be 1x1 or larger.")
scale = torch.tensor(0.5 / spatial_scale).to(dtype=torch.float)
rois = g.op("Sub", rois, scale)
return g.op('RoiAlign',
input,
rois,
batch_indices,
spatial_scale_f=spatial_scale,
output_height_i=pooled_height,
output_width_i=pooled_width,
sampling_ratio_i=sampling_ratio)
def roi_align(g, input, rois, spatial_scale, pooled_height, pooled_width,
sampling_ratio, aligned):
if (aligned):
raise RuntimeError(
'Unsupported: ONNX export of roi_align with aligned')
batch_indices = _cast_Long(
g,
squeeze(
g,
select(
g, rois, 1,
g.op('Constant',
value_t=torch.tensor([0], dtype=torch.long))), 1),
False)
rois = select(
g, rois, 1,
g.op('Constant',
value_t=torch.tensor([1, 2, 3, 4], dtype=torch.long)))
return g.op('RoiAlign',
input,
rois,
batch_indices,
spatial_scale_f=spatial_scale,
output_height_i=pooled_height,
output_width_i=pooled_width,
sampling_ratio_i=sampling_ratio)
def embedding_bag(g,
embedding_matrix,
indices,
offsets,
scale_grad_by_freq,
mode,
sparse,
per_sample_weights,
include_last_offset,
padding_idx):
if scale_grad_by_freq and sym_help._training_mode:
return sym_help._onnx_unsupported("embedding_bag with scale_grad_by_freq for training mode")
if padding_idx is not None and padding_idx >= 0:
raise RuntimeError("embedding_bag with padding_idx")
from torch.onnx.symbolic_opset9 import select
import warnings
warnings.warn("Export of embedding_bag with dynamic input/offsets shape is not supported in opset 10. "
"Please use opset 11 or higher to export model for dynamic input shape.'")
offsets_dim_0 = sym_help._get_tensor_dim_size(offsets, 0)
if offsets_dim_0 is not None:
if include_last_offset:
offset_len = offsets_dim_0 - 1
offsets_extended = offsets
else:
offset_len = offsets_dim_0
offsets_extended = [offsets, g.op("Constant", value_t=torch.tensor([maxsize]))]
offsets_extended = g.op("Concat", *offsets_extended, axis_i=0)
list_ = []
for i in range(offset_len):
start_ = sym_help._unsqueeze_helper(g, select(g, offsets_extended, torch.tensor(0), torch.tensor(i)), [0])
end_ = sym_help._unsqueeze_helper(g, select(g, offsets_extended, torch.tensor(0), torch.tensor(i + 1)), [0])
axes_ = g.op("Constant", value_t=torch.tensor([0]))
indices_row = g.op("Slice", indices, start_, end_, axes_)
embeddings = g.op("Gather", embedding_matrix, indices_row)
if not sym_help._is_none(per_sample_weights):
per_sample_weights_row = g.op("Slice", per_sample_weights, start_, end_, axes_)
per_sample_weights_row = sym_help._unsqueeze_helper(g, per_sample_weights_row, [1])
embeddings = g.op("Mul", embeddings, per_sample_weights_row)
if mode == 0:
embeddings = sym_help._reducesum_helper(g, embeddings, axes_i=[0], keepdims_i=0)
elif mode == 1:
embeddings = g.op("ReduceMean", embeddings, axes_i=[0], keepdims_i=0)
else:
embeddings = g.op("ReduceMax", embeddings, axes_i=[0], keepdims_i=0)
embeddings = sym_help._unsqueeze_helper(g, embeddings, [0])
list_.append(embeddings)
output = g.op("Concat", *list_, axis_i=0)
# aten::embedding_bag returns a tuple of 4 elements: output, offset2bag, bag_size, max_indices.
# But the last three outputs are not used in torch.nn.EmbeddingBag or torch.nn.functional.embedding_bag.
return output, None, None, None
else:
return sym_help._onnx_unsupported("embedding_bag with unknown shape of offsets for opset 10 is not supported. "
"please use opset 11 or higher.")
def symbolic(g, bboxes, scores, iou_threshold, offset):
from ..onnx import is_custom_op_loaded
has_custom_op = is_custom_op_loaded()
# TensorRT nms plugin is aligned with original nms in ONNXRuntime
is_trt_backend = os.environ.get('ONNX_BACKEND') == 'MMCVTensorRT'
if has_custom_op and (not is_trt_backend):
return g.op('mmcv::NonMaxSuppression',
bboxes,
scores,
iou_threshold_f=float(iou_threshold),
offset_i=int(offset))
else:
from torch.onnx.symbolic_opset9 import select, squeeze, unsqueeze
boxes = unsqueeze(g, bboxes, 0)
scores = unsqueeze(g, unsqueeze(g, scores, 0), 0)
max_output_per_class = g.op('Constant',
value_t=torch.tensor([sys.maxsize],
dtype=torch.long))
iou_threshold = g.op('Constant',
value_t=torch.tensor([iou_threshold],
dtype=torch.float))
nms_out = g.op('NonMaxSuppression', boxes, scores,
max_output_per_class, iou_threshold)
return squeeze(
g,
select(
g, nms_out, 1,
g.op('Constant',
value_t=torch.tensor([2], dtype=torch.long))), 1)
def symbolic(g, bboxes, scores, iou_threshold, offset):
from ..onnx import is_custom_op_loaded
has_custom_op = is_custom_op_loaded()
if has_custom_op:
return g.op('mmcv::NonMaxSuppression',
bboxes,
scores,
iou_threshold_f=float(iou_threshold),
offset_i=int(offset))
else:
from torch.onnx.symbolic_opset9 import select, squeeze, unsqueeze
boxes = unsqueeze(g, bboxes, 0)
scores = unsqueeze(g, unsqueeze(g, scores, 0), 0)
max_output_per_class = g.op('Constant',
value_t=torch.tensor([sys.maxsize],
dtype=torch.long))
iou_threshold = g.op('Constant',
value_t=torch.tensor([iou_threshold],
dtype=torch.float))
nms_out = g.op('NonMaxSuppression', boxes, scores,
max_output_per_class, iou_threshold)
return squeeze(
g,
select(
g, nms_out, 1,
g.op('Constant',
value_t=torch.tensor([2], dtype=torch.long))), 1)
def symbolic_multi_label_nms(g, boxes, scores, iou_threshold):
boxes = unsqueeze(g, boxes, 0)
scores = unsqueeze(g, unsqueeze(g, scores, 0), 0)
max_output_per_class = g.op('Constant', value_t=torch.tensor([sys.maxsize], dtype=torch.long))
iou_threshold = g.op('Constant', value_t=torch.tensor([iou_threshold], dtype=torch.float))
nms_out = g.op('NonMaxSuppression', boxes, scores, max_output_per_class, iou_threshold)
return squeeze(g, select(g, nms_out, 1, g.op('Constant', value_t=torch.tensor([2], dtype=torch.long))), 1)
def roi_align(g, input, rois, spatial_scale, pooled_height, pooled_width,
sampling_ratio):
batch_indices = _cast_Long(
g,
squeeze(
g,
select(
g, rois, 1,
g.op('Constant',
value_t=torch.tensor([0], dtype=torch.long))), 1),
False)
rois = select(
g, rois, 1,
g.op('Constant',
value_t=torch.tensor([1, 2, 3, 4], dtype=torch.long)))
return g.op('RoiAlign',
input,
rois,
batch_indices,
spatial_scale_f=spatial_scale,
output_height_i=pooled_height,
output_width_i=pooled_width,
sampling_ratio_i=sampling_ratio)
def symbolic(g, bboxes, scores, iou_threshold, offset):
from torch.onnx.symbolic_opset9 import select, squeeze, unsqueeze
boxes = unsqueeze(g, bboxes, 0)
scores = unsqueeze(g, unsqueeze(g, scores, 0), 0)
max_output_per_class = g.op(
'Constant', value_t=torch.tensor([sys.maxsize], dtype=torch.long))
iou_threshold = g.op(
'Constant',
value_t=torch.tensor([iou_threshold], dtype=torch.float))
nms_out = g.op('NonMaxSuppression', boxes, scores,
max_output_per_class, iou_threshold)
return squeeze(
g,
select(
g, nms_out, 1,
g.op('Constant', value_t=torch.tensor([2], dtype=torch.long))),
1)
def symbolic_nmsfilt(g, boxes, scores, iou_threshold, score_threshold, max_output_boxes):
# if should return all
if max_output_boxes <= 0:
max_output_boxes = 10000
shape = g.op("Shape", scores) # original shape
boxes = view(g, boxes, (1, -1, 4))
max_output_per_class = g.op('Constant', value_t=torch.tensor([max_output_boxes], dtype=torch.long))
iou_threshold = g.op('Constant', value_t=torch.tensor([iou_threshold], dtype=torch.float))
score_threshold = g.op('Constant', value_t=torch.tensor([score_threshold], dtype=torch.float))
# center_point_box == 1 is for our center_x, centr_y, width, height format
nms_out = g.op('NonMaxSuppression',
boxes, view(g, scores, (1, 1, -1)), max_output_per_class, iou_threshold, score_threshold,
center_point_box_i=1)
idx = view(g, select(g, nms_out, 1, g.op('Constant', value_t=torch.tensor([2], dtype=torch.long))), (-1,))
scores = view(g, scores, (-1,))
flat_shape = g.op("Shape", scores)
src = index_select(g, scores, 0, idx)
src = view(g, src, (-1,))
filt = g.op("ConstantOfShape", flat_shape)
filt = scatter(g, filt, 0, idx, src)
return view(g, filt, shape)
def symbolic_nms(g, boxes, scores, iou_threshold, max_output_boxes):
# if should return all
if max_output_boxes <= 0:
max_output_boxes = 10000
boxes = view(g, boxes, (1, -1, 4))
max_output_per_class = g.op('Constant',
value_t=torch.tensor([max_output_boxes],
dtype=torch.long))
iou_threshold = g.op('Constant',
value_t=torch.tensor([iou_threshold],
dtype=torch.float))
# center_point_box == 1 is for our center_x, centr_y, width, height format
nms_out = g.op('NonMaxSuppression',
boxes,
view(g, scores, (1, 1, -1)),
max_output_per_class,
iou_threshold,
center_point_box_i=1)
idx = select(
g, nms_out, 1,
g.op('Constant', value_t=torch.tensor([2], dtype=torch.long)))
return view(g, idx, (-1, ))
def _size_helper(g, self, dim):
full_shape = g.op("Shape", self)
from torch.onnx.symbolic_opset9 import select
return select(g, full_shape, g.op("Constant", value_t=torch.tensor([0])), dim)
def embedding_bag(g, embedding_matrix, indices, offsets, scale_grad_by_freq,
mode, sparse, per_sample_weights, include_last_offset):
if scale_grad_by_freq and sym_help._training_mode:
return sym_help._onnx_unsupported(
'embedding_bag with scale_grad_by_freq for training mode')
from torch.onnx.symbolic_opset9 import size, div, select
# Check if initial indices was 2D. In functional.py:
# offsets is set to torch.arange(0, indices.numel(), indices.size(1))
# Then indices is reshaped to 1D: indices.reshape(-1)
if len(list(indices.node().inputs())) > 0 and indices.node().inputs().__next__().type().sizes() is not None \
and len(indices.node().inputs().__next__().type().sizes()) == 2:
# Assert include_last_offset is False
assert not include_last_offset
embeddings = g.op("Gather", embedding_matrix, indices)
dim_0 = size(g, offsets, g.op("Constant",
value_t=torch.LongTensor([0])))
dim_1 = div(
g, size(g, indices, g.op("Constant",
value_t=torch.LongTensor([0]))), dim_0)
dim_2 = g.op("Constant", value_t=torch.LongTensor([-1]))
shape = [dim_0, dim_1, dim_2]
shape = g.op("Concat", *shape, axis_i=0)
if not sym_help._is_none(per_sample_weights):
per_sample_weights = g.op("Unsqueeze",
per_sample_weights,
axes_i=[1])
embeddings = g.op("Mul", embeddings, per_sample_weights)
embeddings = g.op("Reshape", embeddings, shape)
if mode == 0:
embeddings = g.op("ReduceSum",
embeddings,
axes_i=[1],
keepdims_i=0)
elif mode == 1:
embeddings = g.op("ReduceMean",
embeddings,
axes_i=[1],
keepdims_i=0)
else:
embeddings = g.op("ReduceMax",
embeddings,
axes_i=[1],
keepdims_i=0)
# aten::embedding_bag returns a tuple of 4 elements: output, offset2bag, bag_size, max_indices.
# But the last three outputs are not used in torch.nn.EmbeddingBag or torch.nn.functional.embedding_bag.
return embeddings, None, None, None
elif offsets.type().sizes() is not None:
if include_last_offset:
offset_len = offsets.type().sizes()[0] - 1
offsets_extended = offsets
else:
offset_len = offsets.type().sizes()[0]
offsets_extended = [
offsets,
g.op("Constant", value_t=torch.tensor([maxsize]))
]
offsets_extended = g.op("Concat", *offsets_extended, axis_i=0)
list_ = []
for i in range(offset_len):
start_ = g.op("Unsqueeze",
select(g, offsets_extended, torch.tensor(0),
torch.tensor(i)),
axes_i=[0])
end_ = g.op("Unsqueeze",
select(g, offsets_extended, torch.tensor(0),
torch.tensor(i + 1)),
axes_i=[0])
axes_ = g.op("Constant", value_t=torch.tensor([0]))
indices_row = g.op("Slice", indices, start_, end_, axes_)
embeddings = g.op("Gather", embedding_matrix, indices_row)
if not sym_help._is_none(per_sample_weights):
per_sample_weights_row = g.op("Slice", per_sample_weights,
start_, end_, axes_)
per_sample_weights_row = g.op("Unsqueeze",
per_sample_weights_row,
axes_i=[1])
embeddings = g.op("Mul", embeddings, per_sample_weights_row)
if mode == 0:
embeddings = g.op("ReduceSum",
embeddings,
axes_i=[0],
keepdims_i=0)
elif mode == 1:
embeddings = g.op("ReduceMean",
embeddings,
axes_i=[0],
keepdims_i=0)
else:
embeddings = g.op("ReduceMax",
embeddings,
axes_i=[0],
keepdims_i=0)
embeddings = g.op("Unsqueeze", embeddings, axes_i=[0])
list_.append(embeddings)
output = g.op("Concat", *list_, axis_i=0)
# aten::embedding_bag returns a tuple of 4 elements: output, offset2bag, bag_size, max_indices.
# But the last three outputs are not used in torch.nn.EmbeddingBag or torch.nn.functional.embedding_bag.
return output, None, None, None
else:
return sym_help._onnx_unsupported(
'embedding_bag with unknown shape of indices')