Google Git
Sign in
android / platform / external / pytorch / f3b8a470e1 / . / torch / onnx / symbolic_opset9.py
blob: 38b021650a0b8974fb37d36c439de2a12f64ad74 [file] [log] [blame]
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from __future__ import unicode_literals
import torch
from torch._C import ListType, OptionalType
from torch.nn.modules.utils import _single, _pair, _triple
import torch.onnx
# This import monkey-patches graph manipulation methods on Graph, used for the
# ONNX symbolics
import torch.onnx.utils
from functools import partial
from functools import wraps
import torch.onnx.symbolic_helper as sym_help
from torch.onnx.symbolic_helper import parse_args, _parse_arg, _unimplemented
import numpy
import math
import warnings
# EDITING THIS FILE? READ THIS FIRST!
# see Note [Edit Symbolic Files] in symbolic_helper.py
# This file exports ONNX ops for opset 9
# Opset 9 is supported by ONNX release 1.4.1
# release on 01/23/19
# Note [Pointwise by scalar]
# ~~~~~~~~~~~~~~~~~~~~~~~~~~
# What happens if you add a tensor with a constant (e.g., x + 2)? There are
# some moving parts to implementing the ONNX translation in this case:
#
# - By the time we get the scalar in a symbolic function here, it is no longer
# a Python long/float, but a PyTorch tensor with numel == 1 (eventually, we
# want it to be a zero dim tensor but this change has not happened yet.)
# However, the type of this scalar is *exactly* what the user wrote in
# Python, which may not match the tensor it is being added to. PyTorch
# will do implicit conversions on scalars; however, ONNX will not, so
# we must do the conversion ourselves. This is what _if_scalar_type_as
# does.
#
# - Dispatch to these functions takes advantage an outrageous coincidence
# between the tensor and scalar name. When we add two tensors together,
# you get the dispatch:
#
# add(*[self, other], **{"alpha": alpha})
#
# When you add a tensor and a scalar, you get the dispatch:
#
# add(*[self], **{"other": other, "alpha": alpha})
#
# By having the argument name line up with the name of the scalar attribute
# if it exists, we can write a single function for both overloads.
#
# used to represent "missing" optional inputs
def unused(g):
n = g.op("prim::Constant")
n.setType(OptionalType.ofTensor())
return n
def _shape_as_tensor(g, input):
return g.op('Shape', input)
def _reshape_from_tensor(g, input, shape):
return g.op('Reshape', input, shape)
def reshape(g, self, shape):
return view(g, self, shape)
def reshape_as(g, self, other):
shape = g.op('Shape', other)
return reshape(g, self, shape)
def add(g, self, other, alpha=None):
# default alpha arg is to allow no-alpha add (aten add st overload no alpha)
if alpha and sym_help._scalar(sym_help._maybe_get_scalar(alpha)) != 1:
return _unimplemented("add", "alpha != 1")
return g.op("Add", self, other)
def sub(g, self, other, alpha=None):
# default alpha arg is to allow no-alpha sub (aten sub st overload no alpha)
if alpha and sym_help._scalar(sym_help._maybe_get_scalar(alpha)) != 1:
return _unimplemented("sub", "alpha != 1")
return g.op("Sub", self, other)
def rsub(g, self, other, alpha=None):
return sub(g, other, self, alpha=alpha)
def mul(g, self, other):
return g.op("Mul", self, other)
def div(g, self, other):
return g.op("Div", self, other)
def floor_divide(g, self, other):
out = div(g, self, other)
# the correct operation is truncate, which is not supported in ONNX,
# we cannot call floor since it will behave differently for negative numbers
# (eg. -0.1 should become -0 )
# - if scalar_type information are not available, assume that
# we need to call floor (treat as float)
out = g.op("Cast", out, to_i=sym_help.cast_pytorch_to_onnx['Long'])
# Matching PyTorch's behavior:
# - if self is fp the output's type is self's type
# - if self is not fp and other is fp, the output is of type 'Float'
# - self is not fp and other is not fp, the output's type is self's output type
# - the output type defaults to Float
scalar_type = self.type().scalarType()
if scalar_type is not None:
if not sym_help._is_fp(self) and \
other.type().scalarType() is not None and \
sym_help._is_fp(other):
out = g.op("Cast", out, to_i=sym_help.cast_pytorch_to_onnx['Float'])
else:
out = g.op("Cast", out, to_i=sym_help.cast_pytorch_to_onnx[scalar_type])
else:
out = g.op("Cast", out, to_i=sym_help.cast_pytorch_to_onnx['Float'])
return out
# Division where both inputs are cast to floating types
# If both inputs are floating, performs div as usual
# If only one input is a floating type, the other input is cast to its type
# If neither input is a floating type, both inputs are cast to the default scalar type
def true_divide(g, self, other):
# Case 1: both values are floating
# Performs div as usual
if sym_help._is_fp(self) and sym_help._is_fp(other):
return div(g, self, other)
# Case 2: self is floating, other is not
# Casts other to self's dtype
if sym_help._is_fp(self):
g.op("Cast", other, to_i=sym_help.cast_pytorch_to_onnx[self.type().scalarType()])
return div(g, self, other)
# Case 3: other is floating, self is not
# Casts self to other's dtype
if sym_help._is_fp(other):
g.op("Cast", other, to_i=sym_help.cast_pytorch_to_onnx[other.type().scalarType()])
return div(g, self, other)
# Case 4: neither is floating
# Casts both inputs to the default scalar type
scalar_type = torch.get_default_dtype()
onnx_scalar_type = sym_help.cast_pytorch_to_onnx['Float']
assert scalar_type is torch.float or scalar_type is torch.double
if torch.get_default_dtype() is torch.double:
onnx_scalar_type = sym_help.cast_pytorch_to_onnx['Double']
g.op("Cast", self, to_i=onnx_scalar_type)
g.op("Cast", other, to_i=onnx_scalar_type)
return div(g, self, other)
def reciprocal(g, self):
return g.op("Div", torch.ones(1), self)
@parse_args('v', 'i')
def cat(g, tensor_list, dim):
tensors = sym_help._unpack_list(tensor_list)
return g.op("Concat", *tensors, axis_i=dim)
@parse_args('v', 'i')
def stack(g, tensor_list, dim):
unsqueezed = [g.op("Unsqueeze", t, axes_i=[dim]) for t in sym_help._unpack_list(tensor_list)]
return g.op("Concat", *unsqueezed, axis_i=dim)
def mm(g, self, other):
# Create a dummy C tensor. Only needed for API purposes, the value is
# since beta = 0
C = g.op("Constant", value_t=torch.tensor([1]))
return g.op("Gemm", self, other, C, beta_f=0.0, alpha_f=1.0)
def bmm(g, self, other):
return g.op("MatMul", self, other)
def matmul(g, self, other):
return g.op("MatMul", self, other)
@parse_args('v', 'v', 'v', 't', 't')
def addmm(g, self, mat1, mat2, beta, alpha):
return g.op("Gemm", mat1, mat2, self, beta_f=sym_help._scalar(beta), alpha_f=sym_help._scalar(alpha))
def neg(g, self):
return g.op("Neg", self)
def sqrt(g, self):
return g.op("Sqrt", self)
def rsqrt(g, self):
return div(g, sym_help._if_scalar_type_as(g, torch.ones(1), self), sqrt(g, self))
def tanh(g, self):
return g.op("Tanh", self)
def sin(g, self):
return g.op("Sin", self)
def cos(g, self):
return g.op("Cos", self)
def tan(g, self):
return g.op("Tan", self)
def asin(g, self):
return g.op("Asin", self)
def acos(g, self):
return g.op("Acos", self)
def atan(g, self):
return g.op("Atan", self)
def sigmoid(g, self):
return g.op("Sigmoid", self)
def sign(g, self):
return g.op("Sign", self)
def _slice(g, input, axes, starts, ends):
assert len(starts) == len(ends)
if len(starts) == 1 and starts[0] == 0 and ends[0] == 9223372036854775807:
return input
return g.op("Slice", input, axes_i=axes, starts_i=starts, ends_i=ends)
def _reduce_op_symbolic(onnx_op_name, allow_multi_dim_support=True):
def symbolic(g, self, dim=None, keepdim=None):
if dim is None:
# all-reduce path
return g.op(onnx_op_name, self, keepdims_i=0)
else:
# dim-reduce path
desc = 'is' if allow_multi_dim_support else 'i'
dim, keepdim = sym_help._get_const(dim, desc, 'dim'), sym_help._get_const(keepdim, 'i', 'keepdim')
dim_list = dim if allow_multi_dim_support else [dim]
return g.op(onnx_op_name, self, axes_i=dim_list, keepdims_i=keepdim)
return symbolic
def overload_by_arg_count(fn):
@wraps(fn)
def wrapper(g, *args):
overloads = fn(g, *args)
last_exception = None
for overload in overloads:
arg_descriptors = overload._arg_descriptors
if len(arg_descriptors) == len(args):
return overload(g, *args)
raise NotImplementedError("Unknown aten::{} signature".format(fn.__name__))
return wrapper
def _reduce_with_dtype(onnx_op, name, allow_multi_dim_support=True):
symbolic = _reduce_op_symbolic(onnx_op, allow_multi_dim_support=allow_multi_dim_support)
@overload_by_arg_count
def reduce(g, *args, **kwargs):
@parse_args('v', 'none')
def reduce_nodim(g, self, dtype):
if dtype.node().kind() != 'prim::Constant':
return _unimplemented(name, "dtype")
return symbolic(g, self)
dim_desc = 'is' if allow_multi_dim_support else 'i'
@parse_args('v', dim_desc, 'i', 'none')
def reduce_dim(g, self, dim, keepdim, dtype):
if dtype.node().kind() != 'prim::Constant':
return _unimplemented(name, "dtype")
return symbolic(g, self, dim, keepdim)
return reduce_nodim, reduce_dim
return reduce
sum = _reduce_with_dtype('ReduceSum', 'sum')
mean = _reduce_with_dtype('ReduceMean', 'mean')
prod = _reduce_with_dtype('ReduceProd', 'prod', allow_multi_dim_support=False) # torch.prod does not support multidimensional 'dim'
@parse_args('v', 'i', 'none')
def cumsum(g, input, dim, dtype):
if dtype.node().kind() != 'prim::Constant':
return _unimplemented(name, "dtype")
return g.op("ATen", input, operator_s="cumsum", dim_i=dim)
def _sample_dirichlet(g, self, generator):
if not sym_help._is_none(generator):
return _unimplemented('_sample_dirichlet',
'We are not able to export generator')
return g.op("ATen", self, operator_s="_sample_dirichlet")
def _standard_gamma(g, self, generator):
if not sym_help._is_none(generator):
return _unimplemented('_standard_gamma',
'We are not able to export generator')
return g.op("ATen", self, operator_s="_standard_gamma")
def t(g, self):
return g.op("Transpose", self, perm_i=(1, 0))
def expand(g, self, size, implicit):
size = sym_help._maybe_get_const(size, 'is')
if not sym_help._is_value(size):
size = g.op("Constant", value_t=torch.LongTensor(size))
elif sym_help._is_packed_list(size):
# Expand with -1 dim value means dim is unchanged.
# Since onnx::expand supports two-way broadcasting,
# -1 dim value can be exported to onnx as 1
size = view(g, stack(g, size, 0), [-1])
dtype = 4 # dim type is int64
ones = ones_like(g, size, dtype)
neg_ones = mul(g, ones, g.op("Constant", value_t=torch.tensor(-1)))
size = where(g, g.op("Equal", size, neg_ones), ones, size)
return g.op("Expand", self, size)
def expand_as(g, self, other):
shape = g.op("Shape", other)
return g.op("Expand", self, shape)
def embedding(g, weight, indices, padding_idx, scale_grad_by_freq, sparse):
return g.op("Gather", weight, indices)
@parse_args('v', 'v', 'v', 'i', 'i', 'i', 'v', 'i')
def embedding_bag(g,
embedding_matrix,
indices,
offsets,
scale_grad_by_freq,
mode,
sparse,
per_sample_weights,
include_last_offset):
if not sym_help._is_none(per_sample_weights):
raise RuntimeError('Unsupported: ONNX export of embedding_bag '
'with per_sample_weights')
return g.op("ATen",
embedding_matrix,
indices,
offsets,
operator_s="embedding_bag",
outputs=4,
scale_grad_by_freq_i=scale_grad_by_freq,
mode_i=mode,
sparse_i=sparse,
include_last_offset_i=include_last_offset)
def size(g, self, dim):
if sym_help._maybe_get_const(dim, 'i') < 0:
rank = self.type().dim()
if rank:
dim = sym_help._maybe_get_const(dim, 'i') + rank
dim = g.op("Constant", value_t=torch.tensor(dim))
return sym_help._size_helper(g, self, dim)
@parse_args('v', 'i', 'i')
def transpose(g, self, dim0, dim1):
if dim0 == dim1: # micro-optimization
return self
# NB: Transpose in ONNX is actually a Permute
if self.isCompleteTensor():
axes = list(range(self.type().dim()))
axes[dim0], axes[dim1] = axes[dim1], axes[dim0]
return g.op("Transpose", self, perm_i=axes)
else:
# if we don't have dim information we cannot
# output a permute so use ATen instead
return g.op("ATen", self, operator_s="transpose", dim0_i=dim0, dim1_i=dim1)
@parse_args('v', 'is')
def permute(g, self, dims):
if dims == list(range(0, len(dims))):
return self
return g.op("Transpose", self, perm_i=dims)
def view(g, self, size):
size = sym_help._maybe_get_const(size, 'is')
if sym_help._is_value(size):
shape = size
else:
if self.isCompleteTensor():
self_sizes = self.type().sizes()
if self_sizes and len(size) == 2 and self_sizes[0] == size[0]:
return g.op("Flatten", self, axis_i=1)
shape = g.op("Constant", value_t=torch.LongTensor(size))
return g.op("Reshape", self, shape)
def prim_ConstantSplit(g, self, split_size, dim):
size = self.type().sizes()[dim]
splits = [split_size] * (size // split_size)
leftover = size % split_size
if leftover:
splits.append(leftover)
return g.op("Split", self, split_i=splits, axis_i=dim, outputs=len(splits))
# TODO: It would be better to export this as a chunk directly, as this is
# less sensitive to changes in input size.
# TODO: Once we have proper scoping, stop reimplementing chunk, delete this
# method, and use the desugared version
def prim_ConstantChunk(g, self, chunks, dim):
split_size = (self.type().sizes()[dim] + chunks - 1) // chunks
return prim_ConstantSplit(g, self, split_size, dim)
def split(g, self, split_size_or_sizes, dim):
if sym_help._is_value(split_size_or_sizes) and split_size_or_sizes.node().kind() != 'onnx::Constant':
raise RuntimeError("ONNX symbolic expected a constant value of the {} argument, got `{}`"
.format('split_size_or_sizes', split_size_or_sizes))
split_val = split_size_or_sizes.node()['value']
if split_val.dim() > 0:
return split_with_sizes(g, self, split_size_or_sizes, dim)
split_size = sym_help._get_const(split_size_or_sizes, 'i', 'split_size')
dim = sym_help._get_const(dim, 'i', 'dim')
size = self.type().sizes()[dim]
splits = [split_size] * (size // split_size)
leftover = size % split_size
if leftover:
splits.append(leftover)
return g.op("Split", self, split_i=splits, axis_i=dim, outputs=1)
@parse_args('v', 'is', 'i')
def split_with_sizes(g, self, split_sizes, dim):
return g.op("Split", self, split_i=split_sizes, axis_i=dim, outputs=1)
@parse_args('v', 'i')
def unbind(g, self, dim=0):
# NOTE: This conversion of this node is handled in onnx peephole pass.
# Due to that an additional Squeeze node needs to be inserted for each output from unbind.
return g.op("aten::unbind", self, axis_i=dim)
@parse_args('v', 'i', 'v')
def select(g, self, dim, index):
index = sym_help._maybe_get_scalar(index)
if (not sym_help._is_value(index)) and (index < 0):
if index == -1:
end_index = 9223372036854775807
else:
end_index = index + 1
slice_node = sym_help._slice_helper(g, self, axes=[dim], starts=[index], ends=[end_index])
return g.op("Squeeze", slice_node, axes_i=[dim])
else:
return g.op("Gather", self, index, axis_i=dim)
def squeeze(g, self, dim=None):
if dim is None:
dims = []
for i, size in enumerate(self.type().sizes()):
if size == 1:
dims.append(i)
else:
dims = [sym_help._get_const(dim, 'i', 'dim')]
# Handle negative dims
for i, dim in enumerate(dims):
if dim < 0:
rank = self.type().dim()
if rank:
warnings.warn("ONNX export squeeze with negative axis " + str(dim) +
" might cause the onnx model to be incorrect. " +
"Negative axis is not supported in ONNX. " +
"Axis is converted to " + str(dim + rank) +
" based on input shape at export time. " +
"Passing an tensor of different rank in execution will be incorrect.")
dims[i] += rank
else:
return _unimplemented('squeeze', 'negative axis with unknown input rank')
return g.op("Squeeze", self, axes_i=dims)
def prelu(g, self, weight):
if self.isCompleteTensor():
self_sizes = self.type().sizes()
if self_sizes and len(self_sizes) > 2:
weight = g.op("Unsqueeze", weight, axes_i=list(range(1, len(self_sizes) - 1)))
return g.op("PRelu", self, weight)
def relu(g, input):
return g.op("Relu", input)
def ceil(g, input):
return g.op("Ceil", input)
def floor(g, input):
return g.op("Floor", input)
@parse_args('v', 't', 't')
def threshold(g, self, threshold, value):
# See Note [Export inplace]
if sym_help._scalar(threshold) != 0:
return _unimplemented("threshold", "non-zero threshold")
if sym_help._scalar(value) != 0:
return _unimplemented("threshold", "non-zero value")
return g.op("Relu", self)
def leaky_relu(g, input, negative_slope, inplace=False):
negative_slope = sym_help._get_const(negative_slope, 't', 'negative_slope')
# See Note [Export inplace]
# TODO: Talk to ONNX about unconditional cast of scalar to float
return g.op("LeakyRelu", input, alpha_f=sym_help._scalar(negative_slope))
@parse_args('v', 'i')
def glu(g, input, dim):
assert input.type().sizes()[dim] % 2 == 0
first, second = g.op('Split', input, axis_i=dim, outputs=2)
return g.op('Mul', first, g.op('Sigmoid', second))
@parse_args('v', 'i', 'none')
def softmax(g, input, dim, dtype=None):
# Softmax does normalization at vector level.
# PyTorch and ONNX use different strategies to split the input tensor into vectors.
# Thus dim and axis have different meanings.
# PyTorch slices the input tensor into vectors along the `dim`-th dimension.
# ONNX reshapes the input into a 2-D tensor, and `axis` indicates where the input is coerced.
# If input is a 2 x 3 tensor:
# input = [[1.0, 1.0, 1.0],
# [1.0, 1,0, 1,0]]
# with dim = 0, the result is:
# result = [[0.5, 0.5, 0.5],
# [0.5, 0.5, 0.5]]
# with axis = 0, the result is:
# result = [[0.167, 0.167, 0.167],
# [0.167, 0.167, 0.167]]
# So only when dim and axis both equal to ndim - 1 (the last dimension),
# their semantics are equivalent.
# So use softmax when dim and axis both equal to ndim - 1
# otherwise compute softmax using a subgraph with other operators
input_dim = input.type().dim()
if input_dim:
# TODO: remove this as onnx opset 11 spec allows negative axes
if dim < 0:
dim = input_dim + dim
if input_dim == dim + 1:
softmax = g.op('Softmax', input, axis_i=dim)
if dtype and dtype.node().kind() != 'prim::Constant':
parsed_dtype = sym_help._get_const(dtype, 'i', 'dtype')
softmax = g.op("Cast", softmax, to_i=sym_help.scalar_type_to_onnx[parsed_dtype])
return softmax
exp = g.op('Exp', input)
sum = g.op('ReduceSum', exp, axes_i=[dim])
softmax = g.op('Div', exp, sum)
if dtype and dtype.node().kind() != 'prim::Constant':
parsed_dtype = sym_help._get_const(dtype, 'i', 'dtype')
softmax = g.op("Cast", softmax, to_i=sym_help.scalar_type_to_onnx[parsed_dtype])
return softmax
@parse_args('v', 't', 'v')
def softplus(g, self, beta, threshold):
if beta != 1:
return _unimplemented("beta", "has to be 1")
return g.op('Softplus', self)
def get_pool_ceil_padding(input, kernel_size, stride, padding):
dim = input.type().sizes()[-len(padding):]
ceiled_output_dim = [int(math.ceil((dim[i] + 2 * padding[i] - kernel_size[i]) / float(stride[i]))) + 1
for i in range(0, len(padding))]
# ensure last pooling starts inside
ceiled_output_dim = [ceiled_output_dim[i] - 1
if (((ceiled_output_dim[i] - 1) * stride[i]) >= (dim[i] + padding[i]))
else ceiled_output_dim[i]
for i in range(0, len(ceiled_output_dim))]
padding_ceil = [0
if (stride[i] == 1)
else
(kernel_size[i] - (dim[i] + 2 * padding[i] - ((ceiled_output_dim[i] - 1) * stride[i] + 1)))
for i in range(0, len(padding))]
# ensure padding is not > kernel_size
padding_ceil = [(int(padding_ceil[i]) if padding_ceil[i] < kernel_size[i] - 1 else int(kernel_size[i] - 1))
if ((padding_ceil[i] + 2 * padding[i]) >= (kernel_size[i]))
else
int(padding_ceil[i])
for i in range(0, len(padding_ceil))]
return padding_ceil
def _max_pool(name, tuple_fn, ndims, return_indices):
@parse_args('v', 'is', 'is', 'is', 'is', 'i')
def symbolic_fn(g, input, kernel_size, stride, padding, dilation, ceil_mode):
if ceil_mode and not input.isCompleteTensor():
return _unimplemented(name, "input size not accessible")
if set(tuple_fn(dilation)) != {1}:
return _unimplemented(name, "dilation")
if not stride:
stride = kernel_size
padding = tuple(tuple_fn(padding))
if ceil_mode:
padding_ceil = get_pool_ceil_padding(input, kernel_size, stride, padding)
padding = padding + tuple(numpy.add(padding_ceil, padding))
else:
padding = padding * 2
kwargs = {
'kernel_shape_i': tuple_fn(kernel_size),
'pads_i': padding,
'strides_i': tuple_fn(stride),
}
# easy but hacky way to get flattened indices values
# to be used to convert the indices values to non-flattened.
# In ONNX the indices are computed as a flatten 1-D tensor,
# so the values in indices are in [0, N x C x D1 x ... x Dn).
# To convert the indices to the same format used by Pytorch,
# we first execute a maxpool with a kernel and stride of 1 on the same input.
# This will result in a tensor of indices in which each index will have it's own value.
# Using this tensor as a reference, we extract the first index of each axis and substract
# it from each index of this axis in the indices to convert.
# This step will result in a tensor were each dimension has values of indices within
# the dimension it is in.
# For more information :
# https://github.com/pytorch/pytorch/pull/16455#issuecomment-460776407
if return_indices:
r, indices = g.op("MaxPool", input, outputs=2, **kwargs)
_, flattened_indices = g.op("MaxPool", input, outputs=2,
kernel_shape_i=[1 for _ in range(ndims)],
strides_i=[1 for _ in range(ndims)])
# convert indices to have non-flattened indices values
s = sym_help._slice_helper(g, flattened_indices, axes=[2 + i for i in range(ndims)],
starts=tuple_fn(0), ends=tuple_fn(1))
indices = sub(g, indices, s)
return r, indices
else:
r = g.op("MaxPool", input, outputs=1, **kwargs)
return r
return symbolic_fn
max_pool1d = _max_pool("max_pool1d", _single, 1, return_indices=False)
max_pool2d = _max_pool("max_pool2d", _pair, 2, return_indices=False)
max_pool3d = _max_pool("max_pool3d", _triple, 3, return_indices=False)
max_pool1d_with_indices = _max_pool("max_pool1d_with_indices", _single, 1, return_indices=True)
max_pool2d_with_indices = _max_pool("max_pool2d_with_indices", _pair, 2, return_indices=True)
max_pool3d_with_indices = _max_pool("max_pool3d_with_indices", _triple, 3, return_indices=True)
def _avg_pool(name, tuple_fn):
@parse_args('v', 'is', 'is', 'is', 'i', 'i', 'none')
def symbolic_fn(g, input, kernel_size, stride, padding, ceil_mode, count_include_pad, divisor_override=None):
if ceil_mode and not input.isCompleteTensor():
return _unimplemented(name, "input size not accessible")
if not stride:
stride = kernel_size
padding = sym_help._avgpool_helper(tuple_fn, padding, kernel_size, stride, divisor_override, name)
if ceil_mode:
padding_ceil = get_pool_ceil_padding(input, kernel_size, stride, padding)
if count_include_pad:
input = g.op("Pad", input,
pads_i=((0,) * 2 + padding) * 2,
mode_s='constant',
value_f=0.)
padding = (0,) * len(padding)
if ceil_mode:
padding = padding + tuple(numpy.add(padding_ceil, padding))
else:
padding = padding * 2
output = g.op("AveragePool", input,
kernel_shape_i=tuple_fn(kernel_size),
strides_i=tuple_fn(stride),
pads_i=padding)
return output
return symbolic_fn
avg_pool1d = _avg_pool('avg_pool1d', _single)
avg_pool2d = _avg_pool('avg_pool2d', _pair)
avg_pool3d = _avg_pool('avg_pool3d', _triple)
def _adaptive_pool(name, type, tuple_fn, fn=None):
@parse_args('v', 'is')
def symbolic_fn(g, input, output_size):
# _adaptive_pool is supported for cases where output_size is 1 for all dimensions,
# by executing a GlobalPool.
# It is also supported for cases where the output size is a factor of the input size.
# For these cases the stride and kernel size are uniform along all the indices of
# the same dimension, which makes it possible to export it to ONNX.
# for MaxPool, GlobalMaxPool does not return indices,
# so we try using max_poolxd_with_indices, and if it is not possible
# (input is not a complete tensor or output size not factor of input size)
# then we call GlobalAveragePool and return None for the indices
if output_size == [1] * len(output_size) and type == "AveragePool":
return g.op("GlobalAveragePool", input)
if not input.isCompleteTensor():
if output_size == [1] * len(output_size):
return g.op("GlobalMaxPool", input), None
return _unimplemented(name, 'input size not accessible')
dim = input.type().sizes()[2:]
# verify if output size % input size = 0 for all dim
mod = [dim[i] % output_size[i] for i in range(0, len(dim))]
if mod != [0] * len(mod):
if output_size == [1] * len(output_size):
return g.op("GlobalMaxPool", input), None
return _unimplemented(name, 'output size that are not factor of input size')
k = [int(dim[i] / output_size[i]) for i in range(0, len(dim))]
# call max_poolxd_with_indices to get indices in the output
if type == "MaxPool":
return fn(g, input, k, k, (0,) * len(dim), (1,) * len(dim), False)
output = g.op(type, input,
kernel_shape_i=tuple_fn(k),
strides_i=tuple_fn(k))
return output
return symbolic_fn
adaptive_avg_pool1d = _adaptive_pool('adaptive_avg_pool1d', "AveragePool", _single)
adaptive_avg_pool2d = _adaptive_pool('adaptive_avg_pool2d', "AveragePool", _pair)
adaptive_avg_pool3d = _adaptive_pool('adaptive_avg_pool3d', "AveragePool", _triple)
adaptive_max_pool1d = _adaptive_pool('adaptive_max_pool1d', "MaxPool", _single, max_pool1d_with_indices)
adaptive_max_pool2d = _adaptive_pool('adaptive_max_pool2d', "MaxPool", _pair, max_pool2d_with_indices)
adaptive_max_pool3d = _adaptive_pool('adaptive_max_pool3d', "MaxPool", _triple, max_pool3d_with_indices)
# Generate paddings in ONNX order based on pad in pytorch.
# Arguments:
# dim: the dimension of the tensor.
# pad: the paddings in pytorch.
# The order is dim_n_begin, dim_n_end, dim_n-1_begin, dim_n-1_end, ...
def _prepare_onnx_paddings(dim, pad):
assert isinstance(dim, int)
# The desired order of paddings is
# dim_0_begin, dim_1_begin, ... , dim_0_end, ..., dim_n_end.
# n is the dimension of input.
# assume zero-dimensions in the beginning
paddings = list(pad[:]) + [0] * (dim * 2 - len(pad))
# reverse order and collate first beginnings and then ends
paddings = paddings[-2::-2] + paddings[-1::-2]
return paddings
@parse_args('v', 'is', 'f')
def constant_pad_nd(g, input, padding, value):
mode = "constant"
paddings = _prepare_onnx_paddings(input.type().dim(), padding)
return g.op("Pad", input, pads_i=paddings, mode_s=mode, value_f=value)
@parse_args('v', 'is')
def reflection_pad(g, input, padding):
mode = "reflect"
paddings = _prepare_onnx_paddings(input.type().dim(), padding)
return g.op("Pad", input, pads_i=paddings, mode_s=mode)
@parse_args('v', 'is')
def replication_pad(g, input, padding):
mode = "edge"
paddings = _prepare_onnx_paddings(input.type().dim(), padding)
return g.op("Pad", input, pads_i=paddings, mode_s=mode)
reflection_pad1d = reflection_pad
reflection_pad2d = reflection_pad
reflection_pad3d = reflection_pad
replication_pad1d = replication_pad
replication_pad2d = replication_pad
replication_pad3d = replication_pad
def _interpolate(name, dim, interpolate_mode):
def symbolic_fn(g, input, output_size, *args):
scales, align_corners = sym_help._get_interpolate_attributes(g, interpolate_mode, args)
sym_help._interpolate_warning(interpolate_mode)
align_corners = sym_help._maybe_get_scalar(align_corners)
if align_corners:
return _unimplemented(name, "align_corners == True")
if scales is None:
scales = sym_help._interpolate_size_to_scales(g, input, output_size, dim)
return g.op("Upsample", input, scales, mode_s=interpolate_mode)
return symbolic_fn
upsample_nearest1d = _interpolate('upsample_nearest1d', 3, "nearest")
upsample_nearest2d = _interpolate('upsample_nearest2d', 4, "nearest")
upsample_nearest3d = _interpolate('upsample_nearest3d', 5, "nearest")
upsample_linear1d = _interpolate('upsample_linear1d', 3, "linear")
upsample_bilinear2d = _interpolate('upsample_bilinear2d', 4, "linear")
upsample_trilinear3d = _interpolate('upsample_trilinear3d', 5, "linear")
def __interpolate(g, input, size, scale_factor, mode , align_corners, recompute_scale_factor):
scales, mode = sym_help._interpolate_get_scales_and_mode(g, input, size, scale_factor,
mode , align_corners)
return g.op("Upsample", input, scales, mode_s=mode)
@parse_args('v')
def bitwise_not(g, inp):
if inp.type().scalarType() != 'Bool':
return _unimplemented("bitwise_not", "non-bool tensor")
return g.op("Not", inp)
def wrap_logical_op_with_cast_to(to_type):
def decorator(fn):
def wrap_with_cast(g, input, other):
return g.op("Cast", fn(g, input, other), to_i=sym_help.cast_pytorch_to_onnx[to_type])
return wrap_with_cast
return decorator
def wrap_logical_op_with_cast_to_and_from(to_type):
def decorator(fn):
def wrap_with_cast(g, input, other):
to_cast_func = globals()['_cast_{}'.format(to_type)]
from_cast_func = wrap_logical_op_with_cast_to(input.type().scalarType())(fn)
return from_cast_func(g, to_cast_func(g, input, False), to_cast_func(g, other, False))
return wrap_with_cast
return decorator
def wrap_logical_op_with_negation(func):
def wrap_with_not(g, input, other):
return g.op("Not", func(g, input, other))
return wrap_with_not
def eq(g, self, other):
return g.op("Equal", self, other)
@wrap_logical_op_with_negation
def ne(g, self, other):
return g.op("Equal", self, other)
def gt(g, input, other):
return gt_impl(g, input, other)
def gt_impl(g, input, other):
return g.op("Greater", input, other)
def lt(g, input, other):
return lt_impl(g, input, other)
def lt_impl(g, input, other):
return g.op("Less", input, other)
@wrap_logical_op_with_negation
def ge(g, input, other):
return lt_impl(g, input, other)
@wrap_logical_op_with_negation
def le(g, input, other):
return gt_impl(g, input, other)
@wrap_logical_op_with_cast_to_and_from('Bool')
def __and_(g, input, other):
return g.op('And', input, other)
@wrap_logical_op_with_cast_to_and_from('Bool')
def __or_(g, input, other):
return g.op('Or', input, other)
def __rshift_(g, self, other):
# make sure to cast other to self's type
# (when self is long, make sure that other is not float)
if other.type().scalarType() != self.type().scalarType():
other = g.op("Cast", other, to_i=sym_help.cast_pytorch_to_onnx[self.type().scalarType()])
two = g.op('Constant', value_t=torch.tensor(2, dtype=torch.float32))
# exponent (same type as self) has to be float or double in onnx::Pow
if not sym_help._is_fp(self):
other = g.op("Cast", other, to_i=sym_help.cast_pytorch_to_onnx['Float'])
two_pow = g.op('Pow', two, other)
rshift = g.op('Div', self, two_pow)
return rshift
def __lshift_(g, self, other):
# make sure to cast other to self's type
# (when self is long, make sure that other is not float)
if other.type().scalarType() != self.type().scalarType():
other = g.op("Cast", other, to_i=sym_help.cast_pytorch_to_onnx[self.type().scalarType()])
two = g.op('Constant', value_t=torch.tensor(2, dtype=torch.float32))
# exponent (same type as self) has to be float or double in onnx::Pow
if not sym_help._is_fp(self):
other = g.op("Cast", other, to_i=sym_help.cast_pytorch_to_onnx['Float'])
two_pow = g.op('Pow', two, other)
lshift = g.op('Mul', self, two_pow)
return lshift
def where(g, condition, self, other):
return g.op("Where", condition, self, other)
@parse_args('v', 'i', 'none')
def log_softmax(g, input, dim, dtype=None):
# PyTorch dim and ONNX axis have different meanings.
# See Softmax comment for details.
# TODO: remove this as onnx opset 11 spec allows negative axes
input_dim = input.type().dim()
if input_dim is None:
return _unimplemented("dim",
"ONNX and PyTorch use different strategies to split the input. "
"Input rank must be known at export time.")
if dim < 0:
dim = input_dim + dim
is_transpose_required = (input_dim != dim + 1)
# ONNX only supports log_softmax with dim = -1. Transpose must be added before and after log_softmax to support other cases.
if is_transpose_required:
axes = list(range(input_dim))
axes[dim], axes[-1] = axes[-1], axes[dim]
input = g.op("Transpose", input, perm_i=axes)
dim = input_dim - 1
return_op = g.op("LogSoftmax", input, axis_i=dim)
if dtype and dtype.node().kind() != 'prim::Constant':
return_op = g.op("Cast", return_op, to_i=sym_help.scalar_type_to_onnx[dtype])
if is_transpose_required:
return_op = g.op("Transpose", return_op, perm_i=axes)
return return_op
@parse_args('v', 'v', 'v', 'is', 'is', 'is', 'i', 'is', 'i', 'i', 'i', 'i')
def _convolution(g, input, weight, bias, stride, padding, dilation,
transposed, output_padding, groups, benchmark, deterministic, cudnn_enabled):
weight_size = weight.type().sizes()
args = [input, weight]
# ONNX only supports 1D bias
if not sym_help._is_none(bias) and bias.type().dim() == 1:
args.append(bias)
kwargs = {"kernel_shape_i": weight_size[2:],
"strides_i": stride,
# NB: ONNX supports asymmetric padding, whereas PyTorch supports only
# symmetric padding
"pads_i": padding + padding,
"dilations_i": dilation,
"group_i": groups}
if any(o != 0 for o in output_padding):
# ONNX supports both output_shape and output_padding. they are equivalent expressive.
# output_padding is more straightforward, so we use it here.
# output_shape = stride * (input_shape - 1) + output_padding + kernel_shape - padding * 2
assert transposed
assert len(stride) == len(output_padding)
kwargs["output_padding_i"] = output_padding
n = g.op("ConvTranspose" if transposed else "Conv", *args, **kwargs)
if not sym_help._is_none(bias) and bias.type().dim() != 1:
return g.op("Add", n, bias)
else:
return n
@parse_args('v', 'v', 'v', 'is', 'is', 'is', 'i')
def conv1d(g, input, weight, bias, stride, padding, dilation, groups):
return _convolution(g, input, weight, bias, stride, padding, dilation, False, (), groups, None, None, None)
@parse_args('v', 'v', 'v', 'is', 'is', 'is', 'i')
def conv2d(g, input, weight, bias, stride, padding, dilation, groups):
return _convolution(g, input, weight, bias, stride, padding, dilation, False, (), groups, None, None, None)
@parse_args('v', 'v', 'v', 'is', 'is', 'is', 'i')
def conv3d(g, input, weight, bias, stride, padding, dilation, groups):
return _convolution(g, input, weight, bias, stride, padding, dilation, False, (), groups, None, None, None)
@parse_args('v', 'v', 'v', 'is', 'is', 'is', 'i', 'is')
def conv_transpose1d(g, input, weight, bias, stride, padding, output_padding, groups, dilation):
return _convolution(g, input, weight, bias, stride, padding, dilation, True, output_padding, groups, None, None, None)
@parse_args('v', 'v', 'v', 'is', 'is', 'is', 'i', 'is')
def conv_transpose2d(g, input, weight, bias, stride, padding, output_padding, groups, dilation):
return _convolution(g, input, weight, bias, stride, padding, dilation, True, output_padding, groups, None, None, None)
@parse_args('v', 'v', 'v', 'is', 'is', 'is', 'i', 'is')
def conv_transpose3d(g, input, weight, bias, stride, padding, output_padding, groups, dilation):
return _convolution(g, input, weight, bias, stride, padding, dilation, True, output_padding, groups, None, None, None)
@parse_args('v', 'v', 'v', 'v', 'v', 'i', 'f', 'f', 'i')
def batch_norm(g, input, weight, bias, running_mean, running_var, training, momentum, eps, cudnn_enabled):
input_sizes = input.type().sizes()
if weight is None or sym_help._is_none(weight):
assert len(input_sizes) > 1
weight_value = torch.tensor([1.] * input_sizes[1]).type(
'torch.' + input.type().scalarType() + 'Tensor')
weight = g.op("Constant", value_t=weight_value)
if bias is None or sym_help._is_none(bias):
assert len(input_sizes) > 1
bias_value = torch.tensor([0.] * input_sizes[1]).type(
'torch.' + input.type().scalarType() + 'Tensor')
bias = g.op("Constant", value_t=bias_value)
out = g.op("BatchNormalization", input, weight, bias, running_mean, running_var,
epsilon_f=eps,
momentum_f=1 - momentum,
outputs=1 if not training else 5)
if not training:
return out
else:
res, new_running_mean, new_running_var, saved_mean, saved_var = out
new_running_mean.setType(running_mean.type())
new_running_var.setType(running_var.type())
saved_mean.setDebugName("batch_norm_dead_output-" + saved_mean.debugName())
saved_var.setDebugName("batch_norm_dead_output-" + saved_var.debugName())
return res
@parse_args('v', 'is', 'v', 'v', 'f', 'i')
def layer_norm(g, input, normalized_shape, weight, bias, eps, cudnn_enable):
if sym_help._operator_export_type == torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK:
return g.op("ATen", input, weight, bias, normalized_shape_i=normalized_shape,
eps_f=eps, cudnn_enable_i=cudnn_enable, operator_s="layer_norm")
axes = [-i for i in range(len(normalized_shape), 0, -1)]
two_cst = g.op("Constant", value_t=torch.tensor(2.))
eps_cst = g.op("Constant", value_t=torch.tensor(eps))
mean = g.op("ReduceMean", input, axes_i=axes)
numerator = sub(g, input, mean)
# variance = e((x - e(x))^2), and (x - e(x)) is the numerator in the layer_norm formula
variance = g.op("ReduceMean", pow(g, numerator, two_cst), axes_i=axes)
denominator = sqrt(g, add(g, variance, eps_cst))
layer_norm = div(g, numerator, denominator)
if not (weight is None or sym_help._is_none(weight)):
layer_norm = mul(g, layer_norm, weight)
if not (bias is None or sym_help._is_none(bias)):
layer_norm = add(g, layer_norm, bias)
return layer_norm
@parse_args('v', 'v', 'v', 'v', 'v', 'i', 'f', 'f', 'i')
def instance_norm(g, input, weight, bias, running_mean, running_var, use_input_stats, momentum, eps, cudnn_enabled):
input_sizes = input.type().sizes()
if weight is None or sym_help._is_none(weight):
assert len(input_sizes) > 1
weight_value = torch.tensor([1.] * input_sizes[1]).type(
'torch.' + input.type().scalarType() + 'Tensor')
weight = g.op("Constant", value_t=weight_value)
if bias is None or sym_help._is_none(bias):
assert len(input_sizes) > 1
bias_value = torch.tensor([0.] * input_sizes[1]).type(
'torch.' + input.type().scalarType() + 'Tensor')
bias = g.op("Constant", value_t=bias_value)
return g.op("InstanceNormalization", input, weight, bias, epsilon_f=eps)
@parse_args('v', 'i', 'i', 'i')
def unfold(g, input, dimension, size, step):
if sym_help._operator_export_type == torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK:
return g.op("ATen", input, operator_s="unfold", dimension_i=dimension, size_i=size, step_i=step)
if input.isCompleteTensor():
sizedim = input.type().sizes()[dimension]
low_indices = range(0, sizedim, step)
hi_indices = range(size, sizedim + 1, step)
stack = [sym_help._slice_helper(g, input, axes=[dimension], starts=[low], ends=[hi])
for low, hi in zip(low_indices, hi_indices)]
ndim = input.type().dim()
perm = list(range(0, ndim))
perm.append(perm.pop(dimension))
unsqueeze = [g.op("Unsqueeze", g.op("Transpose", t, perm_i=perm), axes_i=[dimension]) for t in stack]
return g.op("Concat", *unsqueeze, axis_i=dimension)
else:
return _unimplemented("Unfold", "input size not accessible")
@parse_args('v', 't', 't', 't')
def elu(g, input, alpha, scale, input_scale):
if scale and scale != 1.:
return _unimplemented("scale", "does not support scale in Elu")
if input_scale and input_scale != 1.:
return _unimplemented("input_scale", "does not support input_scale in Elu")
# See Note [Export inplace]
return g.op("Elu", input, alpha_f=sym_help._scalar(alpha))
def selu(g, input):
return g.op("Selu", input)
@parse_args('v', 'i', 'v')
def index_select(g, self, dim, index):
# In case of a scalar index, index_select returns a tensor with the same rank as the input.
# To match this behavior in ONNX, we make index a 1D tensor so that the following gather
# also produces a tensor with the same rank as the input.
index_const = sym_help._maybe_get_scalar(index)
index_dim = index.type().dim()
if not sym_help._is_value(index_const):
# Index is a constant scalar. Make it a size 1 constant tensor.
index = g.op("Constant", value_t=torch.LongTensor([index_const]))
elif index_dim is not None:
if index_dim == 0:
# Index is a scalar. Reshape it to a size 1 tensor.
index = g.op("Reshape", index, g.op("Constant", value_t=torch.LongTensor([1])))
return g.op("Gather", self, index, axis_i=dim)
def index_put(g, self, indices_list_value, values, accumulate):
indices_list = sym_help._unpack_list(indices_list_value)
args = [self] + indices_list + [values, accumulate]
return g.op("ATen", *args, operator_s='index_put')
def index_fill(g, self, dim, index, value):
dim_value = sym_help._parse_arg(dim, 'i')
if sym_help._operator_export_type == torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK:
return g.op("ATen", self, index, value, dim_i=dim_value, operator_s="index_fill")
expanded_index_shape, expanded_index = sym_help._index_fill_reshape_helper(g, self, dim, index)
value = sym_help._maybe_get_scalar(value)
value = sym_help._if_scalar_type_as(g, value, self)
expanded_value = expand(g, value, expanded_index_shape, None)
return scatter(g, self, dim, expanded_index, expanded_value)
def index_copy(g, self, dim, index, source):
dim_value = sym_help._parse_arg(dim, 'i')
if sym_help._operator_export_type == torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK:
return g.op("ATen", self, index, source, dim_i=dim_value, operator_s="index_copy")
expanded_index_shape, expanded_index = sym_help._index_fill_reshape_helper(g, self, dim, index)
return scatter(g, self, dim, expanded_index, source)
def type_as(g, self, other):
if self.isCompleteTensor() and other.isCompleteTensor() and self.type().scalarType() == other.type().scalarType():
return self
if other.isCompleteTensor():
other_type_name = other.type().scalarType()
return g.op("Cast", self, to_i=sym_help.cast_pytorch_to_onnx[other_type_name])
else:
# We don't know the type of other, bail by emitting ATen
return g.op("ATen", self, other, operator_s="type_as")
@parse_args('v', 'v', 'i', 'f')
def cosine_similarity(g, x1, x2, dim, eps):
return g.op("ATen", x1, x2, dim_i=dim, eps_f=eps, operator_s="cosine_similarity")
# ignore clone operators that are inserted by PyTorch autograd
def clone(g, input, unused_memory_format):
return input
def abs(g, self):
return g.op("Abs", self)
def log(g, self):
return g.op("Log", self)
def log1p(g, self):
return log(g, add(g, sym_help._if_scalar_type_as(g, torch.ones(1), self), self))
def pow(g, self, exponent):
return g.op("Pow", self, exponent)
def clamp(g, self, min, max):
# min or max may be None that we need to dispatch to
# Clip separately, as ONNX does not have None syntax
if sym_help._is_none(min):
return clamp_max(g, self, max)
elif sym_help._is_none(max):
return clamp_min(g, self, min)
else:
min = _parse_arg(min, 'f')
max = _parse_arg(max, 'f')
return g.op("Clip", self, min_f=min, max_f=max)
@parse_args('v', 'f')
def clamp_min(g, self, min):
return g.op("Clip", self, min_f=min)
@parse_args('v', 'f')
def clamp_max(g, self, max):
return g.op("Clip", self, max_f=max)
# torch.max (same for torch.min) actually has two interfaces smashed together:
# torch.max(x, dim, keepdim) and torch.max(x, y)
def max(g, self, dim_or_y=None, keepdim=None):
# torch.max(input)
if dim_or_y is None and keepdim is None:
return g.op("ReduceMax", self, keepdims_i=0)
# torch.max(input, other)
if keepdim is None:
return g.op("Max", self, dim_or_y)
# torch.max(input, dim, keepdim)
else:
dim = sym_help._get_const(dim_or_y, 'i', 'dim')
keepdim = sym_help._get_const(keepdim, 'i', 'keepdim')
max = g.op("ReduceMax", self, axes_i=[dim], keepdims_i=keepdim)
indices = g.op('ArgMax', self, axis_i=dim, keepdims_i=keepdim)
return max, indices
def min(g, self, dim_or_y=None, keepdim=None):
# torch.min(input)
if dim_or_y is None and keepdim is None:
return g.op("ReduceMin", self, keepdims_i=0)
# torch.min(input, other)
if keepdim is None:
return g.op("Min", self, dim_or_y)
# torch.min(input, dim, keepdim)
else:
dim = sym_help._get_const(dim_or_y, 'i', 'dim')
keepdim = sym_help._get_const(keepdim, 'i', 'keepdim')
min = g.op("ReduceMin", self, axes_i=[dim], keepdims_i=keepdim)
indices = g.op('ArgMin', self, axis_i=dim, keepdims_i=keepdim)
return min, indices
def exp(g, self):
return g.op("Exp", self)
@parse_args('v', 'f', 'i')
def dropout(g, input, p, train):
if not train: # in eval mode, dropout is non-op
return input
warnings.warn("Dropout is a training op and should not be exported in inference mode. "
"Make sure to call eval() on the model, and to export it with param training=False.")
r, _ = g.op("Dropout", input, ratio_f=p, outputs=2)
return r
def _unsupported_dropout(name):
@parse_args('v', 'f', 'i')
def feature_dropout(g, input, p, train):
# NB: In inference mode, FeatureDropout is exported as an identity op.
if train:
return _unimplemented(name, "training mode")
return input
return feature_dropout
feature_dropout = _unsupported_dropout("feature_dropout")
alpha_dropout = _unsupported_dropout("alpha_dropout")
feature_alpha_dropout = _unsupported_dropout("feature_alpha_dropout")
# See Note [Export inplace]
dropout_ = dropout
feature_dropout_ = feature_dropout
alpha_dropout_ = alpha_dropout
feature_alpha_dropout_ = feature_alpha_dropout
@parse_args('v', 't', 'is', 'i')
def norm(g, self, p, dim, keepdim):
if p == 1:
f = _reduce_op_symbolic("ReduceL1")
elif p == 2:
f = _reduce_op_symbolic("ReduceL2")
else:
raise RuntimeError("ONNX export only p-norms with p of 1 or 2")
return f(g, self, dim=dim, keepdim=keepdim)
@parse_args('v', 'v', 'v', 'i')
def conv_tbc(g, input, weight, bias, pad):
return g.op("ATen", input, weight, bias, operator_s="conv_tbc", pad_i=pad)
@parse_args('v', 'i', 'i')
def _unique(g, input, sorted, return_inverse):
return g.op("ATen", input, operator_s="_unique", sorted_i=sorted,
return_inverse_i=return_inverse, outputs=2)
@parse_args('v', 'i', 'i', 'i')
def _unique2(g, input, sorted, return_inverse, return_counts):
return g.op("ATen", input, operator_s="_unique2", sorted_i=sorted,
return_inverse_i=return_inverse, return_counts_i=return_counts,
outputs=3)
for k, v in sym_help.cast_pytorch_to_onnx.items():
name = '_cast_{}'.format(k)
globals()[name] = parse_args('v', 'i')(partial(sym_help._cast_func_template, v))
@parse_args('v', 'i', 'v', 'v', 'v', 'v')
def empty(g, sizes, dtype, layout, device, pin_memory=False, memory_format=None):
return zeros(g, sizes, dtype, layout, device, pin_memory)
@parse_args('v', 'i', 'v', 'v', 'v', 'v')
def empty_like(g, input, dtype=None, layout=None, device=None, pin_memory=False, memory_format=None):
return zeros_like(g, input, dtype, layout, device, pin_memory)
def scalar_tensor(g, scalar, dtype, *options):
dtype = sym_help._get_const(dtype, 'i', 'dtype')
if dtype is None:
dtype = 6 # float
scalar = g.op("Cast", scalar, to_i=sym_help.scalar_type_to_onnx[dtype])
return scalar
@parse_args('v', 'i', 'v', 'v', 'v')
def zeros(g, sizes, dtype, layout, device, pin_memory=False):
# NOTE: no way to set device, layout and pin_memory in ONNX, so we ignore it
if dtype is None:
dtype = 6 # float
return g.op("ConstantOfShape", sizes,
value_t=torch.tensor([0], dtype=sym_help.scalar_type_to_pytorch_type[dtype]))
@parse_args('v', 'i', 'v', 'v', 'v', 'v')
def zeros_like(g, input, dtype=None, layout=None, device=None, pin_memory=False, memory_format=None):
shape = g.op("Shape", input)
if dtype is None:
dtype = 6 # float
return g.op("ConstantOfShape", shape,
value_t=torch.tensor([0], dtype=sym_help.scalar_type_to_pytorch_type[dtype]))
@parse_args('v', 'v', 'i', 'v', 'v', 'v')
def new_zeros(g, self, sizes, dtype, layout, device, pin_memory=False):
if dtype is None:
dtype = self.type().scalarType()
dtype = sym_help.scalar_type_to_onnx.index(sym_help.cast_pytorch_to_onnx[dtype])
return zeros(g, sizes, dtype, layout, device, pin_memory)
@parse_args('v', 'i', 'v', 'v', 'v')
def ones(g, sizes, dtype, layout, device, pin_memory=False):
if dtype is None:
dtype = 6 # float
return g.op("ConstantOfShape", sizes,
value_t=torch.tensor([1], dtype=sym_help.scalar_type_to_pytorch_type[dtype]))
@parse_args('v', 'i', 'v', 'v', 'v', 'v')
def ones_like(g, input, dtype=None, layout=None, device=None, pin_memory=False, memory_format=None):
shape = g.op("Shape", input)
if dtype is None:
dtype = 6 # float
return g.op("ConstantOfShape", shape,
value_t=torch.tensor([1], dtype=sym_help.scalar_type_to_pytorch_type[dtype]))
def full(g, sizes, value, dtype, layout, device, pin_memory=False):
if dtype is None:
dtype = 6 # float
const_value = sym_help._maybe_get_const(value, 't')
if sym_help._is_value(const_value):
tmp = zeros(g, sizes, dtype, layout, device)
return add(g, tmp, value, g.op("Constant", value_t=torch.tensor(1)))
else:
dtype = sym_help._get_const(dtype, 'i', 'dtype')
return g.op("ConstantOfShape", sizes,
value_t=torch.tensor([const_value], dtype=sym_help.scalar_type_to_pytorch_type[dtype]))
@parse_args('v', 'f', 'i', 'v', 'v', 'v', 'v')
def full_like(g, input, fill_value, dtype=None, layout=None, device=None, pin_memory=False, memory_format=None):
shape = g.op("Shape", input)
if dtype is None:
dtype = 6 # float
return g.op("ConstantOfShape", shape,
value_t=torch.tensor([fill_value], dtype=sym_help.scalar_type_to_pytorch_type[dtype]))
@parse_args('v', 'v', 'v', 'v', 'i')
def slice(g, self, dim, start, end, step):
if step != 1:
_unimplemented("slice", "step!=1 is currently not supported")
if start.node().kind() != 'onnx::Constant' or \
end.node().kind() != 'onnx::Constant' or dim.node().kind() != 'onnx::Constant':
if sym_help._operator_export_type == torch.onnx.OperatorExportTypes.ONNX:
raise RuntimeError('Unsupported: ONNX export of Slice with dynamic inputs. DynamicSlice '
'is a deprecated experimental op. Please use statically allocated '
'variables or export to a higher opset version.')
else:
start_unsqueezed = g.op("Unsqueeze", start, axes_i=[0])
end_unsqueezed = g.op("Unsqueeze", end, axes_i=[0])
dim_unsqueezed = g.op("Unsqueeze", dim, axes_i=[0])
return g.op("DynamicSlice", self, start_unsqueezed, end_unsqueezed, dim_unsqueezed)
else:
start = _parse_arg(start, 'i')
end = _parse_arg(end, 'i')
dim = _parse_arg(dim, 'i')
return sym_help._slice_helper(g, self, axes=[dim], starts=[start], ends=[end])
@parse_args('v', 'f', 'f')
def hardtanh(g, self, min_val, max_val):
return g.op("Clip", self, min_f=min_val, max_f=max_val)
def alias(g, self):
return self
@parse_args('v', 'i')
def unsqueeze(g, self, dim):
# Handle negative dim
if dim < 0:
rank = self.type().dim()
if rank:
warnings.warn("ONNX export unsqueeze with negative axis " + str(dim) +
" might cause the onnx model to be incorrect. " +
"Negative axis is not supported in ONNX. " +
"Axis is converted to " + str(dim + rank + 1) +
" based on input shape at export time. " +
"Passing an tensor of different rank in execution will be incorrect.")
dim = dim + rank + 1
else:
return _unimplemented('unsqueeze', 'negative axis with unknown input rank')
return g.op("Unsqueeze", self, axes_i=[dim])
@parse_args('v', 'i', 'i', 'none')
def sort(g, self, dim, decending, out=None):
if out is not None:
_unimplemented("Sort", "Out parameter is not supported for sort")
if not self.isCompleteTensor():
return _unimplemented("Sort", "input size not accessible")
return g.op("TopK", self, k_i=self.type().sizes()[dim], axis_i=dim, outputs=2)
@parse_args('v', 'i', 'i', 'i', 'i', 'none')
def topk(g, self, k, dim, largest, sorted, out=None):
if out is not None:
_unimplemented("TopK", "Out parameter is not supported for topk")
if not largest:
_unimplemented("TopK", "Ascending TopK is not supported")
return g.op("TopK", self, k_i=k, axis_i=dim, outputs=2)
def to(g, self, *args):
# ONNX doesn't have a concept of a device, so we ignore device casts
if len(args) == 4:
if args[0].type().isSubtypeOf(ListType.ofInts()):
# aten::to(Tensor, Device, bool, bool, memory_format)
return self
else:
# aten::to(Tensor, ScalarType, bool, bool, memory_format)
dtype = sym_help._get_const(args[0], 'i', 'dtype')
# memory_format is ignored
return g.op("Cast", self, to_i=sym_help.scalar_type_to_onnx[dtype])
elif len(args) == 5:
# aten::to(Tensor, Device, ScalarType, bool, bool, memory_format)
dtype = sym_help._get_const(args[1], 'i', 'dtype')
# memory_format is ignored
return g.op("Cast", self, to_i=sym_help.scalar_type_to_onnx[dtype])
elif len(args) == 6:
# aten::to(Tensor, ScalarType, Layout, Device, bool, bool, memory_format) -> Tensor
dtype = sym_help._get_const(args[0], 'i', 'dtype')
# Layout, device and memory_format are ignored
return g.op("Cast", self, to_i=sym_help.scalar_type_to_onnx[dtype])
elif len(args) == 7:
# aten::to(Tensor, ScalarType, Layout, Device, bool, bool, bool, memory_format) -> Tensor
dtype = sym_help._get_const(args[0], 'i', 'dtype')
# Layout, device and memory_format are ignored
return g.op("Cast", self, to_i=sym_help.scalar_type_to_onnx[dtype])
else:
raise NotImplementedError("Unknown aten::to signature")
def repeat(g, self, repeats):
if not sym_help._is_value(repeats):
repeats = g.op("Constant", value_t=torch.LongTensor(repeats))
const_repeats = sym_help._maybe_get_const(repeats, 'is')
if self.isCompleteTensor() and not sym_help._is_value(const_repeats):
sizes = self.type().sizes()
diff_dims = len(const_repeats) - len(sizes)
if diff_dims > 0:
self = view(g, self, [1] * diff_dims + sizes)
return g.op("Tile", self, repeats)
@parse_args('v', 'i')
def pixel_shuffle(g, self, upscale_factor):
dims = self.type().sizes()
if len(dims) != 4:
return _unimplemented("pixel_shuffle", "only support 4d input")
output_channel = dims[1] // upscale_factor // upscale_factor
after_view = view(g, self, [-1, output_channel, upscale_factor, upscale_factor,
dims[2], dims[3]])
after_transpose = g.op("Transpose", after_view, perm_i=[0, 1, 4, 2, 5, 3])
return view(g, after_transpose,
[-1, output_channel, dims[2] * upscale_factor, dims[3] *
upscale_factor])
def _generic_rnn(g, variant, input, initial_states, all_weights, has_biases,
num_layers, dropout, train, bidirectional, batch_first=None, batch_sizes=None):
warnings.warn("Exporting a model to ONNX with a batch_size other than 1, " +
"with a variable length with " + variant + " can cause an error " +
"when running the ONNX model with a different batch size. " +
"Make sure to save the model with a batch size of 1, " +
"or define the initial states (h0/c0) as inputs of the model. ")
onnxActivations = ['Relu', 'Tanh', 'Sigmoid', 'Affine', 'LeakyRelu', 'ThresholdedRelu',
'ScaledTanh', 'HardSigmoid', 'Elu', 'Softsign', 'Softplus']
variantToOnnxActivationMap = dict(zip([act_fun.lower() for act_fun in onnxActivations], onnxActivations))
weights_per_layer = 4 if has_biases else 2
assert len(all_weights) == num_layers * weights_per_layer * (1 + bidirectional)
layer_weights = [all_weights[i:i + weights_per_layer] for i in range(0, len(all_weights), weights_per_layer)]
if batch_first:
# batch, seq, feat -> seq, batch, feat
input = g.op('Transpose', input, perm_i=[1, 0, 2])
if dropout and train:
return _unimplemented("RNN/GRU/LSTM", "dropout in training mode")
if variant.startswith('RNN'):
nonlinearity = variantToOnnxActivationMap[variant[4:].lower()]
variant = 'RNN'
w_hh = all_weights[1]
hidden_size = w_hh.type().sizes()[1]
unidirectional = not bidirectional
prev_output = input
h_outs = []
if variant == 'RNN' or variant == 'GRU':
h0 = initial_states
elif variant == 'LSTM':
h0, c0 = initial_states
c_outs = []
sequence_lens = unused(g) if batch_sizes is None else batch_sizes
if variant == 'GRU':
# pytorch is reset, input, hidden
# onnx is input, reset, hidden
reform_permutation = [(1, 2), (0, 1), (2, 3)]
elif variant == 'LSTM':
# pytorch is input, forget, cell, output.
# onnx is input, output, forget, cell.
reform_permutation = [(0, 1), (3, 4), (1, 3)]
def reform_weights(g, w, n, intervals):
slices = [sym_help._slice_helper(g, w, axes=[0], starts=[x * n], ends=[y * n]) for x, y in intervals]
return g.op('Concat', *slices, axis_i=0)
def transform_weights(layer_index):
if variant == 'RNN':
weight_ih, weight_hh, bias_ih, bias_hh = layer_weights[layer_index]
elif variant == 'GRU' or variant == 'LSTM':
weight_ih, weight_hh, bias_ih, bias_hh = \
[reform_weights(g, w, hidden_size, reform_permutation) for w in layer_weights[layer_index]]
bias_concat = g.op('Concat', bias_ih, bias_hh, axis_i=0)
return tuple(g.op('Unsqueeze', x, axes_i=[0]) for x in (weight_ih, weight_hh, bias_concat))
def retrieve_state(x, start, end):
return x if num_layers == 1 else sym_help._slice_helper(g, x, axes=[0], starts=[start], ends=[end])
for i in range(num_layers):
if unidirectional:
weight_ih, weight_hh, bias_concat = transform_weights(i)
state_indices = i, i + 1
else:
weight_ih_f, weight_hh_f, bias_f = transform_weights(2 * i)
weight_ih_b, weight_hh_b, bias_b = transform_weights(2 * i + 1)
weight_ih = g.op('Concat', weight_ih_f, weight_ih_b, axis_i=0)
weight_hh = g.op('Concat', weight_hh_f, weight_hh_b, axis_i=0)
bias_concat = g.op('Concat', bias_f, bias_b, axis_i=0)
state_indices = 2 * i, 2 * i + 2
inputs = [prev_output, weight_ih, weight_hh, bias_concat, sequence_lens]
inputs.append(retrieve_state(h0, *state_indices))
if variant == 'LSTM':
inputs.append(retrieve_state(c0, *state_indices))
extra_kwargs = {} if unidirectional else {'direction_s': 'bidirectional'}
if variant == 'RNN':
if bidirectional:
activation = [nonlinearity, nonlinearity]
else:
activation = [nonlinearity]
prev_output, h_out = g.op('RNN', *inputs, outputs=2,
hidden_size_i=hidden_size,
activations_s=activation,
**extra_kwargs)
elif variant == 'GRU':
prev_output, h_out = g.op('GRU', *inputs, outputs=2,
hidden_size_i=hidden_size,
linear_before_reset_i=1,
**extra_kwargs)
elif variant == 'LSTM':
prev_output, h_out, c_out = g.op('LSTM', *inputs, outputs=3,
hidden_size_i=hidden_size,
**extra_kwargs)
if bidirectional:
# The ONNX RNN/GRU/LSTM produce an output of dimensions
# seq_len, num_directions, batch, hidden_size
# We have to convert to match pytorch's expected
# seq_len, batch, num_directions * hidden_size
# by first moving num_directions before hidden_size with
# Transpose, and then combining it with hidden_size
# with Reshape.
prev_output = g.op('Transpose', prev_output, perm_i=[0, 2, 1, 3])
prev_output = g.op('Reshape', prev_output, g.op('Constant', value_t=torch.LongTensor([0, 0, -1])))
else:
prev_output = g.op('Squeeze', prev_output, axes_i=[1])
h_outs.append(h_out)
if variant == 'LSTM':
c_outs.append(c_out)
if batch_first:
# seq, batch, num_directions * hidden_size -> batch, seq, num_directions * hidden_size
prev_output = g.op('Transpose', prev_output, perm_i=[1, 0, 2])
h_outs = h_out if num_layers == 1 else g.op('Concat', *h_outs, axis_i=0)
if variant == 'RNN' or variant == 'GRU':
return prev_output, h_outs
elif variant == 'LSTM':
c_outs = c_out if num_layers == 1 else g.op('Concat', *c_outs, axis_i=0)
return prev_output, h_outs, c_outs
@parse_args('v', 'v', 'v', 'i', 'i', 'f', 'i', 'i', 'i')
def _lstm_full(g, input, hidden_v, weight_v, has_biases, num_layers, dropout, train, bidirectional, batch_first):
hidden, weight = sym_help._unpack_list(hidden_v), sym_help._unpack_list(weight_v)
return _generic_rnn(g, 'LSTM', input, hidden, weight, has_biases, num_layers,
dropout, train, bidirectional, batch_first)
@parse_args('v', 'v', 'v', 'v', 'i', 'i', 'f', 'i', 'i')
def _lstm_packed(g, input, batch_sizes, hidden_v, weight_v, has_biases, num_layers, dropout, train, bidirectional):
hidden, weight = sym_help._unpack_list(hidden_v), sym_help._unpack_list(weight_v)
return _generic_rnn(g, 'LSTM', input, hidden, weight, has_biases, num_layers,
dropout, train, bidirectional, batch_sizes=batch_sizes)
def lstm(g, *args):
if sym_help._is_tensor_list(args[3]):
return _lstm_packed(g, *args)
else:
return _lstm_full(g, *args)
def _one_hidden_rnn(kind):
@parse_args('v', 'v', 'v', 'i', 'i', 'f', 'i', 'i', 'i')
def _rnn_full(g, input, hidden, weight_v, has_biases, num_layers, dropout, train, bidirectional, batch_first):
weight = sym_help._unpack_list(weight_v)
return _generic_rnn(g, kind, input, hidden, weight, has_biases, num_layers,
dropout, train, bidirectional, batch_first)
@parse_args('v', 'v', 'v', 'v', 'i', 'i', 'f', 'i', 'i')
def _rnn_packed(g, input, batch_sizes, hidden, weight_v, has_biases, num_layers, dropout, train, bidirectional):
weight = sym_help._unpack_list(weight_v)
return _generic_rnn(g, kind, input, hidden, weight, has_biases, num_layers,
dropout, train, bidirectional, batch_sizes=batch_sizes)
def symbolic(g, *args):
if sym_help._is_tensor_list(args[3]):
return _rnn_packed(g, *args)
else:
return _rnn_full(g, *args)
return symbolic
gru = _one_hidden_rnn('GRU')
rnn_tanh = _one_hidden_rnn('RNN_TANH')
rnn_relu = _one_hidden_rnn('RNN_RELU')
@parse_args('v', 'i')
def _dim_arange(g, like, dim):
like_shape = g.op('Shape', like)
stop = g.op("Gather", like_shape, g.op("Constant", value_t=torch.tensor(dim)), axis_i=0)
if sym_help._operator_export_type == torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK:
return g.op("_caffe2::Range", stop)
else:
# aten::arange(Scalar end, ScalarType dtype, Layout, Device, bool pin_memory)
return arange(g, stop, 4, None, None, None)
def detach(g, input):
# Erase aten::detach nodes because ONNX is inference only
return input
@parse_args('v', 'i')
def contiguous(g, input, memory_format):
if memory_format > 2: # allower values are any, preserve and contiguous_format
raise RuntimeError("onnx memory_format support is not implemented")
return input
@parse_args('v', 'v', 'i')
def _pack_padded_sequence(g, input, lengths, batch_first):
# There currently is no PackPadded operator in ONNX. We rely on an
# optimization pass to remove this later. It is an error if all
# PackPadded operators cannot be optimized out.
if batch_first:
input = g.op('Transpose', input, perm_i=[1, 0, 2])
if not lengths.type().isSubtypeOf(torch._C.TensorType.get()):
raise RuntimeError("Lengths must be a Tensor for ONNX export")
# We know it's a TensorType so this check is now safe.
# It's really only necessary because those operators expand to something that
# only works with int32 types in Caffe2...
if lengths.type().scalarType() != 'Int':
lengths = _cast_Int(g, lengths, False)
return g.op("prim::PackPadded", input, lengths, outputs=2)
@parse_args('v', 'v', 'i', 't', 'v')
def _pad_packed_sequence(g, data, batch_sizes, batch_first, padding_value, total_length):
# Ignore total_length as it is not supported in _symbolic_pad_packed_sequence
# It is only useful/used when training using data_parallel model, so
# It shouldn't be relevant for ONNX anyway
data, lengths = g.op("prim::PadPacked", data, batch_sizes, outputs=2)
if batch_first:
data = g.op('Transpose', data, perm_i=[1, 0, 2])
return data, lengths
def randn(g, shapes, dtype, *options):
dtype = sym_help._get_const(dtype, 'i', 'dtype')
if dtype is None:
dtype = 6 # float
if sym_help._is_packed_list(shapes):
shape_const = g.op("ConstantOfShape", shapes,
value_t=torch.tensor([0], dtype=sym_help.scalar_type_to_pytorch_type[6]))
return g.op('RandomNormalLike', shape_const, dtype_i=sym_help.scalar_type_to_onnx[dtype])
shape = sym_help._get_const(shapes, "is", "randn")
return g.op('RandomNormal', shape_i=shape)
def rand(g, shapes, dtype, *options):
dtype = sym_help._get_const(dtype, 'i', 'dtype')
if dtype is None:
dtype = 6 # float
if sym_help._is_packed_list(shapes):
shape_const = g.op("ConstantOfShape", shapes,
value_t=torch.tensor([0], dtype=sym_help.scalar_type_to_pytorch_type[6]))
return g.op('RandomUniformLike', shape_const, dtype_i=sym_help.scalar_type_to_onnx[dtype])
shape = sym_help._get_const(shapes, "is", "rand")
return g.op('RandomUniform', shape_i=shape)
def randn_like(g, self, dtype, layout=None, device=None, pin_memory=False, memory_format=None):
dtype = sym_help._get_const(dtype, 'i', 'dtype')
if dtype is None:
dtype = 6 # float
return g.op('RandomNormalLike', self, dtype_i=sym_help.scalar_type_to_onnx[dtype])
def rand_like(g, self, dtype, layout=None, device=None, pin_memory=False, memory_format=None):
dtype = sym_help._get_const(dtype, 'i', 'dtype')
if dtype is None:
dtype = 6 # float
return g.op('RandomUniformLike', self, dtype_i=sym_help.scalar_type_to_onnx[dtype])
@parse_args('v', 'f', 'f', 'i', 'none')
def rrelu(g, input, lower, upper, training, generator):
p = g.op('RandomUniformLike', input, high_f=upper, low_f=lower)
return g.op('PRelu', input, p)
@parse_args('v')
def log_sigmoid(g, input):
p = g.op('Sigmoid', input)
return g.op('Log', p)
@parse_args('v')
def erf(g, input):
return g.op('Erf', input)
@parse_args('v', 'i', 'i')
def flatten(g, input, start_dim, end_dim):
dim = input.type().dim()
# TODO: remove this as onnx opset 11 spec allows negative axes
if end_dim < 0 :
end_dim = dim + end_dim
# use ONNX's Flatten operator for cases where the output shape is 2D
if start_dim == 1 and end_dim == dim - 1 :
return g.op("Flatten", input, axis_i=start_dim)
if start_dim == 0 and end_dim == dim - 2 :
return g.op("Flatten", input, axis_i=end_dim + 1)
# use Reshape for cases where the output shape is not 2D
if not input.isCompleteTensor():
return _unimplemented("flatten",
"input size not accessible "
"(consider using reshape op instead of flatten op to export to ONNX)")
input_dims = input.type().sizes()
output_dims = []
for i in range(0, dim):
if start_dim < i and end_dim >= i:
output_dims[start_dim] = output_dims[start_dim] * input_dims[i]
else:
output_dims.append(input_dims[i])
shape = g.op("Constant", value_t=torch.LongTensor(output_dims))
p = _reshape_from_tensor(g, input, shape)
return p
@parse_args('v')
def nonzero(g, input):
return t(g, g.op('NonZero', input))
@parse_args('v')
def isnan(g, input):
output = g.op('IsNaN', input)
return output
@parse_args('v', 'i', 'i', 'i')
def narrow(g, input, dim, start, length):
return sym_help._slice_helper(g, input, axes=[dim], starts=[start], ends=[start + length])
def argmax(g, input, dim, keepdim):
if sym_help._is_none(dim):
flattened = reshape(g, input, (-1,))
return g.op('ArgMax', flattened, axis_i=0, keepdims_i=False)
else:
dim = _parse_arg(dim, 'i')
keepdim = _parse_arg(keepdim, 'i')
return g.op('ArgMax', input, axis_i=dim, keepdims_i=keepdim)
def argmin(g, input, dim, keepdim):
if sym_help._is_none(dim):
flattened = reshape(g, input, (-1,))
return g.op('ArgMin', flattened, axis_i=0, keepdims_i=False)
else:
dim = _parse_arg(dim, 'i')
keepdim = _parse_arg(keepdim, 'i')
return g.op('ArgMin', input, axis_i=dim, keepdims_i=keepdim)
@parse_args('v', 'i', 'v', 'v')
def scatter(g, self, dim, index, src):
return g.op("Scatter", self, index, src, axis_i=dim)
@parse_args('v', 'i', 'v', 'v')
def scatter_add(g, self, dim, index, src):
if not self.isCompleteTensor():
return _unimplemented("scatter_add", "input size not accessible")
dtype = self.type().scalarType()
dtype = sym_help.scalar_type_to_onnx.index(sym_help.cast_pytorch_to_onnx[dtype])
dtype = sym_help.scalar_type_to_pytorch_type[dtype]
sizes = self.type().sizes()
to_add = g.op("Constant", value_t=torch.zeros(sizes, dtype=dtype))
to_add = sym_help._scatter_helper(g, to_add, dim, index, src)
return add(g, self, to_add)
def log2(g, self):
_ln2 = 0.693147180559945309
return g.op('Div', log(g, self), g.op('Constant', value_t=torch.Tensor([_ln2])))
def prim_shape(g, self):
return g.op('Shape', self)
@parse_args('v', 'i')
def one_hot(g, self, num_classes):
values = g.op("Constant", value_t=torch.LongTensor([0, 1]))
depth = g.op("Constant", value_t=torch.LongTensor([num_classes]))
return g.op("OneHot", self, depth, values, axis_i=-1)
@parse_args('v', 'i', 'v', 'v')
def gather(g, self, dim, index, sparse_grad=False):
if sym_help._maybe_get_const(sparse_grad, 'i'):
return _unimplemented("gather", "sparse_grad == True")
# NOTE: This workaround is needed since GatherElement is only supported
# since opset 11, and Gather in ONNX is not the same as torch.gather.
dtype = self.type().scalarType()
values = g.op("Constant", value_t=torch.LongTensor([0, 1]))
depth = size(g, self, g.op("Constant", value_t=torch.LongTensor([dim])))
index = g.op("Cast", g.op("OneHot", index, depth, values, axis_i=dim), to_i=sym_help.cast_pytorch_to_onnx[dtype])
mul = g.op("Mul", g.op("Unsqueeze", self, axes_i=[dim + 1]), index)
return g.op("ReduceSum", mul, axes_i=[dim], keepdims_i=0)
@parse_args('v', 'is', 'b', 'i')
def _std(g, input, dim, unbiased, keepdim):
if input.isCompleteTensor():
sqrd = g.op("Mul", input, input)
if dim is None:
sqrdmean = g.op("ReduceMean", sqrd, keepdims_i=0)
mean = g.op("ReduceMean", input, keepdims_i=0)
redudced_dims = input.type().sizes()
else:
sqrdmean = g.op("ReduceMean", sqrd, axes_i=dim, keepdims_i=keepdim)
mean = g.op("ReduceMean", input, axes_i=dim, keepdims_i=keepdim)
redudced_dims = [input.type().sizes()[i] for i in dim]
meansqrd = g.op("Mul", mean, mean)
var = g.op("Abs", g.op("Sub", sqrdmean, meansqrd))
# This is to correct bias in calculating variance, by dividing it over (N - 1) instead on N
if unbiased:
count = numpy.prod(redudced_dims)
mul = g.op("Mul", var, g.op("Constant", value_t=torch.tensor(count, dtype=torch.float)))
var = g.op("Div", mul, g.op("Constant", value_t=torch.tensor(count - 1, dtype=torch.float)))
std = g.op("Sqrt", var)
return std
else:
_unimplemented("std", "Unknown input rank. Cannot compute std along dimensions.")
# Since position of optional arguments can change for std, this is a hack to find if first argument
# is 'dim' or 'unbiased'. As shown below, 'dim' argument could be listed before 'unbiased' :
# torch.std(input, unbiased=True)
# torch.std(input, dim, keepdim=False, unbiased=True)
def std(g, input, *args):
if args[0].type().isSubtypeOf(ListType.ofInts()):
return _std(g, input, *args)
else:
return _std(g, input, None, args[0], None)
@parse_args('v', 'is', 'i')
def logsumexp(g, input, dim, keepdim):
return g.op('ReduceLogSumExp', input, axes_i=dim, keepdims_i=keepdim)
def arange(g, *args):
if sym_help._operator_export_type == torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK:
return g.op("ATen", *args, operator_s="arange")
def _get_arange_dtype(dtype):
dtype = sym_help._maybe_get_const(dtype, 'i')
if sym_help._is_value(dtype):
dtype = 4 # default to int64
return dtype
if len(args) == 5:
# aten::arange(Scalar end, ScalarType dtype, Layout, Device, bool pin_memory)
dtype = _get_arange_dtype(args[1])
end = g.op("Unsqueeze", args[0], axes_i=[0])
arange_tensor = g.op("Squeeze", nonzero(g, ones(g, end, dtype, *(args[2:]))), axes_i=[1])
return g.op("Cast", arange_tensor, to_i=sym_help.scalar_type_to_onnx[dtype])
elif len(args) == 6:
# aten::arange(Scalar start, Scalar end, ScalarType dtype, Layout, Device, bool pin_memory)
dtype = _get_arange_dtype(args[2])
end = g.op("Unsqueeze", args[1], axes_i=[0])
start = g.op("Unsqueeze", args[0], axes_i=[0])
range_tensor = g.op("Sub", end, start)
arange_tensor = g.op("Add", g.op("Squeeze", nonzero(g, ones(g, range_tensor, dtype, *(args[3:]))), axes_i=[1]), start)
return g.op("Cast", arange_tensor, to_i=sym_help.scalar_type_to_onnx[dtype])
elif len(args) == 7:
# aten::arange(Scalar start, Scalar end, Scalar step, ScalarType dtype, Layout, Device, bool pin_memory)
dtype = _get_arange_dtype(args[3])
step = g.op("Unsqueeze", args[2], axes_i=[0])
end = g.op("Unsqueeze", args[1], axes_i=[0])
start = g.op("Unsqueeze", args[0], axes_i=[0])
range_tensor = g.op("Div", g.op("Sub", end, start), step)
arange_tensor = g.op("Squeeze", nonzero(g, ones(g, range_tensor, dtype, *(args[4:]))), axes_i=[1])
arange_tensor = g.op("Add", g.op("Mul", arange_tensor, step), start)
return g.op("Cast", arange_tensor, to_i=sym_help.scalar_type_to_onnx[dtype])
else:
raise NotImplementedError("Unknown aten::arange signature taking " + str(len(args)) + " arguments.")
def masked_fill(g, self, mask, value):
mask = _cast_Bool(g, mask, False)
value = sym_help._maybe_get_scalar(value)
return g.op('Where', mask, sym_help._if_scalar_type_as(g, value, self), self)
def index(g, self, index):
if sym_help._operator_export_type == torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK:
return g.op("ATen", self, index, operator_s="index")
if sym_help._is_packed_list(index):
indices = sym_help._unpack_list(index)
else:
indices = [index]
def try_mask_to_index(index):
if not sym_help._is_none(index) and (index.type().scalarType() == "Byte" or index.type().scalarType() == "Bool"):
if sym_help._export_onnx_opset_version < 9:
raise RuntimeError("Exporting masked indices are only supported after ONNX opset 9.")
warnings.warn("Exporting aten::index operator with indices of type Byte. "
"Only 1-D indices are supported. In any other case, "
"this will produce an incorrect ONNX graph.")
index = squeeze(g, nonzero(g, index), dim=1)
return index
indices = [try_mask_to_index(idx) for idx in indices]
if len(indices) == 1:
return index_select(g, self, 0, indices[0])
else:
# Multiple tensors as indices. Each tensor could either be
# 1. prim::Constant()
# representing ":" in python indexing. E.g. tensor[:, :]
# 2. prim::Constant[value=...] or tensor output
# representing advanced indexing. E.g. tensor[[0, 1], [2, 0]].
# For more info on advanced indexing,
# check https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html#advanced-indexing
# Consider a general case of
# t: [x_1, y_1, y_2, ..., x_m, ..., y_n]
# where t is a tensor of rank m+n, {x_i} are axes where tensor index is provided, and {y_i} are axes for ":".
# Same results can be achieved through transposing t into
# t: [x_1, x_2, ..., x_m, y_1, y_2, ..., y_n]
# and use gatherND. However ONNX does not have gatherND, to use 1d gather we'll need to flatten t
# and process the tensor indices.
# t: [x_1 * x_2 * ... * x_m, y_1 * y_2 * ... * y_n]
# tensor index = \sum_{i=1}^m (ind_i * \prod_{j=i+1}^m (x_j))
# After gather, reshape and transpose back.
adv_idx_indices = [i for i, idx in enumerate(indices) if not sym_help._is_none(idx)]
if len(adv_idx_indices) == 0:
return self
elif len(adv_idx_indices) == 1:
return index_select(g, self, adv_idx_indices[0], indices[adv_idx_indices[0]])
else:
rank = self.type().dim()
if rank is None:
raise NotImplementedError("Unsupported aten::index operator of advanced indexing on tensor of unknown rank, " +
"try turning on shape and type propagate during export: " +
"torch.onnx._export(..., propagate=True).")
# TODO: If indexing is supported natively in ONNX in future opsets,
# update the warning to recommend exporting with higher opset version.
warnings.warn("Exporting aten::index operator of advanced indexing in opset " +
str(sym_help._export_onnx_opset_version) +
" is achieved by combination of multiple ONNX operators, " +
"including Reshape, Transpose, Concat, and Gather. " +
"If indices include negative values, the exported graph will produce incorrect results.")
rank = self.type().dim()
adv_idx_count = len(adv_idx_indices)
shape_tensor = _shape_as_tensor(g, self)
dim_tensor_list = [
g.op("Gather", shape_tensor, g.op("Constant", value_t=torch.LongTensor([dim])), axis_i=0) for dim in range(rank)
]
self = g.op("Transpose", self, perm_i=adv_idx_indices + [i for i in range(rank) if i not in adv_idx_indices])
self = g.op("Flatten", self, axis_i=adv_idx_count)
# Note that tensor indices will be broadcasted while accumulating. Thus we get the final subarray shape as well.
cum_adv_index = indices[adv_idx_indices[-1]]
multiplier = dim_tensor_list[adv_idx_indices[-1]]
for i in range(adv_idx_count - 2, -1, -1):
adv_index = g.op("Mul", indices[adv_idx_indices[i]], multiplier)
cum_adv_index = g.op("Add", cum_adv_index, adv_index)
multiplier = g.op("Mul", multiplier, dim_tensor_list[adv_idx_indices[i]])
# perform gather
self = index_select(g, self, 0, cum_adv_index)
cum_adv_index_shape_tensor = _shape_as_tensor(g, cum_adv_index)
# check if all advanced indices are consecutive.
# Refer to https://docs.scipy.org/doc/numpy/reference/arrays.indexing.html#combining-advanced-and-basic-indexing
# to understand how the subarray position is decided.
if adv_idx_indices == list(range(adv_idx_indices[0], adv_idx_indices[-1] + 1)):
# unfold regular index axes
folded_adv_idx_shape_list = [g.op("Constant", value_t=torch.LongTensor([-1]))] \
+ [dim_tensor_list[i] for i in range(rank) if i not in adv_idx_indices]
folded_adv_idx_shape = g.op("Concat", *folded_adv_idx_shape_list, axis_i=0)
self = g.op("Reshape", self, folded_adv_idx_shape)
# Transpose folded advanced indexed axis to its original location.
adv_idx_permute = list(range(1, adv_idx_indices[0] + 1)) \
+ [0] + list(range(adv_idx_indices[0] + 1, rank - adv_idx_count + 1))
self = g.op("Transpose", self, perm_i=adv_idx_permute)
# unfold advanced index axes
final_shape_list = [dim_tensor_list[i] for i in range(adv_idx_indices[0])] \
+ [cum_adv_index_shape_tensor] \
+ [dim_tensor_list[i] for i in range(adv_idx_indices[0], rank) if i not in adv_idx_indices]
final_shape = g.op("Concat", *final_shape_list, axis_i=0)
else:
final_shape = g.op(
"Concat",
cum_adv_index_shape_tensor,
*[dim_tensor_list[i] for i in range(rank) if i not in adv_idx_indices],
axis_i=0)
return g.op("Reshape", self, final_shape)
@parse_args('v', 'is', 'i')
def frobenius_norm(g, self, dim=None, keepdim=False):
sqr = g.op('Mul', self, self)
sumsqr = g.op('ReduceSum', sqr, axes_i=dim, keepdims_i=keepdim)
return g.op('Sqrt', sumsqr)
@parse_args('v', 'i', 'b', 'v')
def multinomial(g, input, num_samples, replacement=False, generator=None):
if generator is not None and not sym_help._is_none(generator):
_unimplemented("Multinomial", "generator is not supported for multinomial")
if not replacement and num_samples > 1:
_unimplemented("Multinomial", "replacement=False when num_samples > 1 is not supported for multinomial")
log_input = log(g, input)
return g.op("Multinomial", log_input,
dtype_i=sym_help.cast_pytorch_to_onnx['Long'],
sample_size_i=num_samples)
def baddbmm(g, self, batch1, batch2, beta, alpha):
dtype = self.type().scalarType()
batch_mul = matmul(g, batch1, batch2)
mul_a = mul(g, batch_mul, g.op("Cast", alpha, to_i=sym_help.cast_pytorch_to_onnx[dtype]))
mul_b = mul(g, self, g.op("Cast", beta, to_i=sym_help.cast_pytorch_to_onnx[dtype]))
return add(g, mul_a, mul_b)
def meshgrid(g, tensor_list):
tensors = [view(g, t, torch.LongTensor([-1])) for t in sym_help._unpack_list(tensor_list)]
tensors_shape = [g.op("Shape", t) for t in tensors]
out_shape = g.op("Concat", *tensors_shape, axis_i=0)
out = []
for i, t in enumerate(tensors):
shape_i = [g.op("Constant", value_t=torch.ones(1, dtype=torch.int64))] * len(tensors)
shape_i[i] = tensors_shape[i]
t_reshaped = _reshape_from_tensor(g, t, g.op("Concat", *shape_i, axis_i=0))
out.append(g.op("Expand", t_reshaped, out_shape))
return g.op("prim::ListConstruct", *out)
def remainder(g, input, other):
div = g.op("Div", input, other)
if sym_help._is_fp(input):
div = g.op("Floor", div)
quo = g.op("Mul", div, other)
return g.op("Sub", input, quo)
def gelu(g, self):
_sqrt2 = 1.4142135623730951
erf = g.op('Erf', div(g, self, torch.tensor(_sqrt2)))
erf_plusone = add(g, erf, g.op('Constant', value_t=torch.tensor(1, dtype=torch.float)))
return mul(g, mul(g, self, erf_plusone), g.op('Constant', value_t=torch.tensor(0.5, dtype=torch.float)))
@parse_args('v', 'i', 'v', 'v', 'f', 'i')
def group_norm(g, input, num_groups, weight, bias, eps, cudnn_enabled):
if sym_help._operator_export_type == torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK:
return g.op("ATen", input, weight, bias, num_groups_i=num_groups,
eps_f=eps, cudnn_enabled_i=cudnn_enabled, operator_s="group_norm")
input_sizes = input.type().sizes()
assert input_sizes[1] % num_groups == 0
# 0 in the shape list keeps dimension value unchanged.
shape = [0, num_groups, -1]
input_reshaped = g.op('Reshape', input, g.op('Constant', value_t=torch.LongTensor(shape)))
# C is always divisible by num_groups
# Due to shape difference. we need to apply weight and bias after
# instance norm computation and reshape
weight_ = g.op("Constant", value_t=torch.tensor([1.] * num_groups).type(
'torch.' + input.type().scalarType() + 'Tensor'))
bias_ = g.op("Constant", value_t=torch.tensor([0.] * num_groups).type(
'torch.' + input.type().scalarType() + 'Tensor'))
norm_reshaped = g.op("InstanceNormalization", input_reshaped, weight_, bias_, epsilon_f=eps)
norm = g.op('Reshape', norm_reshaped, g.op("Shape", input))
if weight is None or weight.node().mustBeNone():
weight_value = torch.tensor([1.]).type(
'torch.' + input.type().scalarType() + 'Tensor')
weight = g.op("Constant", value_t=weight_value)
if bias is None or bias.node().mustBeNone():
bias_value = torch.tensor([0.]).type(
'torch.' + input.type().scalarType() + 'Tensor')
bias = g.op("Constant", value_t=bias_value)
# Norm has shape [N, C, *] so we reshape weight and bias to [C, *]
axes = list(range(1, len(input_sizes) - 1))
return add(g, mul(g, norm, g.op("Unsqueeze", weight, axes_i=axes)), g.op("Unsqueeze", bias, axes_i=axes))
@parse_args('v', 'v', 'i')
def _weight_norm(g, weight_v, weight_g, dim):
rank = weight_v.type().dim()
if rank:
# W = g * ((v) / ||v||)
# Compute norm_except_dim for l2 norm. dim = None means over all dims
# torch's weight_norm module sets dim = -1 if it's None.
# This conflicts the logic for negative axes to access dims backwards
# TODO: Might need a fix in torch group_norm module
axes = list(range(rank))
if dim is not None:
if dim < -1:
dim += rank
if dim != -1:
axes.remove(dim)
norm_v = norm(g, weight_v, 2, axes, 1)
div = g.op("Div", weight_v, norm_v)
return g.op("Mul", div, weight_g)
else:
return g.op("ATen", weight_v, weight_g, dim_i=dim, operator_s="_weight_norm")
def dim(g, self):
'''Implement the dim functionality available for a pytorch tensor in ONNX'''
# ONNX does not support dim directly in this opset so we can use 2 ops to get the info
shape = g.op('Shape', self)
return g.op('Size', shape)
def __getitem_(g, self, i):
return select(g, self, g.op("Constant", value_t=torch.tensor([0])), i)
def take(g, self, index):
self_flattened = g.op('Reshape', self, g.op("Constant", value_t=torch.tensor([-1], dtype=torch.int64)))
out = index_select(g, self_flattened, 0, index)
out = reshape_as(g, out, index)
return out