Spaces:
Running
on
Zero
Running
on
Zero
File size: 5,532 Bytes
689a1f3 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 |
# Copyright (c) Facebook, Inc. and its affiliates.
"""
Wrappers around on some nn functions, mainly to support empty tensors.
Ideally, add support directly in PyTorch to empty tensors in those functions.
These can be removed once https://github.com/pytorch/pytorch/issues/12013
is implemented
"""
import warnings
from typing import List, Optional
import torch
from torch.nn import functional as F
from detectron2.utils.env import TORCH_VERSION
def shapes_to_tensor(x: List[int], device: Optional[torch.device] = None) -> torch.Tensor:
"""
Turn a list of integer scalars or integer Tensor scalars into a vector,
in a way that's both traceable and scriptable.
In tracing, `x` should be a list of scalar Tensor, so the output can trace to the inputs.
In scripting or eager, `x` should be a list of int.
"""
if torch.jit.is_scripting():
return torch.as_tensor(x, device=device)
if torch.jit.is_tracing():
assert all(
[isinstance(t, torch.Tensor) for t in x]
), "Shape should be tensor during tracing!"
# as_tensor should not be used in tracing because it records a constant
ret = torch.stack(x)
if ret.device != device: # avoid recording a hard-coded device if not necessary
ret = ret.to(device=device)
return ret
return torch.as_tensor(x, device=device)
def check_if_dynamo_compiling():
if TORCH_VERSION >= (1, 14):
from torch._dynamo import is_compiling
return is_compiling()
else:
return False
def cat(tensors: List[torch.Tensor], dim: int = 0):
"""
Efficient version of torch.cat that avoids a copy if there is only a single element in a list
"""
assert isinstance(tensors, (list, tuple))
if len(tensors) == 1:
return tensors[0]
return torch.cat(tensors, dim)
def empty_input_loss_func_wrapper(loss_func):
def wrapped_loss_func(input, target, *, reduction="mean", **kwargs):
"""
Same as `loss_func`, but returns 0 (instead of nan) for empty inputs.
"""
if target.numel() == 0 and reduction == "mean":
return input.sum() * 0.0 # connect the gradient
return loss_func(input, target, reduction=reduction, **kwargs)
return wrapped_loss_func
cross_entropy = empty_input_loss_func_wrapper(F.cross_entropy)
class _NewEmptyTensorOp(torch.autograd.Function):
@staticmethod
def forward(ctx, x, new_shape):
ctx.shape = x.shape
return x.new_empty(new_shape)
@staticmethod
def backward(ctx, grad):
shape = ctx.shape
return _NewEmptyTensorOp.apply(grad, shape), None
class Conv2d(torch.nn.Conv2d):
"""
A wrapper around :class:`torch.nn.Conv2d` to support empty inputs and more features.
"""
def __init__(self, *args, **kwargs):
"""
Extra keyword arguments supported in addition to those in `torch.nn.Conv2d`:
Args:
norm (nn.Module, optional): a normalization layer
activation (callable(Tensor) -> Tensor): a callable activation function
It assumes that norm layer is used before activation.
"""
norm = kwargs.pop("norm", None)
activation = kwargs.pop("activation", None)
super().__init__(*args, **kwargs)
self.norm = norm
self.activation = activation
def forward(self, x):
# torchscript does not support SyncBatchNorm yet
# https://github.com/pytorch/pytorch/issues/40507
# and we skip these codes in torchscript since:
# 1. currently we only support torchscript in evaluation mode
# 2. features needed by exporting module to torchscript are added in PyTorch 1.6 or
# later version, `Conv2d` in these PyTorch versions has already supported empty inputs.
if not torch.jit.is_scripting():
# Dynamo doesn't support context managers yet
is_dynamo_compiling = check_if_dynamo_compiling()
if not is_dynamo_compiling:
with warnings.catch_warnings(record=True):
if x.numel() == 0 and self.training:
# https://github.com/pytorch/pytorch/issues/12013
assert not isinstance(
self.norm, torch.nn.SyncBatchNorm
), "SyncBatchNorm does not support empty inputs!"
x = F.conv2d(
x, self.weight, self.bias, self.stride, self.padding, self.dilation, self.groups
)
if self.norm is not None:
x = self.norm(x)
if self.activation is not None:
x = self.activation(x)
return x
ConvTranspose2d = torch.nn.ConvTranspose2d
BatchNorm2d = torch.nn.BatchNorm2d
interpolate = F.interpolate
Linear = torch.nn.Linear
def nonzero_tuple(x):
"""
A 'as_tuple=True' version of torch.nonzero to support torchscript.
because of https://github.com/pytorch/pytorch/issues/38718
"""
if torch.jit.is_scripting():
if x.dim() == 0:
return x.unsqueeze(0).nonzero().unbind(1)
return x.nonzero().unbind(1)
else:
return x.nonzero(as_tuple=True)
@torch.jit.script_if_tracing
def move_device_like(src: torch.Tensor, dst: torch.Tensor) -> torch.Tensor:
"""
Tracing friendly way to cast tensor to another tensor's device. Device will be treated
as constant during tracing, scripting the casting process as whole can workaround this issue.
"""
return src.to(dst.device)
|