TorchScript 是 Pytorch 模型的中间表达,能够运行在高性能的环境,比如 C++ 的环境中。
TorchScript 提供了捕获模型定义的工具,即使在 PyTorch 的灵活和动态特性下也是如此。首先来看看跟踪 (tracing) 的功能。
import torch
class MyCell(torch.nn.Module):
def __init__(self):
super(MyCell, self).__init__()
self.linear = torch.nn.Linear(4, 4)
def forward(self, x, h):
new_h = torch.tanh(self.linear(x) + h)
return new_h, new_h
my_cell = MyCell()
x, h = torch.rand(3, 4), torch.rand(3, 4)
traced_cell = torch.jit.trace(my_cell, (x, h))
print("=== Tracing ===")
print(traced_cell)
print("=== Tracing call ===")
print(traced_cell(x, h))
print("=== Tracing code ===")
print(traced_cell.code)
print("=== Tracing graph ===")
torch.jit.trace
调用了模块,在运行模块时记录了操作,并创建了 torch.jit.ScriptModule
的一个实例(TracedModule 的一个实例)。
TorchScript 在中间表示(IR)中记录其定义,通常在深度学习中称为 graph
。我们可以使用 .graph
属性检查这个 graph
。
graph
是一个非常底层的表示形式,包含的大部分信息对最终用户并不有用。相反,我们可以使用 .code
属性来给出代码的 Python 语法解释。
输出如下:
=== Tracing ===
<class 'torch.jit._trace.TopLevelTracedModule'>
MyCell(
original_name=MyCell
(linear): Linear(original_name=Linear)
)
=== Tracing call ===
(tensor([[ 0.8078, 0.3965, -0.2237, 0.5210],
[ 0.6914, -0.4180, 0.1072, 0.7222],
[ 0.8706, -0.1659, -0.6110, 0.7282]], grad_fn=<TanhBackward0>), tensor([[ 0.8078, 0.3965, -0.2237, 0.5210],
[ 0.6914, -0.4180, 0.1072, 0.7222],
[ 0.8706, -0.1659, -0.6110, 0.7282]], grad_fn=<TanhBackward0>))
=== Tracing code ===
def forward(self,
x: Tensor,
h: Tensor) -> Tuple[Tensor, Tensor]:
linear = self.linear
_0 = torch.tanh(torch.add((linear).forward(x, ), h))
return (_0, _0)
=== Tracing graph ===
graph(%self.1 : __torch__.MyCell,
%x : Float(3, 4, strides=[4, 1], requires_grad=0, device=cpu),
%h : Float(3, 4, strides=[4, 1], requires_grad=0, device=cpu)):
%linear : __torch__.torch.nn.modules.linear.Linear = prim::GetAttr[name="linear"](%self.1)
%20 : Tensor = prim::CallMethod[name="forward"](%linear, %x)
%11 : int = prim::Constant[value=1]() # /path/to/test.py:9:0
%12 : Float(3, 4, strides=[4, 1], requires_grad=1, device=cpu) = aten::add(%20, %h, %11) # /path/to/test.py:9:0
%13 : Float(3, 4, strides=[4, 1], requires_grad=1, device=cpu) = aten::tanh(%12) # /path/to/test.py:9:0
%14 : (Float(3, 4, strides=[4, 1], requires_grad=1, device=cpu), Float(3, 4, strides=[4, 1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%13, %13)
return (%14)
那么这么做这一切的原因是什么呢?有几个原因:
- TorchScript 代码可以在自己的解释器中调用,这基本上是一个受限制的 Python 解释器。这个解释器不会获取全局解释器锁,因此许多请求可以在同一实例上同时处理。
- 这种格式允许我们将整个模型保存到磁盘并加载到另一个环境中,例如在使用 Python 以外的语言编写的服务器中。
- TorchScript 为我们提供了一种表示形式,在这种表示形式中,我们可以对代码进行编译器优化,以提供更高效的执行。
- TorchScript 允许我们与许多后端/设备运行时进行交互,这些运行时需要比单个操作符更广泛的程序视图。
调用 traced_cell 产生的结果与Python模块相同。
print(my_cell(x, h))
print(traced_cell(x, h))
(tensor([[-0.0060, -0.3173, -0.1889, -0.4302],
[ 0.1956, 0.3746, 0.0606, -0.1401],
[ 0.7122, -0.5074, 0.6233, 0.4109]], grad_fn=<TanhBackward0>), tensor([[-0.0060, -0.3173, -0.1889, -0.4302],
[ 0.1956, 0.3746, 0.0606, -0.1401],
[ 0.7122, -0.5074, 0.6233, 0.4109]], grad_fn=<TanhBackward0>))
(tensor([[-0.0060, -0.3173, -0.1889, -0.4302],
[ 0.1956, 0.3746, 0.0606, -0.1401],
[ 0.7122, -0.5074, 0.6233, 0.4109]],
grad_fn=<DifferentiableGraphBackward>), tensor([[-0.0060, -0.3173, -0.1889, -0.4302],
[ 0.1956, 0.3746, 0.0606, -0.1401],
[ 0.7122, -0.5074, 0.6233, 0.4109]],
grad_fn=<DifferentiableGraphBackward>))
虽然 graph 的输出是一个比较底层的代码,还是稍微看看里面内容的含义。
里面用到了底层的库 ATen。ATen 来自于 A TENsor library for C++11 的缩写,是一个张量库,几乎所有 PyTorch 中的其他 Python 和 C++ 接口都是在其之上构建的。它提供了一个核心 Tensor 类,上面定义了许多数百个操作。这些操作中的大多数都有 CPU 和 GPU 实现,Tensor 类会根据其类型动态分派到相应的实现。使用 ATen 的一个小例子如下所示:
#include <ATen/ATen.h>
at::Tensor a = at::ones({2, 2}, at::kInt);
at::Tensor b = at::randn({2, 2});
auto c = a + b.to(at::kInt);
上面 graph 中 aten::add
实现了如下的计算 {aten::add, "${0} + ${2}*${1}"}
,也就是输入 0 加上输入 2 乘以输入 1。其中输入 1 为常数 1,所以结果为两个输入相加。
graph(%self.1 : __torch__.MyCell,
%x : Float(3, 4, strides=[4, 1], requires_grad=0, device=cpu),
%h : Float(3, 4, strides=[4, 1], requires_grad=0, device=cpu)):
# 定义的 Linear 层,从 %self.1 获取属性 linear
%linear : __torch__.torch.nn.modules.linear.Linear = prim::GetAttr[name="linear"](%self.1)
# 调用 %linear 的前向函数,保存到 %20 中
%20 : Tensor = prim::CallMethod[name="forward"](%linear, %x)
# 常数 Tensor 1,保存到 %11 中
%11 : int = prim::Constant[value=1]() # /path/to/test.py:9:0
# 计算 %20 + %h * %11 保存到 %12 中
%12 : Float(3, 4, strides=[4, 1], requires_grad=1, device=cpu) = aten::add(%20, %h, %11) # /path/to/test.py:9:0
# 计算 tanh(%12) 保存到 %13 中
%13 : Float(3, 4, strides=[4, 1], requires_grad=1, device=cpu) = aten::tanh(%12) # /path/to/test.py:9:0
# 构建 tuple(%13, %13) 作为返回,保存到 %14 中
%14 : (Float(3, 4, strides=[4, 1], requires_grad=1, device=cpu), Float(3, 4, strides=[4, 1], requires_grad=1, device=cpu)) = prim::TupleConstruct(%13, %13)
return (%14)
增加一个控制流的子模块。
import torch
class MyDecisionGate(torch.nn.Module):
def forward(self, x):
if x.sum() > 0:
return x
else:
return -x
class MyCell(torch.nn.Module):
def __init__(self, dg):
super(MyCell, self).__init__()
self.dg = dg
self.linear = torch.nn.Linear(4, 4)
def forward(self, x, h):
new_h = torch.tanh(self.dg(self.linear(x)) + h)
return new_h, new_h
x, h = torch.rand(3, 4), torch.rand(3, 4)
my_cell = MyCell(MyDecisionGate())
traced_cell = torch.jit.trace(my_cell, (x, h))
print(traced_cell.dg.code)
print(traced_cell.code)
执行结果如下:
/path/to/control_flow.py:5: TracerWarning: Converting a tensor to a Python boolean might cause the trace to be incorrect. We can't record the data flow of Python values, so this value will be treated as a constant in the future. This means that the trace might not generalize to other inputs!
if x.sum() > 0:
def forward(self,
argument_1: Tensor) -> NoneType:
return None
def forward(self,
x: Tensor,
h: Tensor) -> Tuple[Tensor, Tensor]:
dg = self.dg
linear = self.linear
_0 = (linear).forward(x, )
_1 = (dg).forward(_0, )
_2 = torch.tanh(torch.add(_0, h))
return (_2, _2)
从 .code
输出可以看出,if-else
分支不在其中,为什么呢?跟踪(tracing)确切地做了我们说的事情:运行代码,记录发生的操作并构造一个完全执行这些操作的 ScriptModule
。不幸的是,诸如控制流之类的内容都被删除了。
如何在 TorchScript
中表示这个模块呢?pytorch 提供了一个脚本编译器,它直接分析 Python 源代码,将其转换为 TorchScript。使用脚本编译器将 MyDecisionGate 转换为 TorchScript:
scripted_gate = torch.jit.script(MyDecisionGate())
my_cell = MyCell(scripted_gate)
scripted_cell = torch.jit.script(my_cell)
print(scripted_gate.code)
print(scripted_cell.code)
执行结果如下:
def forward(self,
x: Tensor) -> Tensor:
if bool(torch.gt(torch.sum(x), 0)):
_0 = x
else:
_0 = torch.neg(x)
return _0
def forward(self,
x: Tensor,
h: Tensor) -> Tuple[Tensor, Tensor]:
dg = self.dg
linear = self.linear
_0 = torch.add((dg).forward((linear).forward(x, ), ), h)
new_h = torch.tanh(_0)
return (new_h, new_h)
有些情况需要使用 Tracing 而不是 Scripting(例如,一个模块有许多基于常量值做出的架构决策,我们希望这些值不会出现在 TorchScript 中)。在这种情况下,Scripting 可以与 Tracing 组合使用:torch.jit.script 会将 Tracing 模块的代码内联,而 Tracing 将 scripted 模块的代码内联。
情况一:torch.jit.script 将 Tracing 模块的代码内联
import torch
class MyDecisionGate(torch.nn.Module):
def forward(self, x):
if x.sum() > 0:
return x
else:
return -x
class MyCell(torch.nn.Module):
def __init__(self, dg):
super(MyCell, self).__init__()
self.dg = dg
self.linear = torch.nn.Linear(4, 4)
def forward(self, x, h):
new_h = torch.tanh(self.dg(self.linear(x)) + h)
return new_h, new_h
scripted_gate = torch.jit.script(MyDecisionGate())
class MyRNNLoop(torch.nn.Module):
def __init__(self):
super(MyRNNLoop, self).__init__()
x, h = torch.rand(3, 4), torch.rand(3, 4)
self.cell = torch.jit.trace(MyCell(scripted_gate), (x, h))
def forward(self, xs):
h, y = torch.zeros(3, 4), torch.zeros(3, 4)
for i in range(xs.size(0)):
y, h = self.cell(xs[i], h)
return y, h
rnn_loop = torch.jit.script(MyRNNLoop())
print("=== scripted_gate ===")
print(scripted_gate.code)
print("=== rnn_loop.cell ===")
print(rnn_loop.cell.code)
print("=== rnn_loop ===")
print(rnn_loop.code)
输出结果如下:
=== scripted_gate ===
def forward(self,
x: Tensor) -> Tensor:
if bool(torch.gt(torch.sum(x), 0)):
_0 = x
else:
_0 = torch.neg(x)
return _0
=== rnn_loop.cell ===
def forward(self,
x: Tensor,
h: Tensor) -> Tuple[Tensor, Tensor]:
dg = self.dg
linear = self.linear
_0 = torch.add((dg).forward((linear).forward(x, ), ), h)
_1 = torch.tanh(_0)
return (_1, _1)
=== rnn_loop ===
def forward(self,
xs: Tensor) -> Tuple[Tensor, Tensor]:
h = torch.zeros([3, 4])
y = torch.zeros([3, 4])
y0 = y
h0 = h
for i in range(torch.size(xs, 0)):
cell = self.cell
_0 = (cell).forward(torch.select(xs, 0, i), h0, )
y1, h1, = _0
y0, h0 = y1, h1
return (y0, h0)
torch.select
等价于切片,比如 tensor.select(0, index)
等价于 tensor[index]
参考文档。
情况二:Tracing 将 scripted 模块的代码内联
class WrapRNN(torch.nn.Module):
def __init__(self):
super(WrapRNN, self).__init__()
self.loop = torch.jit.script(MyRNNLoop())
def forward(self, xs):
y, h = self.loop(xs)
return torch.relu(y)
warprnn = WrapRNN()
traced = torch.jit.trace(warprnn(), (torch.rand(10, 3, 4)))
print("=== WrapRNN ===")
print(traced.code)
输出如下:
=== WrapRNN ===
def forward(self,
xs: Tensor) -> Tensor:
loop = self.loop
_0, y, = (loop).forward(xs, )
return torch.relu(y)
当情况需要时,脚本和跟踪都可以使用,并且可以一起使用。
Pytroch 提供 API 将 TorchScript 模型保存到磁盘上的存档格式中,并从中加载。里面包括代码、参数、属性和调试信息,这意味着该存档是模型的一个独立的表示形式,可以在完全分离的进程中加载。
traced.save('wrapped_rnn.pt')
loaded = torch.jit.load('wrapped_rnn.pt')
print(loaded)
print(loaded.code)
xs = torch.rand(3, 3, 4)
print(loaded(xs))
print(warprnn(xs))
输出如下:
RecursiveScriptModule(
original_name=WrapRNN
(loop): RecursiveScriptModule(
original_name=MyRNNLoop
(cell): RecursiveScriptModule(
original_name=MyCell
(dg): RecursiveScriptModule(original_name=MyDecisionGate)
(linear): RecursiveScriptModule(original_name=Linear)
)
)
)
def forward(self,
xs: Tensor) -> Tensor:
loop = self.loop
_0, y, = (loop).forward(xs, )
return torch.relu(y)
tensor([[0.9441, 0.0000, 0.5562, 0.0000],
[0.9451, 0.0000, 0.4321, 0.0000],
[0.9250, 0.0000, 0.0000, 0.0000]], grad_fn=<ReluBackward0>)
tensor([[0.9441, 0.0000, 0.5562, 0.0000],
[0.9451, 0.0000, 0.4321, 0.0000],
[0.9250, 0.0000, 0.0000, 0.0000]], grad_fn=<ReluBackward0>)
保存后的 wrapped_rnn.pt
其实上应该是一个 zip 压缩包。将文件用 unzip 解压后。正如上面所说 torchscript 括代码、参数、属性和调试信息。
Archive: wrapped_rnn.pt.zip
extracting: wrapped_rnn/data/0
extracting: wrapped_rnn/data/1
extracting: wrapped_rnn/data.pkl
inflating: wrapped_rnn/code/__torch__.py
inflating: wrapped_rnn/code/__torch__.py.debug_pkl
inflating: wrapped_rnn/code/__torch__/___torch_mangle_6.py
inflating: wrapped_rnn/code/__torch__/___torch_mangle_6.py.debug_pkl
inflating: wrapped_rnn/code/__torch__/___torch_mangle_3.py
inflating: wrapped_rnn/code/__torch__/___torch_mangle_3.py.debug_pkl
inflating: wrapped_rnn/code/__torch__/torch/nn/modules/linear/___torch_mangle_2.py
inflating: wrapped_rnn/code/__torch__/torch/nn/modules/linear/___torch_mangle_2.py.debug_pkl
extracting: wrapped_rnn/constants.pkl
extracting: wrapped_rnn/version
$ tree .
.
├── code
│ ├── __torch__
│ │ ├── torch
│ │ │ └── nn
│ │ │ └── modules
│ │ │ └── linear
│ │ │ ├── ___torch_mangle_2.py
│ │ │ └── ___torch_mangle_2.py.debug_pkl
│ │ ├── ___torch_mangle_3.py
│ │ ├── ___torch_mangle_3.py.debug_pkl
│ │ ├── ___torch_mangle_6.py
│ │ └── ___torch_mangle_6.py.debug_pkl
│ ├── __torch__.py
│ └── __torch__.py.debug_pkl
├── constants.pkl
├── data
│ ├── 0
│ └── 1
├── data.pkl
└── version
7 directories, 13 files
$ cat code/__torch__.py
class WrapRNN(Module):
__parameters__ = []
__buffers__ = []
training : bool
_is_full_backward_hook : Optional[bool]
loop : __torch__.___torch_mangle_6.MyRNNLoop
def forward(self: __torch__.WrapRNN,
xs: Tensor) -> Tensor:
loop = self.loop
_0, y, = (loop).forward(xs, )
return torch.relu(y)
class MyDecisionGate(Module):
__parameters__ = []
__buffers__ = []
training : bool
_is_full_backward_hook : Optional[bool]
def forward(self: __torch__.MyDecisionGate,
x: Tensor) -> Tensor:
if bool(torch.gt(torch.sum(x), 0)):
_1 = x
else:
_1 = torch.neg(x)
return _1
- https://pytorch.org/tutorials/beginner/Intro_to_TorchScript_tutorial.html
- https://pytorch.org/tutorials/recipes/recipes/saving_and_loading_models_for_inference.html
- https://pytorch.org/docs/stable/jit.html
- https://pytorch.org/docs/master/jit_language_reference.html#language-reference
- https://pytorch.org/cppdocs/