TorchScript is a statically typed subset of Python that can be interpreted by LibTorch without any Python dependency. The torch R package provides interfaces to create, serialize, load and execute TorchScript programs.
Advantages of using TorchScript are:
TorchScript code can be invoked in its own interpreter, which is basically a restricted Python interpreter. This interpreter does not acquire the Global Interpreter Lock, and so many requests can be processed on the same instance simultaneously.
This format allows us to save the whole model to disk and load it into another environment, such as on server written in a language other than R.
TorchScript gives us a representation in which we can do compiler optimizations on the code to make execution more efficient.
TorchScript allows us to interface with many backend/device runtimes that require a broader view of the program than individual operators.
Creating TorchScript programs
Tracing
TorchScript programs can be created from R using tracing. When using tracing, code is automatically converted into this subset of Python by recording only the actual operators on tensors and simply executing and discarding the other surrounding R code.
Currently tracing is the only supported way to create TorchScript programs from R code.
For example, let’s use the jit_trace function to create
a TorchScript program. We pass a regular R function and example
inputs.
fn <- function(x) {
torch_relu(x)
}
traced_fn <- jit_trace(fn, torch_tensor(c(-1, 0, 1)))The jit_trace function has executed the R function with
the example input and recorded all torch operations that occurred during
execution to create a graph. graph is how we call the
intermediate representation of TorchScript programs, and it can be
inspected with:
traced_fn$graph
#> graph(%0 : Float(3, strides=[1], requires_grad=0, device=cpu)):
#> %1 : Float(3, strides=[1], requires_grad=0, device=cpu) = aten::relu(%0)
#> return (%1)The traced function can now be invoked as a regular R function:
traced_fn(torch_randn(3))
#> torch_tensor
#> 0.4826
#> 0.2149
#> 0.1657
#> [ CPUFloatType{3} ]It’s also possible to trace nn_modules() defined in R,
for example:
module <- nn_module(
initialize = function() {
self$linear1 <- nn_linear(10, 10)
self$linear2 <- nn_linear(10, 1)
},
forward = function(x) {
x %>%
self$linear1() %>%
nnf_relu() %>%
self$linear2()
}
)
traced_module <- jit_trace(module(), torch_randn(10, 10))When using jit_trace with a nn_module only
the forward method is traced. However, by default, one pass
will be conducted in ‘train’ mode, and another one in ‘eval’ mode, which
is different from the PyTorch behavior. One can opt out of this by
specifying respect_mode = FALSE which will only trace the
forward pass in the mode the network is currently in. You can use the
jit_trace_module function to pass example inputs to other
methods. Traced modules look like normal nn_modules(), and
can be called the same way:
traced_module(torch_randn(3, 10))
#> torch_tensor
#> -0.1682
#> -0.2424
#> 0.0244
#> [ CPUFloatType{3,1} ][ grad_fn = <AddmmBackward0> ]Limitations of tracing
- Tracing will not record any control flow like if-statements or loops. When this control flow is constant across your module, this is fine and it often inlines the control flow decisions. But sometimes the control flow is actually part of the model itself. For instance, a recurrent network is a loop over the (possibly dynamic) length of an input sequence. For example:
# fn does does an operation for each dimension of a tensor
fn <- function(x) {
x %>%
torch_unbind(dim = 1) %>%
lapply(function(x) x$sum()) %>%
torch_stack(dim = 1)
}
# we trace using as an example a tensor with size (10, 5, 5)
traced_fn <- jit_trace(fn, torch_randn(10, 5, 5))
# applying it with a tensor with different size returns an error.
traced_fn(torch_randn(11, 5, 5))
#> Error in cpp_call_traced_fn(ptr, inputs): The following operation failed in the TorchScript interpreter.
#> Traceback of TorchScript (most recent call last):
#> RuntimeError: Expected 10 elements in a list but found 11- In the returned
ScriptModule, operations that have different behaviors in training and eval modes will always behave as if it were in the mode it was in during tracing, no matter which mode theScriptModuleis in. For example:
traced_dropout <- jit_trace(nn_dropout(), torch_ones(5,5))
traced_dropout(torch_ones(3,3))
#> torch_tensor
#> 0 0 0
#> 0 0 2
#> 0 0 2
#> [ CPUFloatType{3,3} ]
traced_dropout$eval()
#> [1] FALSE
# even after setting to eval mode, dropout is applied
traced_dropout(torch_ones(3,3))
#> torch_tensor
#> 1 1 1
#> 1 1 1
#> 1 1 1
#> [ CPUFloatType{3,3} ]- Tracing proegrams can only take tensors and lists of tensors as input and return tensors and lists of tensors. For example:
fn <- function(x, y) {
x + y
}
jit_trace(fn, torch_tensor(1), 1)
#> Error in cpp_trace_function(tr_fn, list(...), .compilation_unit, strict, : Only tensors or (possibly nested) dict or tuples of tensors can be inputs to traced functions. Got float
#> Exception raised from addInput at /Users/runner/work/libtorch-mac-m1/libtorch-mac-m1/pytorch/torch/csrc/jit/frontend/tracer.cpp:422 (most recent call first):
#> frame #0: std::__1::shared_ptr<c10::(anonymous namespace)::PyTorchStyleBacktrace> std::__1::make_shared[abi:ue170006]<c10::(anonymous namespace)::PyTorchStyleBacktrace, c10::SourceLocation&, void>(c10::SourceLocation&) + 121 (0x111416639 in libc10.dylib)
#> frame #1: c10::Error::Error(c10::SourceLocation, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>) + 54 (0x111416776 in libc10.dylib)
#> frame #2: c10::detail::torchCheckFail(char const*, char const*, unsigned int, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> const&) + 149 (0x111413035 in libc10.dylib)
#> frame #3: torch::jit::tracer::addInput(std::__1::shared_ptr<torch::jit::tracer::TracingState> const&, c10::IValue const&, c10::Type::SingletonOrSharedTypePtr<c10::Type> const&, torch::jit::Value*) + 6225 (0x128a06cf1 in libtorch_cpu.dylib)
#> frame #4: torch::jit::tracer::addInput(std::__1::shared_ptr<torch::jit::tracer::TracingState> const&, c10::IValue const&, c10::Type::SingletonOrSharedTypePtr<c10::Type> const&, torch::jit::Value*) + 4799 (0x128a0675f in libtorch_cpu.dylib)
#> frame #5: torch::jit::tracer::trace(std::__1::vector<c10::IValue, std::__1::allocator<c10::IValue>>, std::__1::function<std::__1::vector<c10::IValue, std::__1::allocator<c10::IValue>> (std::__1::vector<c10::IValue, std::__1::allocator<c10::IValue>>)> const&, std::__1::function<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>> (at::Tensor const&)>, bool, bool, torch::jit::Module*, std::__1::vector<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>, std::__1::allocator<std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char>>>> const&) + 666 (0x128a03eaa in libtorch_cpu.dylib)
#> frame #6: _lantern_trace_fn + 408 (0x11608de28 in liblantern.dylib)
#> frame #7: cpp_trace_function(Rcpp::Function_Impl<Rcpp::PreserveStorage>, XPtrTorchStack, XPtrTorchCompilationUnit, XPtrTorchstring, bool, XPtrTorchScriptModule, bool, bool) + 601 (0x1142e7959 in torchpkg.so)
#> frame #8: _torch_cpp_trace_function + 719 (0x11410468f in torchpkg.so)
#> frame #9: R_doDotCall + 13245 (0x10e8074bd in libR.dylib)
#> frame #10: bcEval_loop + 146595 (0x10e86f123 in libR.dylib)
#> frame #11: bcEval + 628 (0x10e83cdf4 in libR.dylib)
#> frame #12: Rf_eval + 506 (0x10e83c4fa in libR.dylib)
#> frame #13: R_execClosure + 761 (0x10e83f039 in libR.dylib)
#> frame #14: applyClosure_core + 128 (0x10e83e140 in libR.dylib)
#> frame #15: Rf_eval + 1189 (0x10e83c7a5 in libR.dylib)
#> frame #16: do_eval + 1253 (0x10e843b65 in libR.dylib)
#> frame #17: bcEval_loop + 44444 (0x10e85621c in libR.dylib)
#> frame #18: bcEval + 628 (0x10e83cdf4 in libR.dylib)
#> frame #19: Rf_eval + 506 (0x10e83c4fa in libR.dylib)
#> frame #20: forcePromise + 230 (0x10e83d026 in libR.dylib)
#> frame #21: Rf_eval + 634 (0x10e83c57a in libR.dylib)
#> frame #22: do_withVisible + 57 (0x10e843ef9 in libR.dylib)
#> frame #23: do_internal + 362 (0x10e8bbb6a in libR.dylib)
#> frame #24: bcEval_loop + 45071 (0x10e85648f in libR.dylib)
#> frame #25: bcEval + 628 (0x10e83cdf4 in libR.dylib)
#> frame #26: Rf_eval + 506 (0x10e83c4fa in libR.dylib)
#> frame #27: forcePromise + 230 (0x10e83d026 in libR.dylib)
#> frame #28: Rf_eval + 634 (0x10e83c57a in libR.dylib)
#> frame #29: forcePromise + 230 (0x10e83d026 in libR.dylib)
#> frame #30: bcEval_loop + 19464 (0x10e850088 in libR.dylib)
#> frame #31: bcEval + 628 (0x10e83cdf4 in libR.dylib)
#> frame #32: Rf_eval + 506 (0x10e83c4fa in libR.dylib)
#> frame #33: R_execClosure + 761 (0x10e83f039 in libR.dylib)
#> frame #34: applyClosure_core + 128 (0x10e83e140 in libR.dylib)
#> frame #35: Rf_eval + 1189 (0x10e83c7a5 in libR.dylib)
#> frame #36: do_eval + 1253 (0x10e843b65 in libR.dylib)
#> frame #37: bcEval_loop + 44444 (0x10e85621c in libR.dylib)
#> frame #38: bcEval + 628 (0x10e83cdf4 in libR.dylib)
#> frame #39: Rf_eval + 506 (0x10e83c4fa in libR.dylib)
#> frame #40: R_execClosure + 761 (0x10e83f039 in libR.dylib)
#> frame #41: applyClosure_core + 128 (0x10e83e140 in libR.dylib)
#> frame #42: Rf_eval + 1189 (0x10e83c7a5 in libR.dylib)
#> frame #43: do_begin + 429 (0x10e841a2d in libR.dylib)
#> frame #44: Rf_eval + 990 (0x10e83c6de in libR.dylib)
#> frame #45: R_execClosure + 761 (0x10e83f039 in libR.dylib)
#> frame #46: applyClosure_core + 128 (0x10e83e140 in libR.dylib)
#> frame #47: Rf_eval + 1189 (0x10e83c7a5 in libR.dylib)
#> frame #48: do_docall + 615 (0x10e7cd2a7 in libR.dylib)
#> frame #49: bcEval_loop + 44444 (0x10e85621c in libR.dylib)
#> frame #50: bcEval + 628 (0x10e83cdf4 in libR.dylib)
#> frame #51: Rf_eval + 506 (0x10e83c4fa in libR.dylib)
#> frame #52: R_execClosure + 761 (0x10e83f039 in libR.dylib)
#> frame #53: applyClosure_core + 128 (0x10e83e140 in libR.dylib)
#> frame #54: Rf_eval + 1189 (0x10e83c7a5 in libR.dylib)
#> frame #55: do_docall + 615 (0x10e7cd2a7 in libR.dylib)
#> frame #56: bcEval_loop + 44444 (0x10e85621c in libR.dylib)
#> frame #57: bcEval + 628 (0x10e83cdf4 in libR.dylib)
#> frame #58: Rf_eval + 506 (0x10e83c4fa in libR.dylib)
#> frame #59: R_execClosure + 761 (0x10e83f039 in libR.dylib)
#> frame #60: applyClosure_core + 128 (0x10e83e140 in libR.dylib)
#> frame #61: Rf_eval + 1189 (0x10e83c7a5 in libR.dylib)
#> frame #62: forcePromise + 230 (0x10e83d026 in libR.dylib)
#> :Compiling TorchScript
It’s also possible to create TorchScript programs by compiling TorchScript code. TorchScript code looks a lot like standard python code. For example:
tr <- jit_compile("
def fn (x: Tensor):
return torch.relu(x)
")
tr$fn(torch_tensor(c(-1, 0, 1)))
#> torch_tensor
#> 0
#> 0
#> 1
#> [ CPUFloatType{3} ]Serializing and loading
TorchScript programs can be serialized using the
jit_save function and loaded back from disk with
jit_load.
For example:
fn <- function(x) {
torch_relu(x)
}
tr_fn <- jit_trace(fn, torch_tensor(1))
jit_save(tr_fn, "path.pt")
loaded <- jit_load("path.pt")Loaded programs can be executed as usual:
loaded(torch_tensor(c(-1, 0, 1)))
#> torch_tensor
#> 0
#> 0
#> 1
#> [ CPUFloatType{3} ]Note You can load TorchScript programs that were
created in libraries different than torch for R. Eg, a
TorchScript program can be created in PyTorch with
torch.jit.trace or torch.jit.script, and run
from R.
R objects are automatically converted to their TorchScript
counterpart following the Types table in this document. However,
sometimes it’s necessary to make type annotations with
jit_tuple() and jit_scalar() to disambiguate
the conversion.
Types
The following table lists all TorchScript types and how to convert the to and back to R.
| TorchScript Type | R Description |
|---|---|
Tensor |
A torch_tensor with any shape, dtype or backend. |
Tuple[T0, T1, ..., TN] |
A list() containing subtypes T0,
T1, etc. wrapped with jit_tuple() . |
bool |
A scalar logical value create using jit_scalar. |
int |
A scalar integer value created using jit_scalar. |
float |
A scalar floating value created using jit_scalar. |
str |
A string (ie. character vector of length 1) wrapped in
jit_scalar. |
List[T] |
An R list of which all types are type T . Or numeric
vectors, logical vectors, etc. |
Optional[T] |
Not yet supported. |
Dict[str, V] |
A named list with values of type V . Only
str key values are currently supported. |
T |
Not yet supported. |
E |
Not yet supported. |
NamedTuple[T0, T1, ...] |
A named list containing subtypes T0, T1,
etc. wrapped in jit_tuple(). |