pytorch/test/cpp/api/sequential.cpp
Peter Goldsborough 153e2e96d4 Make Sequential ref-counted (#9151)
Summary:
In the C++ API, `Sequential` currently was not refcounted itself, but stored `shared_ptr<AnyModule>` to get the reference semantics. This is unfortunate because most modules in the API are accessed via `->`, e.g. `Linear l(1, 2); l->forward(...);`. `Sequential` was different in that it had value semantics itself, thus was accessed via `.`.

This PR makes `Sequential` store `AnyModule` (without extra indirection), and uses the same pImpl mechanism we use for all other modules to make `Sequential` have reference semantics itself. This makes it consistent with the rest of the library. It also removes one level of indirection inside of `Sequential`, which is cool.

One thing I had to change was that the `ModuleHolder` with which the whole pImpl thing is implemented previously did some tricks to make `Linear(3, 4)` actually construct `Linear(LinearOptions(3, 4))`. This doesn't work well with `Sequential` since it takes a variadic parameter pack. Instead, I made `ModuleHolder` forward all arguments to the underlying module, and then further pushed the trick to forward parameters to modules' options types into the actual Modules. This adds one constructor per Module in the library. This is not something user modules have to do (unless they want this nice forwarding themselves). It makes the code simpler overall.

ezyang ebetica apaszke
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9151

Reviewed By: ezyang

Differential Revision: D8809298

Pulled By: goldsborough

fbshipit-source-id: da68452c3de912fbc67af330ba93b5220de6909f
2018-07-11 17:24:59 -07:00

276 lines
7.5 KiB
C++

#include <catch.hpp>
#include <torch/nn/modules.h>
#include <torch/nn/modules/linear.h>
#include <torch/nn/modules/sequential.h>
#include <torch/tensor.h>
#include <torch/utils.h>
#include <memory>
#include <vector>
using namespace torch::nn;
using Catch::StartsWith;
TEST_CASE("sequential") {
SECTION("construction from shared pointer") {
struct M : torch::nn::Module {
explicit M(int value_) : value(value_) {}
int value;
int forward() {
return value;
}
};
Sequential sequential(
std::make_shared<M>(1), std::make_shared<M>(2), std::make_shared<M>(3));
REQUIRE(sequential->size() == 3);
}
SECTION("construction from concrete type") {
struct M : torch::nn::Module {
explicit M(int value_) : value(value_) {}
int value;
int forward() {
return value;
}
};
Sequential sequential(M(1), M(2), M(3));
REQUIRE(sequential->size() == 3);
}
SECTION("construction from module holders") {
struct MImpl : torch::nn::Module {
explicit MImpl(int value_) : value(value_) {}
int forward() {
return value;
}
int value;
};
struct M : torch::nn::ModuleHolder<MImpl> {
using torch::nn::ModuleHolder<MImpl>::ModuleHolder;
using torch::nn::ModuleHolder<MImpl>::get;
};
Sequential sequential(M(1), M(2), M(3));
REQUIRE(sequential->size() == 3);
}
SECTION("push_back") {
struct M : torch::nn::Module {
explicit M(int value_) : value(value_) {}
int forward() {
return value;
}
int value;
};
Sequential sequential;
REQUIRE(sequential->size() == 0);
REQUIRE(sequential->is_empty());
sequential->push_back(Linear(3, 4));
REQUIRE(sequential->size() == 1);
sequential->push_back(std::make_shared<M>(1));
REQUIRE(sequential->size() == 2);
sequential->push_back(M(2));
REQUIRE(sequential->size() == 3);
}
SECTION("access") {
struct M : torch::nn::Module {
explicit M(int value_) : value(value_) {}
int forward() {
return value;
}
int value;
};
std::vector<std::shared_ptr<M>> modules = {
std::make_shared<M>(1), std::make_shared<M>(2), std::make_shared<M>(3)};
Sequential sequential;
for (auto& module : modules) {
sequential->push_back(module);
}
REQUIRE(sequential->size() == 3);
SECTION("at()") {
SECTION("returns the correct module for a given index") {
for (size_t i = 0; i < modules.size(); ++i) {
REQUIRE(&sequential->at<M>(i) == modules[i].get());
}
}
SECTION("throws for a bad index") {
REQUIRE_THROWS_WITH(
sequential->at<M>(modules.size() + 1),
StartsWith("Index out of range"));
REQUIRE_THROWS_WITH(
sequential->at<M>(modules.size() + 1000000),
StartsWith("Index out of range"));
}
}
SECTION("ptr()") {
SECTION("returns the correct module for a given index") {
for (size_t i = 0; i < modules.size(); ++i) {
REQUIRE(sequential->ptr(i).get() == modules[i].get());
REQUIRE(sequential[i].get() == modules[i].get());
REQUIRE(sequential->ptr<M>(i).get() == modules[i].get());
}
}
SECTION("throws for a bad index") {
REQUIRE_THROWS_WITH(
sequential->ptr(modules.size() + 1),
StartsWith("Index out of range"));
REQUIRE_THROWS_WITH(
sequential->ptr(modules.size() + 1000000),
StartsWith("Index out of range"));
}
}
}
SECTION("forward") {
SECTION("calling forward() on an empty sequential is disallowed") {
Sequential empty;
REQUIRE_THROWS_WITH(
empty->forward<int>(),
StartsWith("Cannot call forward() on an empty Sequential"));
}
SECTION("calling forward() on a non-empty sequential chains correctly") {
struct MockModule : torch::nn::Module {
explicit MockModule(int value) : expected(value) {}
int expected;
int forward(int value) {
REQUIRE(value == expected);
return value + 1;
}
};
Sequential sequential(MockModule{1}, MockModule{2}, MockModule{3});
REQUIRE(sequential->forward<int>(1) == 4);
}
SECTION("calling forward() with the wrong return type throws") {
struct M : public torch::nn::Module {
int forward() {
return 5;
}
};
Sequential sequential(M{});
REQUIRE(sequential->forward<int>() == 5);
REQUIRE_THROWS_WITH(
sequential->forward<float>(),
StartsWith("The type of the return value "
"is int, but you asked for type float"));
}
SECTION("The return type of forward() defaults to Tensor") {
struct M : public torch::nn::Module {
torch::Tensor forward(torch::Tensor v) {
return v;
}
};
Sequential sequential(M{});
auto variable = torch::ones({3, 3}, torch::requires_grad());
REQUIRE(sequential->forward(variable).equal(variable));
}
}
SECTION("returns the last value") {
torch::manual_seed(0);
Sequential sequential(Linear(10, 3), Linear(3, 5), Linear(5, 100));
auto x = torch::randn({1000, 10}, torch::requires_grad());
auto y = sequential->forward(x);
REQUIRE(y.ndimension() == 2);
REQUIRE(y.size(0) == 1000);
REQUIRE(y.size(1) == 100);
}
SECTION("can hold other important modules") {
Sequential sequential(
Linear(10, 3),
Conv2d(1, 2, 3),
Dropout(0.5),
BatchNorm(5),
Embedding(4, 10),
LSTM(4, 5));
}
SECTION("converts at::Tensor to torch::Tensor correctly") {
struct M : torch::nn::Module {
torch::Tensor forward(torch::Tensor input) {
return input;
}
};
Sequential sequential(M{});
torch::Tensor variable = torch::ones(5);
REQUIRE(sequential->forward(variable).sum().toCFloat() == 5);
at::Tensor tensor_that_is_actually_a_variable = variable * 2;
REQUIRE(
sequential->forward(tensor_that_is_actually_a_variable)
.sum()
.toCFloat() == 10);
}
SECTION("extend() pushes modules from other Sequential") {
struct A : torch::nn::Module {
int forward(int x) {
return x;
}
};
struct B : torch::nn::Module {
int forward(int x) {
return x;
}
};
struct C : torch::nn::Module {
int forward(int x) {
return x;
}
};
struct D : torch::nn::Module {
int forward(int x) {
return x;
}
};
Sequential a(A{}, B{});
Sequential b(C{}, D{});
a->extend(*b);
REQUIRE(a->size() == 4);
REQUIRE(a[0]->as<A>());
REQUIRE(a[1]->as<B>());
REQUIRE(a[2]->as<C>());
REQUIRE(a[3]->as<D>());
REQUIRE(b->size() == 2);
REQUIRE(b[0]->as<C>());
REQUIRE(b[1]->as<D>());
std::vector<std::shared_ptr<A>> c = {std::make_shared<A>(),
std::make_shared<A>()};
b->extend(c);
REQUIRE(b->size() == 4);
REQUIRE(b[0]->as<C>());
REQUIRE(b[1]->as<D>());
REQUIRE(b[2]->as<A>());
REQUIRE(b[3]->as<A>());
}
SECTION("has reference semantics") {
Sequential first(Linear(2, 3), Linear(4, 4), Linear(4, 5));
Sequential second(first);
REQUIRE(first.get() == second.get());
REQUIRE(first->size() == second->size());
REQUIRE(std::equal(
first->begin(),
first->end(),
second->begin(),
[](const AnyModule& first, const AnyModule& second) {
return &first == &second;
}));
}
}