Use default dtype for torch::tensor(floating_point_values) and torch::tensor(empty braced-init-list) when dtype is not specified (#29632)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29632

This PR is BC-breaking in the following way:

Previously, C++ `torch::tensor` with a floating-point literal with no suffix (e.g. `torch::tensor(1.1)`) or a (nested) braced-init-list of
floating-point literals with no suffix (e.g. `torch::tensor({{1.1, 2.2}})` produces a tensor with dtype `at::kDouble`. After this PR, it produces a tensor with dtype `torch::get_default_dtype()`, matching Python `torch.tensor` behavior.

Test Plan: Imported from OSS

Differential Revision: D18465819

Pulled By: yf225

fbshipit-source-id: 6834fe50335c677bc3832f2a5e9cf8d1ede9f665
This commit is contained in:
Will Feng
2019-11-13 15:14:08 -08:00
committed by Facebook Github Bot
parent 3fb9bbc99b
commit 2bcac59a30
9 changed files with 174 additions and 72 deletions
+8 -7
View File
@@ -92,16 +92,16 @@ void check_exact_values(
assign_parameter(
parameters,
"0.weight",
torch::tensor({-0.2109, -0.4976, -0.1413, -0.3420, -0.2524, 0.6976}));
torch::tensor({-0.2109, -0.4976, -0.1413, -0.3420, -0.2524, 0.6976}, torch::kFloat64));
assign_parameter(
parameters, "0.bias", torch::tensor({-0.1085, -0.2979, 0.6892}));
parameters, "0.bias", torch::tensor({-0.1085, -0.2979, 0.6892}, torch::kFloat64));
assign_parameter(
parameters, "2.weight", torch::tensor({-0.0508, -0.3941, -0.2843}));
assign_parameter(parameters, "2.bias", torch::tensor({-0.0711}));
parameters, "2.weight", torch::tensor({-0.0508, -0.3941, -0.2843}, torch::kFloat64));
assign_parameter(parameters, "2.bias", torch::tensor({-0.0711}, torch::kFloat64));
auto optimizer = OptimizerClass(parameters.values(), options);
torch::Tensor input =
torch::tensor({0.1, 0.2, 0.3, 0.4, 0.5, 0.6}).reshape({3, 2});
torch::tensor({0.1, 0.2, 0.3, 0.4, 0.5, 0.6}, torch::kFloat64).reshape({3, 2});
for (size_t i = 0; i < kIterations; ++i) {
optimizer.zero_grad();
@@ -116,8 +116,9 @@ void check_exact_values(
expected_parameters.at(i / kSampleEvery).size() == parameters.size());
for (size_t p = 0; p < parameters.size(); ++p) {
ASSERT_TRUE(parameters[p]->defined());
auto computed = parameters[p]->flatten();
auto expected = expected_parameters.at(i / kSampleEvery).at(p);
// Always compare using double dtype, regardless of the original dtype of the tensors
auto computed = parameters[p]->flatten().to(torch::kFloat64);
auto expected = expected_parameters.at(i / kSampleEvery).at(p).to(torch::kFloat64);
if (!computed.allclose(expected, /*rtol=*/1e-3, /*atol=*/5e-4)) {
std::cout << "Iteration " << i << ": " << computed
<< " != " << expected << " (parameter " << p << ")"