Skip to content

C++ APIs Transformer NN Module Top Layer #44333

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
glaringlee wants to merge 7 commits into from
Closed

C++ APIs Transformer NN Module Top Layer #44333

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
af8ee38
C++ APIs Transformer NN Module Top Layer
Sep 8, 2020
9ceeacc
Update on "C++ APIs Transformer NN Module Top Layer"
Sep 9, 2020
51e7de6
Update on "C++ APIs Transformer NN Module Top Layer"
Sep 9, 2020
ab06d3b
Update on "C++ APIs Transformer NN Module Top Layer"
Sep 10, 2020
3912517
Update on "C++ APIs Transformer NN Module Top Layer"
Sep 10, 2020
907d4de
Update on "C++ APIs Transformer NN Module Top Layer"
Sep 11, 2020
fd9f2aa
Update on "C++ APIs Transformer NN Module Top Layer"
Sep 11, 2020
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
197 changes: 149 additions & 48 deletions test/cpp/api/transformer.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,16 @@ using namespace torch::nn;

struct TransformerTest : torch::test::SeedingFixture {};

// a generic function to set constants for parameters so we have fixed result for deterministic test
template<typename Model>
void set_parameter_to_constants(Model& model, const torch::TensorOptions& tensor_options) {
torch::NoGradGuard guard;
for (auto& p : model->parameters()) {
auto sz = p.view(-1).size(0);
p.copy_(torch::cos(torch::arange(0, sz, tensor_options).view(p.sizes())));
}
}

// a generic function to provide consistent encoder/decoder layer for all the transformer tests
template<typename T_LAYER, typename T_OPTIONS>
T_LAYER get_a_test_layer(const torch::TensorOptions& tensor_options) {
Expand All @@ -23,13 +33,7 @@ T_LAYER get_a_test_layer(const torch::TensorOptions& tensor_options) {
}

// set constant weights of the model
{
torch::NoGradGuard guard;
for (auto& p : layer->parameters()) {
auto sz = p.view(-1).size(0);
p.copy_(torch::cos(torch::arange(0, sz, tensor_options).view(p.sizes())));
}
}
set_parameter_to_constants<T_LAYER>(layer, tensor_options);

return layer;
}
Expand Down Expand Up @@ -579,25 +583,26 @@ TEST_F(TransformerTest, PrettyPrintTransformerEncoder) {
")");
}


TEST_F(TransformerTest, PrettyPrintTransformerDecoderLayer) {
ASSERT_EQ(
c10::str(TransformerDecoderLayer(4, 2)),
"torch::nn::TransformerDecoderLayerImpl(\n"
" (self_attn): torch::nn::MultiheadAttention(\n"
" (out_proj): torch::nn::Linear(in_features=4, out_features=4, bias=true)\n"
" )\n"
" (dropout1): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (norm1): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
" (multihead_attn): torch::nn::MultiheadAttention(\n"
" (out_proj): torch::nn::Linear(in_features=4, out_features=4, bias=true)\n"
" )\n"
" (dropout2): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (norm2): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
" (linear1): torch::nn::Linear(in_features=4, out_features=2048, bias=true)\n"
" (dropout): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (linear2): torch::nn::Linear(in_features=2048, out_features=4, bias=true)\n"
" (dropout3): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (norm1): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
" (norm2): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
" (norm3): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
" (dropout1): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (dropout2): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (dropout3): torch::nn::Dropout(p=0.1, inplace=false)\n"
")");
}

Expand Down Expand Up @@ -978,8 +983,7 @@ void transformer_decoder_test_helper(bool is_cuda) {

// Multiple layers with norm
norm = LayerNorm(LayerNormOptions({decoder_layer.get()->options.d_model()}));
model = TransformerDecoder(TransformerDecoderOptions(decoder_layer, 6)
.norm(AnyModule(norm)));
model = TransformerDecoder(TransformerDecoderOptions(decoder_layer, 6).norm(AnyModule(norm)));
if (is_cuda) {
model->to(torch::kCUDA);
}
Expand Down Expand Up @@ -1036,54 +1040,151 @@ TEST_F(TransformerTest, PrettyPrintTransformerDecoder) {
" (layers): torch::nn::ModuleList(\n"
" (0): torch::nn::TransformerDecoderLayerImpl(\n"
" (self_attn): torch::nn::MultiheadAttention(\n"
" (out_proj): torch::nn::Linear(in_features=4, out_features=4,"
" bias=true)\n"
" (out_proj): torch::nn::Linear(in_features=4, out_features=4, bias=true)\n"
" )\n"
" (dropout1): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (norm1): torch::nn::LayerNorm([4], eps=1e-05,"
" elementwise_affine=true)\n"
" (multihead_attn): torch::nn::MultiheadAttention(\n"
" (out_proj): torch::nn::Linear(in_features=4, out_features=4,"
" bias=true)\n"
" (out_proj): torch::nn::Linear(in_features=4, out_features=4, bias=true)\n"
" )\n"
" (dropout2): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (norm2): torch::nn::LayerNorm([4], eps=1e-05,"
" elementwise_affine=true)\n"
" (linear1): torch::nn::Linear(in_features=4, out_features=2048,"
" bias=true)\n"
" (linear1): torch::nn::Linear(in_features=4, out_features=2048, bias=true)\n"
" (dropout): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (linear2): torch::nn::Linear(in_features=2048, out_features=4,"
" bias=true)\n"
" (linear2): torch::nn::Linear(in_features=2048, out_features=4, bias=true)\n"
" (norm1): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
" (norm2): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
" (norm3): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
" (dropout1): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (dropout2): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (dropout3): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (norm3): torch::nn::LayerNorm([4], eps=1e-05,"
" elementwise_affine=true)\n"
" )\n"
" (1): torch::nn::TransformerDecoderLayerImpl(\n"
" (self_attn): torch::nn::MultiheadAttention(\n"
" (out_proj): torch::nn::Linear(in_features=4, out_features=4,"
" bias=true)\n"
" (out_proj): torch::nn::Linear(in_features=4, out_features=4, bias=true)\n"
" )\n"
" (dropout1): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (norm1): torch::nn::LayerNorm([4], eps=1e-05,"
" elementwise_affine=true)\n"
" (multihead_attn): torch::nn::MultiheadAttention(\n"
" (out_proj): torch::nn::Linear(in_features=4, out_features=4,"
" bias=true)\n"
" (out_proj): torch::nn::Linear(in_features=4, out_features=4, bias=true)\n"
" )\n"
" (dropout2): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (norm2): torch::nn::LayerNorm([4], eps=1e-05,"
" elementwise_affine=true)\n"
" (linear1): torch::nn::Linear(in_features=4, out_features=2048,"
" bias=true)\n"
" (linear1): torch::nn::Linear(in_features=4, out_features=2048, bias=true)\n"
" (dropout): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (linear2): torch::nn::Linear(in_features=2048, out_features=4,"
" bias=true)\n"
" (linear2): torch::nn::Linear(in_features=2048, out_features=4, bias=true)\n"
" (norm1): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
" (norm2): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
" (norm3): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
" (dropout1): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (dropout2): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (dropout3): torch::nn::Dropout(p=0.1, inplace=false)\n"
" (norm3): torch::nn::LayerNorm([4], eps=1e-05,"
" elementwise_affine=true)\n"
" )\n"
" )\n"
" (norm): torch::nn::LayerNorm([4], eps=1e-05,"
" elementwise_affine=true)\n"
" (norm): torch::nn::LayerNorm([4], eps=1e-05, elementwise_affine=true)\n"
")");
}

void transformer_test_helper(bool is_cuda) {
// this is a deterministic test for Transformere
torch::Device device = is_cuda ? torch::kCUDA : torch::kCPU;
torch::TensorOptions tensor_options = torch::TensorOptions().dtype(torch::kFloat32).device(device);

// transformer created encoder/decoder
Transformer model(TransformerOptions()
.d_model(4)
.nhead(2)
.num_encoder_layers(2)
.num_decoder_layers(1)
.dim_feedforward(16)
.dropout(0.0)
.activation(torch::kReLU));

set_parameter_to_constants<Transformer>(model, tensor_options);
if (tensor_options.device() == torch::kCUDA) {
model->to(torch::kCUDA);
}

// transformer with customized encoder/decoder
LayerNorm enorm(LayerNormOptions({4}));
TransformerEncoder encoder(TransformerEncoderOptions(
TransformerEncoderLayerOptions(4, 2).dim_feedforward(16).dropout(0.0), 2).norm(AnyModule(enorm)));

LayerNorm dnorm(LayerNormOptions({4}));
TransformerDecoder decoder(TransformerDecoderOptions(
TransformerDecoderLayerOptions(4, 2).dim_feedforward(16).dropout(0.0), 1).norm(AnyModule(dnorm)));

Transformer model_cus(TransformerOptions()
.d_model(4)
.nhead(2)
.custom_encoder(AnyModule(encoder))
.custom_decoder(AnyModule(decoder)));

set_parameter_to_constants<Transformer>(model_cus, tensor_options);
if (tensor_options.device() == torch::kCUDA) {
model_cus->to(torch::kCUDA);
}

// test cases
torch::Tensor src = torch::tensor({
{{1.0, 2.0, 3.0, 4.0}, {5.0, 6.0, 7.0, 8.0}},
{{9.0, 10.0, 11.0, 12.0}, {13.0, 14.0, 15.0, 16.0}},
{{17.0, 18.0, 19.0, 20.0}, {21.0, 22.0, 23.0, 24.0}}}, tensor_options);

torch::Tensor tgt = torch::tensor({
{{1.0, 2.0, 3.0, 4.0}, {5.0, 6.0, 7.0, 8.0}},
{{9.0, 10.0, 11.0, 12.0}, {13.0, 14.0, 15.0, 16.0}}}, tensor_options);

torch::Tensor ref_output = torch::tensor({
{{2.695875, 0.347114, -0.044355, -0.549541}, {2.696091, 0.347015, -0.044770, -0.548522}},
{{2.695875, 0.347114, -0.044355, -0.549541}, {2.696091, 0.347015, -0.044770, -0.548522}}}, tensor_options);
torch::Tensor result = model(src, tgt);
torch::Tensor result_cus = model_cus(src, tgt);
ASSERT_EQ(result.sizes(), ref_output.sizes());
ASSERT_TRUE(result.equal(result_cus));
ASSERT_TRUE(torch::allclose(result, ref_output, 1e-7, 1e-5, /*equal_nan=*/true));

torch::Tensor src_mask = Transformer::Impl::generate_square_subsequent_mask(src.size(0)).to(tensor_options);
ref_output = torch::tensor({
{{2.695875, 0.347114, -0.044355, -0.549541}, {2.696091, 0.347015, -0.044770, -0.548522}},
{{2.695875, 0.347114, -0.044355, -0.549541}, {2.696091, 0.347015, -0.044770, -0.548522}}}, tensor_options);
result = model(src, tgt, src_mask);
result_cus = model_cus(src, tgt, src_mask);
ASSERT_EQ(result.sizes(), ref_output.sizes());
ASSERT_TRUE(result.equal(result_cus));
ASSERT_TRUE(torch::allclose(result, ref_output, 1e-7, 1e-5, /*equal_nan=*/true));

torch::Tensor tgt_key_padding_mask = torch::zeros({tgt.size(1), tgt.size(0)}, tensor_options) == 1;
tgt_key_padding_mask[0][0] = 1;
tgt_key_padding_mask[1][1] = 1;
ref_output = torch::tensor({
{{2.696114, 0.347004, -0.044813, -0.548417}, {2.696091, 0.347015, -0.044770, -0.548522}},
{{2.696114, 0.347004, -0.044813, -0.548417}, {2.696091, 0.347015, -0.044770, -0.548522}}}, tensor_options);
result = model(src, tgt, src_mask, torch::Tensor(), torch::Tensor(), torch::Tensor(), tgt_key_padding_mask);
result_cus = model_cus(src, tgt, src_mask, torch::Tensor(), torch::Tensor(), torch::Tensor(), tgt_key_padding_mask);
ASSERT_EQ(result.sizes(), ref_output.sizes());
ASSERT_TRUE(result.equal(result_cus));
ASSERT_TRUE(torch::allclose(result, ref_output, 1e-7, 1e-5, /*equal_nan=*/true));
}
Copy link
Contributor

@zou3519 zou3519 Sep 10, 2020

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'd also test the error conditions (e.g., src and target don't have the same size in the batch dim) with an ASSERT_RAISES to check that it doesn't crash in an unfortunate way

Copy link
Contributor Author

@glaringlee glaringlee Sep 10, 2020

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

will add these tests


TEST_F(TransformerTest, Transformer) {
transformer_test_helper(false);
}

TEST_F(TransformerTest, Transformer_CUDA) {
transformer_test_helper(true);
}

TEST_F(TransformerTest, TransformerArgsCorrectness) {
Transformer model(TransformerOptions()
.d_model(4)
.nhead(2)
.num_encoder_layers(2)
.num_decoder_layers(1)
.dim_feedforward(16)
.dropout(0.0)
.activation(torch::kReLU));

torch::Tensor src = torch::randn({2, 3, 4});
torch::Tensor tgt = torch::randn({3, 2, 4});

ASSERT_THROWS_WITH(model(src, tgt), "src and tgt should have equal batch size");

tgt = torch::randn({2, 3, 3});
ASSERT_THROWS_WITH(model(src, tgt), "src and tgt should have same feature size as d_model");

src = torch::randn({2, 3});
ASSERT_THROWS_WITH(model(src, tgt), "src and tgt should have 3 dimensions");
}
1 change: 1 addition & 0 deletions torch/csrc/api/include/torch/nn/modules.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -32,3 +32,4 @@
#include <torch/nn/modules/normalization.h>
#include <torch/nn/modules/transformerlayer.h>
#include <torch/nn/modules/transformercoder.h>
#include <torch/nn/modules/transformer.h>
Loading