m53297601
/
Paddle
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into fix_infershape_bug2

Browse Source
shanyi15-patch-1
xuezhong 6 years ago
parent 41740519df 5d45eb06f9
commit 742d758747
50 changed files with 1112 additions and 417 deletions
Show Stats Download Patch File Download Diff File

12
paddle/fluid/API.spec
Unescape View File

@ -277,6 +277,7 @@ paddle.fluid.layers.has_nan (ArgSpec(args=['x'], varargs=None, keywords=None, de
paddle.fluid.layers.isfinite (ArgSpec(args=['x'], varargs=None, keywords=None, defaults=None), ('document', '0a437011c3906079fd8947ed3e52d292'))
paddle.fluid.layers.range (ArgSpec(args=['start', 'end', 'step', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '2ec937ede953ded2fdff2675883900bb'))
paddle.fluid.layers.linspace (ArgSpec(args=['start', 'stop', 'num', 'dtype'], varargs=None, keywords=None, defaults=None), ('document', '495e21e9a848c2d075a102802fc67756'))
paddle.fluid.layers.zeros_like (ArgSpec(args=['x', 'out'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c7e4cfffc93ae89c8f6f53b6d650f923'))
paddle.fluid.layers.While.__init__ (ArgSpec(args=['self', 'cond', 'is_test', 'name'], varargs=None, keywords=None, defaults=(False, None)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.layers.While.block (ArgSpec(args=['self'], varargs=None, keywords=None, defaults=None), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
paddle.fluid.layers.Switch.__init__ (ArgSpec(args=['self', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@ -286,7 +287,11 @@ paddle.fluid.layers.increment (ArgSpec(args=['x', 'value', 'in_place'], varargs=
paddle.fluid.layers.array_write (ArgSpec(args=['x', 'i', 'array'], varargs=None, keywords=None, defaults=(None,)), ('document', '40b6d15f4c86b2b09df340d7778ad713'))
paddle.fluid.layers.create_array (ArgSpec(args=['dtype'], varargs=None, keywords=None, defaults=None), ('document', '2d4f20087080ba5105b55205ad5c5b6a'))
paddle.fluid.layers.less_than (ArgSpec(args=['x', 'y', 'force_cpu', 'cond'], varargs=None, keywords=None, defaults=(None, None)), ('document', '067bbc799c66289ca8b8924c26b6673f'))
paddle.fluid.layers.less_equal (ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', 'd6b173ae1a149e0bdfe7b8bf69285957'))
paddle.fluid.layers.greater_than (ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', '2c9bd414caa6c615539018d27001b44c'))
paddle.fluid.layers.greater_equal (ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', '62c667d24e7b07e166b47a53b61b2ff4'))
paddle.fluid.layers.equal (ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', '80c29b1dc64718f0116de90d1ac88a77'))
paddle.fluid.layers.not_equal (ArgSpec(args=['x', 'y', 'cond'], varargs=None, keywords=None, defaults=(None,)), ('document', '56148fb1024687a08e96af79bdc5c929'))
paddle.fluid.layers.array_read (ArgSpec(args=['array', 'i'], varargs=None, keywords=None, defaults=None), ('document', 'dd68bead34dfbaf6b0a163fc1cc3c385'))
paddle.fluid.layers.array_length (ArgSpec(args=['array'], varargs=None, keywords=None, defaults=None), ('document', 'ffb8b9578ec66db565b223d313aa82a2'))
paddle.fluid.layers.IfElse.__init__ (ArgSpec(args=['self', 'cond', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '6adf97f83acf6453d4a6a4b1070f3754'))
@ -319,6 +324,7 @@ paddle.fluid.layers.atan (ArgSpec(args=['x', 'name'], varargs=None, keywords=Non
paddle.fluid.layers.tanh_shrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '1e521554b9fdda9061ec6d306f0709b7'))
paddle.fluid.layers.softshrink (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '9eef31597bbafa2bd49691e072296e13'))
paddle.fluid.layers.sqrt (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'e9e27491c39ac74d0b1ffe506aec0ebb'))
paddle.fluid.layers.rsqrt (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c445467ebe58b3c0d7f0bba7795b6f56'))
paddle.fluid.layers.abs (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '64650ac42cf82e9920cb0b172b1d29fd'))
paddle.fluid.layers.ceil (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c75d67dc5fe28f68e4cfffead4f698ad'))
paddle.fluid.layers.floor (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '647b16c5da5ef909649ae02abb434973'))
@ -331,13 +337,13 @@ paddle.fluid.layers.reciprocal (ArgSpec(args=['x', 'name'], varargs=None, keywor
paddle.fluid.layers.square (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '48dfb45d773dbc30126c3a7f777de5ee'))
paddle.fluid.layers.softplus (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '459c5781e9d1dd88283b7c5769d7872a'))
paddle.fluid.layers.softsign (ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '80846bcd4bd457207457a6d5411f4148'))
paddle.fluid.layers.uniform_random (ArgSpec(args=['shape', 'dtype', 'min', 'max', 'seed'], varargs=None, keywords=None, defaults=('float32', -1.0, 1.0, 0)), ('document', '308b619af849caa82bbc31e897f5e641'))
paddle.fluid.layers.uniform_random (ArgSpec(args=['shape', 'dtype', 'min', 'max', 'seed'], varargs=None, keywords=None, defaults=('float32', -1.0, 1.0, 0)), ('document', 'a8c4e972b7d6742c838a37abf407ed9a'))
paddle.fluid.layers.hard_shrink (ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,)), ('document', 'c142f5884f3255e0d6075c286bbd531e'))
paddle.fluid.layers.cumsum (ArgSpec(args=['x', 'axis', 'exclusive', 'reverse'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '944d7c03057f5fc88bc78acd4d82f926'))
paddle.fluid.layers.thresholded_relu (ArgSpec(args=['x', 'threshold'], varargs=None, keywords=None, defaults=(None,)), ('document', '90566ea449ea4c681435546e2f70610a'))
paddle.fluid.layers.prior_box (ArgSpec(args=['input', 'image', 'min_sizes', 'max_sizes', 'aspect_ratios', 'variance', 'flip', 'clip', 'steps', 'offset', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, [1.0], [0.1, 0.1, 0.2, 0.2], False, False, [0.0, 0.0], 0.5, None, False)), ('document', '14cac0ee643fa6e026ad82aeeee75bd8'))
paddle.fluid.layers.density_prior_box (ArgSpec(args=['input', 'image', 'densities', 'fixed_sizes', 'fixed_ratios', 'variance', 'clip', 'steps', 'offset', 'flatten_to_2d', 'name'], varargs=None, keywords=None, defaults=(None, None, None, [0.1, 0.1, 0.2, 0.2], False, [0.0, 0.0], 0.5, False, None)), ('document', 'a0d762bb08de9ce93bc780aa57cd5cd9'))
paddle.fluid.layers.multi_box_head (ArgSpec(args=['inputs', 'image', 'base_size', 'num_classes', 'aspect_ratios', 'min_ratio', 'max_ratio', 'min_sizes', 'max_sizes', 'steps', 'step_w', 'step_h', 'offset', 'variance', 'flip', 'clip', 'kernel_size', 'pad', 'stride', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, 0.5, [0.1, 0.1, 0.2, 0.2], True, False, 1, 0, 1, None, False)), ('document', 'a6ab47a2fe681e52fabb7057ddf0efdd'))
paddle.fluid.layers.multi_box_head (ArgSpec(args=['inputs', 'image', 'base_size', 'num_classes', 'aspect_ratios', 'min_ratio', 'max_ratio', 'min_sizes', 'max_sizes', 'steps', 'step_w', 'step_h', 'offset', 'variance', 'flip', 'clip', 'kernel_size', 'pad', 'stride', 'name', 'min_max_aspect_ratios_order'], varargs=None, keywords=None, defaults=(None, None, None, None, None, None, None, 0.5, [0.1, 0.1, 0.2, 0.2], True, False, 1, 0, 1, None, False)), ('document', 'fe9afaee481dd09f28866df22756466f'))
paddle.fluid.layers.bipartite_match (ArgSpec(args=['dist_matrix', 'match_type', 'dist_threshold', 'name'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', '3ddb9b966f193900193a95a3df77c3c1'))
paddle.fluid.layers.target_assign (ArgSpec(args=['input', 'matched_indices', 'negative_indices', 'mismatch_value', 'name'], varargs=None, keywords=None, defaults=(None, None, None)), ('document', 'c0b334f917828f95056f6ebe10907b1c'))
paddle.fluid.layers.detection_output (ArgSpec(args=['loc', 'scores', 'prior_box', 'prior_box_var', 'background_label', 'nms_threshold', 'nms_top_k', 'keep_top_k', 'score_threshold', 'nms_eta'], varargs=None, keywords=None, defaults=(0, 0.3, 400, 200, 0.01, 1.0)), ('document', 'c33093a82a46e3091e789e5572588db1'))
@ -352,7 +358,7 @@ paddle.fluid.layers.generate_mask_labels (ArgSpec(args=['im_info', 'gt_classes',
paddle.fluid.layers.iou_similarity (ArgSpec(args=['x', 'y', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '587845f60c5d97ffdf2dfd21da52eca1'))
paddle.fluid.layers.box_coder (ArgSpec(args=['prior_box', 'prior_box_var', 'target_box', 'code_type', 'box_normalized', 'name', 'axis'], varargs=None, keywords=None, defaults=('encode_center_size', True, None, 0)), ('document', '032d0f4b7d8f6235ee5d91e473344f0e'))
paddle.fluid.layers.polygon_box_transform (ArgSpec(args=['input', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '0e5ac2507723a0b5adec473f9556799b'))
paddle.fluid.layers.yolov3_loss (ArgSpec(args=['x', 'gtbox', 'gtlabel', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'gtscore', 'use_label_smooth', 'name'], varargs=None, keywords=None, defaults=(None, True, None)), ('document', '57fa96922e42db8f064c3fb77f2255e8'))
paddle.fluid.layers.yolov3_loss (ArgSpec(args=['x', 'gt_box', 'gt_label', 'anchors', 'anchor_mask', 'class_num', 'ignore_thresh', 'downsample_ratio', 'gt_score', 'use_label_smooth', 'name'], varargs=None, keywords=None, defaults=(None, True, None)), ('document', '059021025283ad1ee6f4d32228cf3e4e'))
paddle.fluid.layers.yolo_box (ArgSpec(args=['x', 'img_size', 'anchors', 'class_num', 'conf_thresh', 'downsample_ratio', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '5566169a5ab993d177792c023c7fb340'))
paddle.fluid.layers.box_clip (ArgSpec(args=['input', 'im_info', 'name'], varargs=None, keywords=None, defaults=(None,)), ('document', '397e9e02b451d99c56e20f268fa03f2e'))
paddle.fluid.layers.multiclass_nms (ArgSpec(args=['bboxes', 'scores', 'score_threshold', 'nms_top_k', 'keep_top_k', 'nms_threshold', 'normalized', 'nms_eta', 'background_label', 'name'], varargs=None, keywords=None, defaults=(0.3, True, 1.0, 0, None)), ('document', 'ca7d1107b6c5d2d6d8221039a220fde0'))

41
paddle/fluid/inference/tests/api/trt_models_tester.cc
Unescape View File

@ -116,7 +116,7 @@ void compare_continuous_input(std::string model_dir, bool use_tensorrt) {
reinterpret_cast<const PaddlePredictor::Config*>(&analysis_config);
auto native_pred = CreateTestPredictor(config, false);
auto analysis_pred = CreateTestPredictor(config, true);
for (int i = 0; i < 100; i++) {
for (int i = 0; i < 20; i++) {
std::vector<std::vector<PaddleTensor>> inputs_all;
if (!FLAGS_prog_filename.empty() && !FLAGS_param_filename.empty()) {
SetFakeImageInput(&inputs_all, model_dir, true, FLAGS_prog_filename,
@ -133,11 +133,13 @@ void compare_continuous_input(std::string model_dir, bool use_tensorrt) {
TEST(TensorRT_mobilenet, compare) {
std::string model_dir = FLAGS_infer_model + "/mobilenet";
compare(model_dir, /* use_tensorrt */ true);
// Open it when need.
// profile(model_dir, /* use_analysis */ true, FLAGS_use_tensorrt);
}
TEST(TensorRT_resnet50, compare) {
TEST(resnet50, compare_continuous_input) {
std::string model_dir = FLAGS_infer_model + "/resnet50";
compare(model_dir, /* use_tensorrt */ true);
compare_continuous_input(model_dir, true);
}
TEST(TensorRT_resnext50, compare) {
@ -145,24 +147,6 @@ TEST(TensorRT_resnext50, compare) {
compare(model_dir, /* use_tensorrt */ true);
}
TEST(TensorRT_resnext50, profile) {
std::string model_dir = FLAGS_infer_model + "/resnext50";
// Set FLAGS_record_benchmark to true to record benchmark to file.
// FLAGS_record_benchmark=true;
FLAGS_model_name = "resnext50";
profile(model_dir, /* use_analysis */ true, FLAGS_use_tensorrt);
}
TEST(resnext50, compare_analysis_native) {
std::string model_dir = FLAGS_infer_model + "/resnext50";
compare(model_dir, false /*use tensorrt*/);
}
TEST(TensorRT_mobilenet, analysis) {
std::string model_dir = FLAGS_infer_model + "/" + "mobilenet";
compare(model_dir, false /* use_tensorrt */);
}
TEST(AnalysisPredictor, use_gpu) {
std::string model_dir = FLAGS_infer_model + "/" + "mobilenet";
AnalysisConfig config;
@ -180,20 +164,5 @@ TEST(AnalysisPredictor, use_gpu) {
}
}
TEST(TensorRT_mobilenet, profile) {
std::string model_dir = FLAGS_infer_model + "/" + "mobilenet";
profile(model_dir, true, false);
}
TEST(resnet50, compare_continuous_input) {
std::string model_dir = FLAGS_infer_model + "/resnet50";
compare_continuous_input(model_dir, true);
}
TEST(resnet50, compare_continuous_input_native) {
std::string model_dir = FLAGS_infer_model + "/resnet50";
compare_continuous_input(model_dir, false);
}
} // namespace inference
} // namespace paddle

1
paddle/fluid/op_use_default_grad_op_maker.spec
Unescape View File

@ -18,7 +18,6 @@ gru
hierarchical_sigmoid
lrn
lstm_unit
lstmp
max_pool2d_with_index
max_pool3d_with_index
maxout

11
paddle/fluid/operators/activation_op.cc
Unescape View File

@ -227,6 +227,15 @@ $out = \sqrt{x}$
)DOC";
UNUSED constexpr char RsqrtDoc[] = R"DOC(
Rsqrt Activation Operator.
Please make sure input is legal in case of numeric errors.
$out = \frac{1}{\sqrt{x}}$
)DOC";
UNUSED constexpr char AbsDoc[] = R"DOC(
Abs Activation Operator.
@ -575,6 +584,7 @@ REGISTER_ACTIVATION_OP_MAKER(Gelu, GeluDoc);
REGISTER_ACTIVATION_OP_MAKER(Tanh, TanhDoc);
REGISTER_ACTIVATION_OP_MAKER(TanhShrink, TanhShrinkDoc);
REGISTER_ACTIVATION_OP_MAKER(Sqrt, SqrtDoc);
REGISTER_ACTIVATION_OP_MAKER(Rsqrt, RsqrtDoc);
REGISTER_ACTIVATION_OP_MAKER(Abs, AbsDoc);
REGISTER_ACTIVATION_OP_MAKER(Ceil, CeilDoc);
REGISTER_ACTIVATION_OP_MAKER(Floor, FloorDoc);
@ -586,6 +596,7 @@ REGISTER_ACTIVATION_OP_MAKER(Log, LogDoc);
REGISTER_ACTIVATION_OP_MAKER(Square, SquareDoc);
REGISTER_ACTIVATION_OP_MAKER(Softplus, SoftplusDoc);
REGISTER_ACTIVATION_OP_MAKER(Softsign, SoftsignDoc);
} // namespace operators
} // namespace paddle

21
paddle/fluid/operators/activation_op.h
Unescape View File

@ -511,6 +511,26 @@ struct SqrtGradFunctor : public BaseActivationFunctor<T> {
static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
};
// rsqrt(x) = x^(-1/2)
template <typename T>
struct RsqrtFunctor : public BaseActivationFunctor<T> {
template <typename Device, typename X, typename Out>
void operator()(Device d, X x, Out out) const {
out.device(d) = x.rsqrt();
}
};
template <typename T>
struct RsqrtGradFunctor : public BaseActivationFunctor<T> {
template <typename Device, typename X, typename Out, typename dOut,
typename dX>
void operator()(Device d, X x, Out out, dOut dout, dX dx) const {
dx.device(d) = static_cast<T>(-0.5) * dout * out * out * out;
}
static constexpr ActBwdOpFwdDeps FwdDeps() { return kDepOut; }
};
// ceil(x) = ceiling(x)
template <typename T>
struct CeilFunctor : public BaseActivationFunctor<T> {
@ -1191,6 +1211,7 @@ struct SwishGradFunctor : public BaseActivationFunctor<T> {
__macro(atan, Atan, AtanFunctor, AtanGradFunctor); \
__macro(softshrink, SoftShrink, SoftShrinkFunctor, SoftShrinkGradFunctor); \
__macro(sqrt, Sqrt, SqrtFunctor, SqrtGradFunctor); \
__macro(rsqrt, Rsqrt, RsqrtFunctor, RsqrtGradFunctor); \
__macro(abs, Abs, AbsFunctor, AbsGradFunctor); \
__macro(ceil, Ceil, CeilFunctor, ZeroGradFunctor); \
__macro(floor, Floor, FloorFunctor, ZeroGradFunctor); \

8
paddle/fluid/operators/affine_channel_op.cc
Unescape View File

@ -79,9 +79,13 @@ class AffineChannelOp : public framework::OperatorWithKernel {
: x_dims[x_dims.size() - 1]);
PADDLE_ENFORCE_EQ(scale_dims.size(), 1UL);
PADDLE_ENFORCE_EQ(scale_dims[0], C);
PADDLE_ENFORCE_EQ(b_dims.size(), 1UL);
PADDLE_ENFORCE_EQ(b_dims[0], C);
if (ctx->IsRuntime() || scale_dims[0] > 0) {
PADDLE_ENFORCE_EQ(scale_dims[0], C);
}
if (ctx->IsRuntime() || b_dims[0] > 0) {
PADDLE_ENFORCE_EQ(b_dims[0], C);
}
ctx->SetOutputDim("Out", ctx->GetInputDim("X"));
ctx->ShareLoD("X", "Out");

19
paddle/fluid/operators/batch_norm_op.cc
Unescape View File

@ -65,11 +65,22 @@ void BatchNormOp::InferShape(framework::InferShapeContext *ctx) const {
(data_layout == DataLayout::kNCHW ? x_dims[1]
: x_dims[x_dims.size() - 1]);
PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale").size(), 1UL);
PADDLE_ENFORCE_EQ(ctx->GetInputDim("Scale")[0], C);
PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias").size(), 1UL);
PADDLE_ENFORCE_EQ(ctx->GetInputDim("Bias")[0], C);
auto scale_dim = ctx->GetInputDim("Scale");
auto bias_dim = ctx->GetInputDim("Bias");
PADDLE_ENFORCE_EQ(scale_dim.size(), 1UL);
PADDLE_ENFORCE_EQ(scale_dim.size(), 1UL);
bool check = true;
if ((!ctx->IsRuntime()) && (framework::product(scale_dim) <= 0 ||
framework::product(bias_dim) <= 0)) {
check = false;
}
if (check) {
PADDLE_ENFORCE_EQ(scale_dim[0], C);
PADDLE_ENFORCE_EQ(scale_dim[0], C);
}
ctx->SetOutputDim("Y", x_dims);
ctx->SetOutputDim("MeanOut", {C});
ctx->SetOutputDim("VarianceOut", {C});

10
paddle/fluid/operators/concat_op.cc
Unescape View File

@ -49,7 +49,15 @@ class ConcatOp : public framework::OperatorWithKernel {
for (size_t i = 1; i < n; i++) {
for (size_t j = 0; j < in_zero_dims_size; j++) {
if (j == axis) {
out_dims[axis] += ins[i][j];
if (ctx->IsRuntime()) {
out_dims[axis] += ins[i][j];
} else {
if (ins[i][j] == -1) {
out_dims[axis] = -1;
} else {
out_dims[axis] += ins[i][j];
}
}
} else {
if (ctx->IsRuntime()) {
// check all shape in run time

11
paddle/fluid/operators/conv_op.cc
Unescape View File

@ -68,9 +68,14 @@ void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
for (size_t i = 0; i < strides.size(); ++i) {
output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
dilations[i], paddings[i],
strides[i]));
if ((!ctx->IsRuntime()) &&
(in_dims[i + 2] <= 0 || filter_dims[i + 2] <= 0)) {
output_shape.push_back(-1);
} else {
output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
dilations[i], paddings[i],
strides[i]));
}
}
ctx->SetOutputDim("Output", framework::make_ddim(output_shape));
ctx->ShareLoD("Input", "Output");

19
paddle/fluid/operators/conv_shift_op.cc
Unescape View File

@ -36,14 +36,17 @@ class ConvShiftOp : public framework::OperatorWithKernel {
auto y_dims = ctx->GetInputDim("Y");
PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
PADDLE_ENFORCE_EQ(y_dims.size(), 2, "Input(Y)'s rank should be 2.");
PADDLE_ENFORCE_EQ(x_dims[0], y_dims[0],
"The 1st dimension of Input(X) and Input(Y) should "
"be equal.");
PADDLE_ENFORCE_EQ(y_dims[1] % 2, 1,
"The 2nd dimension of Input(Y) should be odd.");
PADDLE_ENFORCE_LE(y_dims[1], x_dims[1],
"The 2nd dimension of Input(Y) should be less than or "
"equal to the 2nd dimension of Input(X).");
if (ctx->IsRuntime() || (x_dims[0] > 0 && y_dims[0] > 0))
PADDLE_ENFORCE_EQ(x_dims[0], y_dims[0],
"The 1st dimension of Input(X) and Input(Y) should "
"be equal.");
if (ctx->IsRuntime() || y_dims[1] > 0)
PADDLE_ENFORCE_EQ(y_dims[1] % 2, 1,
"The 2nd dimension of Input(Y) should be odd.");
if (ctx->IsRuntime() || (x_dims[1] > 0 && y_dims[1] > 0))
PADDLE_ENFORCE_LE(y_dims[1], x_dims[1],
"The 2nd dimension of Input(Y) should be less than or "
"equal to the 2nd dimension of Input(X).");
ctx->ShareDim("X", /*->*/ "Out");
ctx->ShareLoD("X", /*->*/ "Out");
}

32
paddle/fluid/operators/cos_sim_op.cc
Unescape View File

@ -40,17 +40,27 @@ class CosSimOp : public framework::OperatorWithKernel {
auto x_dims = ctx->GetInputDim("X");
auto y_dims = ctx->GetInputDim("Y");
PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(),
"Ranks of Input(X) and Input(Y) must be equal.");
PADDLE_ENFORCE_GE(x_dims.size(), 2,
"Rank of Input(X) must not be less than 2.");
PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 1, x_dims.size()),
framework::slice_ddim(y_dims, 1, y_dims.size()),
"All dimensions except the 1st of Input(X) and Input(Y) "
"must be equal.");
PADDLE_ENFORCE(x_dims[0] == y_dims[0] || y_dims[0] == 1,
"The 1st dimension of Input(Y) must be equal to Input(X) or"
" just 1 (which will be broadcasted to match Input(X)).");
bool check = true;
if ((!ctx->IsRuntime()) &&
(framework::product(x_dims) <= 0 || framework::product(y_dims) <= 0)) {
check = false;
}
if (check) {
PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(),
"Ranks of Input(X) and Input(Y) must be equal.");
PADDLE_ENFORCE_GE(x_dims.size(), 2,
"Rank of Input(X) must not be less than 2.");
PADDLE_ENFORCE_EQ(
framework::slice_ddim(x_dims, 1, x_dims.size()),
framework::slice_ddim(y_dims, 1, y_dims.size()),
"All dimensions except the 1st of Input(X) and Input(Y) "
"must be equal.");
PADDLE_ENFORCE(
x_dims[0] == y_dims[0] || y_dims[0] == 1,
"The 1st dimension of Input(Y) must be equal to Input(X) or"
" just 1 (which will be broadcasted to match Input(X)).");
}
// resize tensor
ctx->SetOutputDim("Out", {x_dims[0], 1});

6
paddle/fluid/operators/detection_map_op.cc
Unescape View File

@ -51,8 +51,10 @@ class DetectionMAPOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE_EQ(label_dims.size(), 2,
"The rank of Input(Label) must be 2, "
"the shape is [N, 6].");
PADDLE_ENFORCE(label_dims[1] == 6 || label_dims[1] == 5,
"The shape of Input(Label) is [N, 6] or [N, 5].");
if (ctx->IsRuntime() || label_dims[1] > 0) {
PADDLE_ENFORCE(label_dims[1] == 6 || label_dims[1] == 5,
"The shape of Input(Label) is [N, 6] or [N, 5].");
}
if (ctx->HasInput("PosCount")) {
PADDLE_ENFORCE(ctx->HasInput("TruePos"),

8
paddle/fluid/operators/distributed_ops/split_byref_op.cc
Unescape View File

@ -31,14 +31,16 @@ class SplitByrefOp : public framework::OperatorWithKernel {
auto in_dims = ctx->GetInputDim("X");
auto outs_names = ctx->Outputs("Out");
size_t num = static_cast<size_t>(ctx->Attrs().Get<int>("num"));
std::vector<int> sections = static_cast<std::vector<int>>(
ctx->Attrs().Get<std::vector<int>>("sections"));
auto sections = ctx->Attrs().Get<std::vector<int>>("sections");
const size_t outs_number = outs_names.size();
std::vector<framework::DDim> outs_dims;
outs_dims.reserve(outs_number);
if (num > 0) {
int64_t in_axis_dim = in_dims[0];
int64_t in_axis_dim = 0;
if (ctx->IsRuntime()) {
in_axis_dim = in_dims[0];
}
PADDLE_ENFORCE_EQ(in_axis_dim % num, 0,
"tensor split does not result"
" in an equal division");

9
paddle/fluid/operators/grid_sampler_op.cc
Unescape View File

@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/grid_sampler_op.h"
#include <memory>
#include "paddle/fluid/framework/op_registry.h"
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cudnn_helper.h"
@ -40,10 +41,12 @@ class GridSampleOp : public framework::OperatorWithKernel {
"Input(X) of GridSampleOp should be 4-D Tensor.");
PADDLE_ENFORCE(grid_dims.size() == 4,
"Input(Grid) of GridSampleOp should be 4-D Tensor.");
PADDLE_ENFORCE(grid_dims[3] == 2, "Input(Grid) dims[3] should be 2.");
PADDLE_ENFORCE_EQ(grid_dims[0], x_dims[0],
"Input(X) and Input(Grid) dims[0] should be equal.");
if (ctx->IsRuntime() || grid_dims[3] > 0) {
PADDLE_ENFORCE(grid_dims[3] == 2, "Input(Grid) dims[3] should be 2.");
}
if (ctx->IsRuntime()) {
PADDLE_ENFORCE_EQ(grid_dims[0], x_dims[0],
"Input(X) and Input(Grid) dims[0] should be equal.");
PADDLE_ENFORCE_EQ(
grid_dims[1], x_dims[2],
"Input(X) dims[2] and Input(Grid) dims[1] should be equal.");

2
paddle/fluid/operators/hierarchical_sigmoid_op.h
Unescape View File

@ -238,6 +238,8 @@ class HierarchicalSigmoidGradOpKernel : public framework::OpKernel<T> {
zero(dev_ctx, w_grad, static_cast<T>(0.0));
bit_code->MulGradWeight(pre_out_grad, w_grad, in);
} else {
PADDLE_ENFORCE(path != nullptr,
"Sparse mode should not be used without custom tree!");
framework::Vector<int64_t> real_rows = PathToRows(*path);
auto* w_grad =
ctx.Output<framework::SelectedRows>(framework::GradVarName("W"));

6
paddle/fluid/operators/interpolate_op.cc
Unescape View File

@ -45,9 +45,14 @@ class InterpolateOp : public framework::OperatorWithKernel {
// round down
out_h = static_cast<int>(dim_x[2] * scale);
out_w = static_cast<int>(dim_x[3] * scale);
// protect when input shape is -1
out_h = out_h > 0 ? out_h : -1;
out_w = out_w > 0 ? out_w : -1;
} else {
out_h = ctx->Attrs().Get<int>("out_h");
out_w = ctx->Attrs().Get<int>("out_w");
PADDLE_ENFORCE_GT(out_h, 0, "out_h should be greater than 0.");
PADDLE_ENFORCE_GT(out_w, 0, "out_w should be greater than 0.");
}
if (ctx->HasInput("OutSize") && ctx->IsRuntime()) {
@ -58,6 +63,7 @@ class InterpolateOp : public framework::OperatorWithKernel {
ctx->ShareLoD("X", "Out");
return;
}
std::vector<int64_t> dim_out({dim_x[0], dim_x[1], out_h, out_w});
ctx->SetOutputDim("Out", framework::make_ddim(dim_out));
}

6
paddle/fluid/operators/kldiv_loss_op.cc
Unescape View File

@ -35,8 +35,10 @@ class KLDivLossOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE_EQ(dim_x.size(), dim_target.size(),
"Input(X) rank and Input(Target) rank should be same.");
for (int i = 0; i < dim_x.size(); i++) {
PADDLE_ENFORCE_EQ(dim_x[i], dim_target[i],
"Input(X) and Input(Target) should in same shape.");
if (ctx->IsRuntime() || (dim_x[i] > 0 && dim_target[i] > 0)) {
PADDLE_ENFORCE_EQ(dim_x[i], dim_target[i],
"Input(X) and Input(Target) should in same shape.");
}
}
auto reduction = ctx->Attrs().Get<std::string>("reduction");

49
paddle/fluid/operators/lstmp_op.cc
Unescape View File

@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/lstmp_op.h"
#include <memory>
#include <string>
namespace paddle {
@ -45,6 +46,7 @@ class LSTMPOp : public framework::OperatorWithKernel {
"Output(BatchHidden) of LSTMP operator should not be null.");
auto in_dims = ctx->GetInputDim("Input");
PADDLE_ENFORCE_EQ(in_dims.size(), 2,
"Input(X)'s rank of LSTMP operator must be 2.");
@ -269,13 +271,47 @@ Users can choose to use fully-connected operator before LSTMP operator.
}
};
class LSTMPGradMaker : public framework::SingleGradOpDescMaker {
public:
using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
protected:
std::unique_ptr<framework::OpDesc> Apply() const override {
auto* grad_op = new framework::OpDesc();
grad_op->SetType("lstmp_grad");
grad_op->SetInput("Weight", Input("Weight"));
grad_op->SetInput("ProjWeight", Input("ProjWeight"));
grad_op->SetInput("Bias", Input("Bias"));
grad_op->SetInput("Projection", Output("Projection"));
grad_op->SetInput("Cell", Output("Cell"));
grad_op->SetInput("BatchGate", Output("BatchGate"));
grad_op->SetInput("BatchCellPreAct", Output("BatchCellPreAct"));
grad_op->SetInput("BatchHidden", Output("BatchHidden"));
grad_op->SetInput("H0", Input("H0"));
grad_op->SetInput("C0", Input("C0"));
grad_op->SetInput(framework::GradVarName("Projection"),
OutputGrad("Projection"));
grad_op->SetOutput(framework::GradVarName("Input"), InputGrad("Input"));
grad_op->SetOutput(framework::GradVarName("Weight"), InputGrad("Weight"));
grad_op->SetOutput(framework::GradVarName("ProjWeight"),
InputGrad("ProjWeight"));
grad_op->SetOutput(framework::GradVarName("Bias"), InputGrad("Bias"));
grad_op->SetOutput(framework::GradVarName("H0"), InputGrad("H0"));
grad_op->SetOutput(framework::GradVarName("C0"), InputGrad("C0"));
grad_op->SetAttrMap(Attrs());
return std::unique_ptr<framework::OpDesc>(grad_op);
}
};
class LSTMPGradOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("Input"),
"Input(Input) of LSTMP operator should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Projection"),
"Input(Projection) of LSTMP operator should not be null.");
PADDLE_ENFORCE(ctx->HasInput("Cell"),
@ -298,7 +334,8 @@ class LSTMPGradOp : public framework::OperatorWithKernel {
ctx->SetOutputDim(g_name, ctx->GetInputDim(name));
};
SetOutGradDim("Input");
ctx->SetOutputDim(framework::GradVarName("Input"),
ctx->GetInputDim("BatchGate"));
SetOutGradDim("Weight");
SetOutGradDim("ProjWeight");
SetOutGradDim("Bias");
@ -310,7 +347,8 @@ class LSTMPGradOp : public framework::OperatorWithKernel {
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
ctx.Input<framework::LoDTensor>("Input")->type(), ctx.device_context());
ctx.Input<framework::LoDTensor>("BatchGate")->type(),
ctx.device_context());
}
};
@ -318,8 +356,7 @@ class LSTMPGradOp : public framework::OperatorWithKernel {
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(lstmp, ops::LSTMPOp, ops::LSTMPOpMaker,
paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(lstmp, ops::LSTMPOp, ops::LSTMPOpMaker, ops::LSTMPGradMaker);
REGISTER_OPERATOR(lstmp_grad, ops::LSTMPGradOp);
REGISTER_OP_CPU_KERNEL(
lstmp, ops::LSTMPKernel<paddle::platform::CPUDeviceContext, float>,

4
paddle/fluid/operators/lstmp_op.h
Unescape View File

@ -267,7 +267,6 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
}
void Compute(const framework::ExecutionContext& ctx) const override {
auto* input = ctx.Input<LoDTensor>("Input");
auto* weight = ctx.Input<Tensor>("Weight");
auto* proj_weight = ctx.Input<Tensor>("ProjWeight");
auto* bias = ctx.Input<Tensor>("Bias");
@ -323,7 +322,8 @@ class LSTMPGradKernel : public framework::OpKernel<T> {
ordered_c0_g.mutable_data<T>(c0_g->dims(), ctx.GetPlace());
}
auto in_dims = input->dims();
// batch_gate dims equal to input dims
auto in_dims = batch_gate->dims();
auto out_dims = cell_out->dims();
framework::DDim proj_dims({in_dims[0], proj_weight->dims()[1]});
int frame_size = static_cast<int>(in_dims[1] / 4);

4
paddle/fluid/operators/merge_lod_tensor_op.cc
Unescape View File

@ -164,7 +164,9 @@ class MergeLoDTensorInferShape : public framework::InferShapeBase {
auto mask_dim = context->GetInputDim("Mask");
PADDLE_ENFORCE_EQ(mask_dim.size(), 2);
PADDLE_ENFORCE_EQ(mask_dim[1], 1);
if (context->IsRuntime() || mask_dim[1] > 0) {
PADDLE_ENFORCE_EQ(mask_dim[1], 1);
}
context->SetOutputDim("Out", context->GetInputDim("InTrue"));
}

40
paddle/fluid/operators/mkldnn/batch_norm_mkldnn_op.cc
Unescape View File

@ -39,13 +39,9 @@ struct bn_type_traits {
class BatchNormMKLDNNHandler : public platform::MKLDNNHandler {
public:
BatchNormMKLDNNHandler(
std::shared_ptr<batch_norm_fwd::primitive_desc> batch_norm_pd,
const platform::MKLDNNDeviceContext &dev_ctx, mkldnn::engine engine,
const std::string &base_key)
: platform::MKLDNNHandler(dev_ctx, engine, base_key) {
batch_norm_pd_ = batch_norm_pd;
}
BatchNormMKLDNNHandler(const platform::MKLDNNDeviceContext &dev_ctx,
mkldnn::engine engine, const std::string &base_key)
: platform::MKLDNNHandler(dev_ctx, engine, base_key) {}
std::shared_ptr<memory> AcquireScaleshiftMemoryFromPrimitive(void *ptr) {
return this->AcquireMemoryFromPrimitive(
@ -62,6 +58,26 @@ class BatchNormMKLDNNHandler : public platform::MKLDNNHandler {
batch_norm_pd_->variance_primitive_desc(), ptr, "@variance_mem_p");
}
std::shared_ptr<batch_norm_fwd::primitive_desc>
AcquireBatchNormPrimitiveDescriptor(const batch_norm_fwd::desc &bn_fwd_desc,
const mkldnn::engine &engine) {
const std::string key_batch_norm_fwd_pd = key_ + "@bn_fwd_pd";
auto batch_norm_pd =
std::static_pointer_cast<batch_norm_fwd::primitive_desc>(
dev_ctx_.GetBlob(key_batch_norm_fwd_pd));
if (batch_norm_pd == nullptr) {
batch_norm_pd_.reset(
new batch_norm_fwd::primitive_desc(bn_fwd_desc, engine));
dev_ctx_.SetBlob(key_batch_norm_fwd_pd, batch_norm_pd_);
} else {
batch_norm_pd_ = batch_norm_pd;
is_reusing_ = true;
}
return batch_norm_pd_;
}
std::shared_ptr<batch_norm_fwd> AcquireTestTrainingBatchNormFwd(
std::shared_ptr<memory> src_memory,
std::shared_ptr<memory> scaleshift_memory,
@ -213,7 +229,7 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
const std::string key = BatchNormMKLDNNHandler::GetHash(
src_tz, epsilon, flags, global_stats, input_format,
ctx.op().Output("SavedMean"));
const std::string key_batch_norm_fwd_pd = key + "@bn_fwd_pd";
BatchNormMKLDNNHandler handler(dev_ctx, mkldnn_engine, key);
auto user_src_md = platform::MKLDNNMemDesc(
{src_tz}, platform::MKLDNNGetDataType<T>(), input_format);
@ -222,13 +238,9 @@ class BatchNormMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
using bn_fwd_types = bn_type_traits<mkldnn::batch_normalization_forward>;
auto batch_norm_fwd_desc =
bn_fwd_types::op_desc{propagation, user_src_md, epsilon, flags};
auto batch_norm_fwd_pd = std::make_shared<batch_norm_fwd::primitive_desc>(
batch_norm_fwd_desc, mkldnn_engine);
// Save conv_pd/src_memory/weights_memory for backward pass
dev_ctx.SetBlob(key_batch_norm_fwd_pd, batch_norm_fwd_pd);
BatchNormMKLDNNHandler handler(batch_norm_fwd_pd, dev_ctx, mkldnn_engine,
key);
auto batch_norm_fwd_pd = handler.AcquireBatchNormPrimitiveDescriptor(
batch_norm_fwd_desc, mkldnn_engine);
auto src_memory =
handler.AcquireSrcMemory(user_src_md, to_void_cast(x_data));

89
paddle/fluid/operators/mkldnn/conv_mkldnn_op.cc
Unescape View File

@ -144,7 +144,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
const std::string key = platform::ConvMKLDNNHandler::GetHash(
src_tz, weights_tz, strides, paddings, dilations, groups,
ctx.op().Input("Input") + ctx.op().Input("Filter"));
const std::string key_conv_pd = key + "@conv_pd";
std::vector<primitive> pipeline;
@ -183,6 +182,8 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
auto dst_md = platform::MKLDNNMemDesc(
dst_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
platform::ConvMKLDNNHandler handler(dev_ctx, mkldnn_engine, key);
// create a conv primitive descriptor and save it for usage in backward
std::shared_ptr<mkldnn::convolution_forward::primitive_desc> conv_pd;
auto fwd_prop_kind = is_test ? mkldnn::prop_kind::forward_inference
@ -191,18 +192,14 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
bias_tz = paddle::framework::vectorize2int(bias->dims());
auto bias_md = platform::MKLDNNMemDesc(
bias_tz, platform::MKLDNNGetDataType<T>(), memory::format::x);
conv_pd = ConvFwdPrimitiveDesc(
conv_pd = handler.AcquireConvolutionPrimitiveDescriptor(
src_md, weights_md, bias_md, dst_md, strides, paddings, mkldnn_engine,
fuse_relu, fuse_residual_conn, fwd_prop_kind);
} else {
conv_pd = ConvFwdPrimitiveDesc(src_md, weights_md, dst_md, strides,
paddings, mkldnn_engine, fuse_relu,
fuse_residual_conn, fwd_prop_kind);
conv_pd = handler.AcquireConvolutionPrimitiveDescriptor(
src_md, weights_md, boost::none, dst_md, strides, paddings,
mkldnn_engine, fuse_relu, fuse_residual_conn, fwd_prop_kind);
}
// Save conv_pd/src_memory/weights_memory for backward pass
if (!is_test) dev_ctx.SetBlob(key_conv_pd, conv_pd);
platform::ConvMKLDNNHandler handler(conv_pd, dev_ctx, mkldnn_engine, key);
// create mkldnn memory from input tensors (data/weights)
auto user_src_memory_p =
@ -633,31 +630,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
}
private:
mkldnn::primitive_attr CreatePostOps(bool fuse_relu,
bool fuse_residual_conn) const {
mkldnn::primitive_attr conv_attr;
mkldnn::post_ops post_operations;
// Fusion with Elementwise layer relies on adding a sum post-operation with
// the scale parameter. It is assumed that when fuse_residual_connection is
// true, the output tensor contains the data coming from residual
// connection. The result of this post_op is:
// Output = scale * Output + Conv_Out.
if (fuse_residual_conn) {
post_operations.append_sum(1.0f);
}
// Fusion with ReLU layer is executed through the PostOps feature. Create a
// PostOps object and configure it to execute an eltwise relu operation.
if (fuse_relu) {
constexpr float scale = 1.0f;
constexpr float negative_slope = 0.0f;
constexpr float placeholder = 0.0f;
post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_relu,
negative_slope, placeholder);
}
conv_attr.set_post_ops(post_operations);
return conv_attr;
}
mkldnn::primitive_attr CreatePostOps(
bool fuse_relu, bool fuse_residual_conn,
const std::vector<float> output_shift_scale, float sum_scale) const {
@ -679,30 +651,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
return conv_attr;
}
std::unique_ptr<mkldnn::convolution_forward::primitive_desc>
ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights,
const memory::desc& dst, const std::vector<int>& strides,
const std::vector<int>& paddings,
const mkldnn::engine& engine, const bool fuse_relu,
const bool fuse_residual_conn,
mkldnn::prop_kind fwd_prop_kind) const {
memory::dims stride_dims = strides;
memory::dims padding_dims = paddings;
auto conv_desc = mkldnn::convolution_forward::desc(
fwd_prop_kind, mkldnn::convolution_direct, src, weights, dst,
stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero);
mkldnn::primitive_attr conv_attr =
CreatePostOps(fuse_relu, fuse_residual_conn);
auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
conv_desc, conv_attr, engine);
return std::unique_ptr<mkldnn::convolution_forward::primitive_desc>(
p_conv_pd);
}
std::unique_ptr<mkldnn::convolution_forward::primitive_desc>
ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights,
const memory::desc& dst, const std::vector<int>& strides,
@ -731,31 +679,6 @@ class ConvMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
p_conv_pd);
}
std::unique_ptr<mkldnn::convolution_forward::primitive_desc>
ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights,
const memory::desc& bias, const memory::desc& dst,
const std::vector<int>& strides,
const std::vector<int>& paddings,
const mkldnn::engine& engine, const bool fuse_relu,
const bool fuse_residual_conn,
mkldnn::prop_kind fwd_prop_kind) const {
memory::dims stride_dims = strides;
memory::dims padding_dims = paddings;
auto conv_desc = mkldnn::convolution_forward::desc(
fwd_prop_kind, mkldnn::convolution_direct, src, weights, bias, dst,
stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero);
mkldnn::primitive_attr conv_attr =
CreatePostOps(fuse_relu, fuse_residual_conn);
auto p_conv_pd = new mkldnn::convolution_forward::primitive_desc(
conv_desc, conv_attr, engine);
return std::unique_ptr<mkldnn::convolution_forward::primitive_desc>(
p_conv_pd);
}
std::unique_ptr<mkldnn::convolution_forward::primitive_desc>
ConvFwdPrimitiveDesc(const memory::desc& src, const memory::desc& weights,
const memory::desc& bias, const memory::desc& dst,

82
paddle/fluid/operators/mkldnn/conv_transpose_mkldnn_op.cc
Unescape View File

@ -12,6 +12,7 @@
See the License for the specific language governing permissions and
limitations under the License. */
#include "boost/optional.hpp"
#include "paddle/fluid/framework/data_layout_transform.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/memory/malloc.h"
@ -124,7 +125,6 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
const std::string key = platform::ConvTransposeMKLDNNHandler::GetHash(
src_tz, weights_tz, strides, paddings, dilations, groups,
ctx.op().Output("Output"));
const std::string key_conv_transpose_pd = key + "@conv_transpose_pd";
std::vector<mkldnn::primitive> pipeline;
@ -153,6 +153,7 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
auto dst_md = platform::MKLDNNMemDesc(
dst_tz, platform::MKLDNNGetDataType<T>(), chosen_memory_format);
platform::ConvTransposeMKLDNNHandler handler(dev_ctx, mkldnn_engine, key);
// create a deconv(conv transpose) primitive descriptor and save it for
// usage in backward
std::shared_ptr<mkldnn::deconvolution_forward::primitive_desc>
@ -163,19 +164,14 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
bias_tz = paddle::framework::vectorize2int(bias->dims());
auto bias_md = platform::MKLDNNMemDesc(
bias_tz, platform::MKLDNNGetDataType<T>(), mkldnn::memory::format::x);
conv_transpose_pd = ConvTransposeFwdPrimitiveDesc(
conv_transpose_pd = handler.AcquireConvolutionPrimitiveDescriptor(
src_md, weights_md, bias_md, dst_md, strides, paddings, mkldnn_engine,
fuse_relu, fwd_prop_kind);
fuse_relu, false, fwd_prop_kind);
} else {
conv_transpose_pd = ConvTransposeFwdPrimitiveDesc(
src_md, weights_md, dst_md, strides, paddings, mkldnn_engine,
fuse_relu, fwd_prop_kind);
conv_transpose_pd = handler.AcquireConvolutionPrimitiveDescriptor(
src_md, weights_md, boost::none, dst_md, strides, paddings,
mkldnn_engine, fuse_relu, false, fwd_prop_kind);
}
// Save conv_pd/src_memory/weights_memory for backward pass
if (!is_test) dev_ctx.SetBlob(key_conv_transpose_pd, conv_transpose_pd);
platform::ConvTransposeMKLDNNHandler handler(conv_transpose_pd, dev_ctx,
mkldnn_engine, key);
// create mkldnn memory from input tensors (data/weights)
auto user_src_memory_p = handler.AcquireSrcMemory(
@ -224,70 +220,6 @@ class ConvTransposeMKLDNNOpKernel : public paddle::framework::OpKernel<T> {
output->set_layout(DataLayout::kMKLDNN);
output->set_format(platform::GetMKLDNNFormat(*dst_memory_p));
}
private:
mkldnn::primitive_attr CreatePostOps(bool fuse_relu) const {
mkldnn::primitive_attr conv_attr;
mkldnn::post_ops post_operations;
// Fusion with ReLU layer is executed through the PostOps feature. Create a
// PostOps object and configure it to execute an eltwise relu operation.
if (fuse_relu) {
constexpr float scale = 1.0f;
constexpr float negative_slope = 0.0f;
constexpr float placeholder = 0.0f;
post_operations.append_eltwise(scale, mkldnn::algorithm::eltwise_relu,
negative_slope, placeholder);
}
conv_attr.set_post_ops(post_operations);
return conv_attr;
}
std::unique_ptr<mkldnn::deconvolution_forward::primitive_desc>
ConvTransposeFwdPrimitiveDesc(
const mkldnn::memory::desc& src, const mkldnn::memory::desc& weights,
const mkldnn::memory::desc& dst, const std::vector<int>& strides,
const std::vector<int>& paddings, const mkldnn::engine& engine,
const bool fuse_relu, mkldnn::prop_kind fwd_prop_kind) const {
mkldnn::memory::dims stride_dims = {strides[0], strides[1]};
mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]};
auto deconv_desc = mkldnn::deconvolution_forward::desc(
fwd_prop_kind, mkldnn::deconvolution_direct, src, weights, dst,
stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero);
mkldnn::primitive_attr deconv_attr = CreatePostOps(fuse_relu);
auto p_conv_transpose_pd =
new mkldnn::deconvolution_forward::primitive_desc(deconv_desc,
deconv_attr, engine);
return std::unique_ptr<mkldnn::deconvolution_forward::primitive_desc>(
p_conv_transpose_pd);
}
std::unique_ptr<mkldnn::deconvolution_forward::primitive_desc>
ConvTransposeFwdPrimitiveDesc(
const mkldnn::memory::desc& src, const mkldnn::memory::desc& weights,
const mkldnn::memory::desc& bias, const mkldnn::memory::desc& dst,
const std::vector<int>& strides, const std::vector<int>& paddings,
const mkldnn::engine& engine, const bool fuse_relu,
mkldnn::prop_kind fwd_prop_kind) const {
mkldnn::memory::dims stride_dims = {strides[0], strides[1]};
mkldnn::memory::dims padding_dims = {paddings[0], paddings[1]};
auto deconv_desc = mkldnn::deconvolution_forward::desc(
fwd_prop_kind, mkldnn::deconvolution_direct, src, weights, bias, dst,
stride_dims, padding_dims, padding_dims, mkldnn::padding_kind::zero);
mkldnn::primitive_attr deconv_attr = CreatePostOps(fuse_relu);
auto p_conv_transpose_pd =
new mkldnn::deconvolution_forward::primitive_desc(deconv_desc,
deconv_attr, engine);
return std::unique_ptr<mkldnn::deconvolution_forward::primitive_desc>(
p_conv_transpose_pd);
}
};
} // namespace operators

38
paddle/fluid/operators/mkldnn/softmax_mkldnn_op.cc
Unescape View File

@ -34,12 +34,9 @@ using platform::to_void_cast;
class SoftmaxMKLDNNHandler : public platform::MKLDNNHandler {
public:
SoftmaxMKLDNNHandler(
std::shared_ptr<mkldnn::softmax_forward::primitive_desc> softmax_pd,
const platform::MKLDNNDeviceContext& dev_ctx, mkldnn::engine engine,
const std::string& base_key)
: platform::MKLDNNHandler(dev_ctx, engine, base_key),
softmax_pd_(softmax_pd) {}
SoftmaxMKLDNNHandler(const platform::MKLDNNDeviceContext& dev_ctx,
mkldnn::engine engine, const std::string& base_key)
: platform::MKLDNNHandler(dev_ctx, engine, base_key) {}
SoftmaxMKLDNNHandler(
std::shared_ptr<mkldnn::softmax_forward::primitive_desc> softmax_pd,
@ -54,6 +51,26 @@ class SoftmaxMKLDNNHandler : public platform::MKLDNNHandler {
key_ += "-BWD";
}
std::shared_ptr<softmax_forward::primitive_desc>
AcquireSoftmaxPrimitiveDescriptor(const softmax_forward::desc& softmax_desc,
const mkldnn::engine& engine) {
const std::string key_softmax_pd = key_ + "@softmax_pd";
auto softmax_pd = std::static_pointer_cast<softmax_forward::primitive_desc>(
dev_ctx_.GetBlob(key_softmax_pd));
if (softmax_pd == nullptr) {
softmax_pd_.reset(
new softmax_forward::primitive_desc(softmax_desc, engine));
dev_ctx_.SetBlob(key_softmax_pd, softmax_pd_);
} else {
softmax_pd_ = softmax_pd;
is_reusing_ = true;
}
return softmax_pd_;
}
std::shared_ptr<mkldnn::softmax_forward> AcquireSoftmax(
std::shared_ptr<mkldnn::memory> dst_memory_p,
std::shared_ptr<mkldnn::memory> src_memory_p) {
@ -138,19 +155,18 @@ class SoftmaxMKLDNNKernel : public paddle::framework::OpKernel<T> {
// Generate keys for storing/retriving primitives for this operator
const std::string key =
platform::MKLDNNHandler::GetHash(softmax_tz, ctx.op().Output("Out"));
const std::string key_softmax_pd = key + "@softmax_pd";
SoftmaxMKLDNNHandler handler(dev_ctx, mkldnn_engine, key);
// Currently only NC data format is supported
auto softmax_md = MKLDNNMemDesc(
{softmax_tz}, platform::MKLDNNGetDataType<T>(), memory::format::nc);
// Normalization is made after innermost dimension eg. C out of NC
auto softmax_desc = softmax_forward::desc(prop_kind::forward_scoring,
softmax_md, 1 /*dim: C*/);
auto softmax_pd = std::make_shared<mkldnn::softmax_forward::primitive_desc>(
softmax_desc, mkldnn_engine);
dev_ctx.SetBlob(key_softmax_pd, softmax_pd);
SoftmaxMKLDNNHandler handler(softmax_pd, dev_ctx, mkldnn_engine, key);
auto softmax_pd =
handler.AcquireSoftmaxPrimitiveDescriptor(softmax_desc, mkldnn_engine);
auto softmax_src_memory_p =
handler.AcquireSrcMemory(softmax_md, to_void_cast<T>(input_data));
auto softmax_dst_memory_p =

12
paddle/fluid/operators/pad2d_op.cc
Unescape View File

@ -483,8 +483,10 @@ class Pad2dOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE_EQ(
paddings_dim.size(), 1,
"Size of Input(Paddings)'s dimension should be equal to 1.");
PADDLE_ENFORCE_EQ(paddings_dim[0], 4,
"Shape of Input(Paddings) should be equal to [4].");
if (ctx->IsRuntime()) {
PADDLE_ENFORCE_EQ(paddings_dim[0], 4,
"Shape of Input(Paddings) should be equal to [4].");
}
out_dims[1] = x_dim[1];
out_dims[2] = x_dim[2];
out_dims[3] = x_dim[3];
@ -504,11 +506,7 @@ class Pad2dOp : public framework::OperatorWithKernel {
}
ctx->SetOutputDim("Out", framework::make_ddim(out_dims));
if (out_dims[0] == x_dim[0]) {
// Only pass LoD when the first dimension is equal between
// output and input.
ctx->ShareLoD("X", /*->*/ "Out");
}
ctx->ShareLoD("X", /*->*/ "Out");
}
protected:

Some files were not shown because too many files have changed in this diff Show More

Write Preview
Loading…
Cancel
Save