diff --git a/paddle/fluid/operators/random_crop_op.cc b/paddle/fluid/operators/random_crop_op.cc index 528a6e4a1b68fe611d104f21bafe970762611a03..123fa44fa3ddbc9343b9629be63fdefdf12b4646 100644 --- a/paddle/fluid/operators/random_crop_op.cc +++ b/paddle/fluid/operators/random_crop_op.cc @@ -37,6 +37,11 @@ class RandomCropOpMaker : public framework::OpProtoAndCheckerMaker { AddOutput("SeedOut", "The random seed after random cropping.") .AsIntermediate(); AddAttr>("shape", "The shape of a cropped instance."); + AddAttr("startup_seed", + "If the input 'Seed' is not initialized, the 'startup_seed' " + "will be used to replace it. Even so, the seed after random " + "crop will also be outputed to the 'SeedOut'.") + .SetDefault(0); AddComment(R"DOC( This operator takes a batch of instance, and do random cropping on each instance. It means that cropping positions differs on each instance, which is determined @@ -49,8 +54,6 @@ class RandomCropOpMaker : public framework::OpProtoAndCheckerMaker { class RandomCropOpInferShape : public framework::InferShapeBase { public: void operator()(framework::InferShapeContext* ctx) const override { - auto seed_dim = ctx->GetInputDim("Seed"); - PADDLE_ENFORCE(seed_dim.size() == 1 && seed_dim[0] == 1); auto shape = ctx->Attrs().Get>("shape"); auto x_dim = ctx->GetInputDim("X"); PADDLE_ENFORCE_GT(x_dim.size(), static_cast(shape.size())); @@ -62,7 +65,6 @@ class RandomCropOpInferShape : public framework::InferShapeBase { out_dim[x_i] = shape[shape_i]; } ctx->SetOutputDim("Out", framework::make_ddim(out_dim)); - ctx->SetOutputDim("SeedOut", framework::make_ddim({1})); } }; diff --git a/paddle/fluid/operators/random_crop_op.h b/paddle/fluid/operators/random_crop_op.h index f3261cbdc986b0cc724315c1eb92b8b84e18c742..d68ba9d661698bb0d33b139f5748daec2ead6595 100644 --- a/paddle/fluid/operators/random_crop_op.h +++ b/paddle/fluid/operators/random_crop_op.h @@ -142,16 +142,22 @@ template class RandomCropKernel : public framework::OpKernel { public: virtual void Compute(const framework::ExecutionContext& ctx) const { - auto& seed_tensor = detail::Ref(ctx.Input("Seed")); int64_t seed = 0; - if (platform::is_cpu_place(seed_tensor.place())) { - seed = *seed_tensor.data(); + auto& seed_tensor = detail::Ref(ctx.Input("Seed")); + if (seed_tensor.IsInitialized()) { + if (platform::is_cpu_place(seed_tensor.place())) { + seed = *seed_tensor.data(); + } else { + LOG(WARNING) << "It is slow to place seed in GPU memory. Please verify " + "your program"; + framework::LoDTensor cpu_seed; + framework::TensorCopySync(seed_tensor, platform::CPUPlace(), &cpu_seed); + seed = *cpu_seed.data(); + } } else { - LOG(WARNING) << "It is slow to place seed in GPU memory. Please verify " - "your program"; - framework::LoDTensor cpu_seed; - framework::TensorCopySync(seed_tensor, platform::CPUPlace(), &cpu_seed); - seed = *cpu_seed.data(); + VLOG(5) << "WARNING: The input 'Seed' is not initialized, use attribute " + "'startup_seed' instead."; + seed = ctx.Attr("startup_seed"); } auto shape = ctx.Attr>("shape"); auto& x = detail::Ref(ctx.Input("X")); @@ -171,7 +177,7 @@ class RandomCropKernel : public framework::OpKernel { engine.discard(functor.prod_batchsize_dims_ * (functor.rank_ - functor.num_batchsize_dims_)); *ctx.Output("SeedOut")->mutable_data( - platform::CPUPlace()) = engine(); + framework::make_ddim({1}), platform::CPUPlace()) = engine(); } }; diff --git a/paddle/fluid/operators/reader/create_custom_reader_op.cc b/paddle/fluid/operators/reader/create_custom_reader_op.cc index 0a02fcdeaa5a6de97d59ddce4f58ad945aa2572a..3f299f6ee29bdce2a5ea2502290bbf7dedb63d0b 100644 --- a/paddle/fluid/operators/reader/create_custom_reader_op.cc +++ b/paddle/fluid/operators/reader/create_custom_reader_op.cc @@ -39,6 +39,7 @@ class CustomReader : public framework::DecoratedReader { const framework::ProgramDesc program_; int sub_block_id_; framework::Executor exe_; + framework::Scope scope_; std::vector source_var_names_; std::vector sink_var_names_; @@ -158,20 +159,20 @@ void CustomReader::ReadNext(std::vector* out) { // The scope for CustomReader's sub-block should be independent and shouldn't // be any other computation scope's child. Otherwise, data preprocessing and // compution cannot be concurrent. - framework::Scope scope; + framework::Scope& exe_scope = scope_.NewScope(); // 1. Copy LoDTensors from underlying reader's output to source variables. for (size_t i = 0; i < source_var_names_.size(); ++i) { - framework::Variable* var = scope.Var(source_var_names_[i]); + framework::Variable* var = exe_scope.Var(source_var_names_[i]); framework::LoDTensor* tensor = var->GetMutable(); tensor->ShareDataWith(underlying_outs[i]); tensor->set_lod(underlying_outs[i].lod()); } // 2. Run the sub-block. - exe_.Run(program_, &scope, sub_block_id_, false, true); + exe_.Run(program_, &exe_scope, sub_block_id_, false, true); // 3. Copy LoDTensors from sink variables to out. out->resize(sink_var_names_.size()); for (size_t i = 0; i < sink_var_names_.size(); ++i) { - const auto& tensor = detail::Ref(scope.FindVar(sink_var_names_[i])) + const auto& tensor = detail::Ref(exe_scope.FindVar(sink_var_names_[i])) .Get(); framework::TensorCopySync(tensor, platform::CPUPlace(), &(*out)[i]); } diff --git a/python/paddle/fluid/layers/nn.py b/python/paddle/fluid/layers/nn.py index f6f188df0d6a9a33f4ad858f00c1ba0fd36661b9..0f9cfdea77990b1a5cf2826f695ec52b06a731e2 100644 --- a/python/paddle/fluid/layers/nn.py +++ b/python/paddle/fluid/layers/nn.py @@ -23,6 +23,7 @@ from layer_function_generator import autodoc, templatedoc from tensor import concat import utils import random +from .. import unique_name __all__ = [ 'fc', @@ -846,7 +847,7 @@ def crf_decoding(input, param_attr, label=None): Returns: Variable: ${viterbi_path_comment} - + Examples: .. code-block:: python @@ -1084,7 +1085,7 @@ def chunk_eval(input, Here is a NER example of labeling for these tagging schemes: .. code-block:: python - + ====== ====== ====== ===== == ============ ===== ===== ===== == ========= Li Ming works at Agricultural Bank of China in Beijing. ====== ====== ====== ===== == ============ ===== ===== ===== == ========= @@ -1110,7 +1111,7 @@ def chunk_eval(input, is the num of chunk types, and `tag_type` get its value from the following table. .. code-block:: python - + Scheme Begin Inside End Single plain 0 - - - IOB 0 1 - - @@ -1146,7 +1147,7 @@ def chunk_eval(input, tuple: tuple containing: precision, recall, f1_score, num_infer_chunks, num_label_chunks, num_correct_chunks - + Examples: .. code-block:: python @@ -1266,7 +1267,7 @@ def sequence_softmax(input, param_attr=None, bias_attr=None, use_cudnn=True): param_attr (ParamAttr|None): attributes for parameter use_cudnn (bool): Use cudnn kernel or not, it is valid only when the cudnn \ library is installed. Default: True - + Returns: Variable: output of sequence_softmax @@ -4143,7 +4144,7 @@ def one_hot(input, depth): Examples: .. code-block:: python - + label = layers.data(name="label", shape=[1], dtype="float32") one_hot_label = layers.one_hot(input=label, depth=10) """ @@ -4862,40 +4863,32 @@ def random_crop(x, shape, seed=None): Returns: ${out_comment} - + Examples: >>> img = fluid.layers.data("img", [3, 256, 256]) >>> cropped_img = fluid.layers.random_crop(img, shape=[3, 224, 224]) """ helper = LayerHelper("random_crop", **locals()) - dtype = helper.input_dtype() + dtype = x.dtype out = helper.create_tmp_variable(dtype) if seed is None: seed = random.randint(-65536, 65535) - + op_attrs = {"shape": shape} if isinstance(seed, int): - seed_value = seed - seed = helper.create_tmp_variable(dtype="int64") - helper.append_op( - type="fill_constant", - inputs={}, - outputs={"Out": seed}, - attrs={ - "dtype": seed.dtype, - "shape": [1], - "value": float(seed_value), - "force_cpu": True - }) + op_attrs["startup_seed"] = seed + seed = helper.create_variable( + name=unique_name.generate("random_crop_seed"), + dtype="int64", + persistable=True) elif not isinstance(seed, Variable): raise ValueError("'seed' must be a Variable or an int.") - seed_out = helper.create_tmp_variable(dtype="int64") helper.append_op( type="random_crop", inputs={"X": x, "Seed": seed}, outputs={"Out": out, - "SeedOut": seed_out}, - attrs={"shape": shape}) + "SeedOut": seed}, + attrs=op_attrs) return out @@ -4961,7 +4954,7 @@ def mean_iou(input, label, num_classes): semantic image segmentation, which first computes the IOU for each semantic class and then computes the average over classes. IOU is defined as follows: - + .. math:: IOU = \\frac{true\_positiv}{(true\_positive + false\_positive + false\_negative)}. @@ -4984,7 +4977,7 @@ def mean_iou(input, label, num_classes): Examples: .. code-block:: python - + iou, wrongs, corrects = fluid.layers.mean_iou(predict, label, num_classes) """ helper = LayerHelper('mean_iou', **locals())