From 15a5bd9cdbdd371fa8d4634bc340223e1c1418ce Mon Sep 17 00:00:00 2001 From: co63oc Date: Sun, 7 Apr 2024 20:34:00 +0800 Subject: [PATCH 1/6] Fix --- paddle/fluid/framework/ir/CMakeLists.txt | 1 - .../framework/ir/graph_pattern_detector.cc | 46 -- .../framework/ir/graph_pattern_detector.h | 25 - .../ir/seqconv_eltadd_relu_fuse_pass.cc | 163 ------- .../ir/seqconv_eltadd_relu_fuse_pass.h | 42 -- .../inference/api/paddle_pass_builder.cc | 3 +- .../sequence_ops/sequence_conv_op.cc | 279 ----------- .../sequence_ops/sequence_conv_op.cu.cc | 25 - .../operators/sequence_ops/sequence_conv_op.h | 204 -------- .../sequence_ops/sequence_conv_op_xpu.cc | 374 --------------- python/paddle/static/nn/__init__.py | 2 - python/paddle/static/nn/sequence_lod.py | 141 ------ .../quantization/quant2_int8_onednn_pass.py | 1 - test/book/notest_understand_sentiment.py | 266 ----------- test/book/test_recommender_system.py | 392 --------------- .../test_seqconv_eltadd_relu_fuse_pass.py | 120 ----- test/legacy_test/CMakeLists.txt | 1 - test/legacy_test/dist_text_classification.py | 242 ---------- test/legacy_test/nets.py | 80 ---- .../test_dist_text_classification.py | 16 - .../test_fusion_seqconv_eltadd_relu_op.py | 95 ---- test/sequence/CMakeLists.txt | 1 - test/sequence/test_sequence_conv.py | 312 ------------ test/xpu/test_sequence_conv_op_xpu.py | 448 ------------------ 24 files changed, 1 insertion(+), 3278 deletions(-) delete mode 100644 paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc delete mode 100644 paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h delete mode 100644 paddle/fluid/operators/sequence_ops/sequence_conv_op.cc delete mode 100644 paddle/fluid/operators/sequence_ops/sequence_conv_op.cu.cc delete mode 100644 paddle/fluid/operators/sequence_ops/sequence_conv_op.h delete mode 100644 paddle/fluid/operators/sequence_ops/sequence_conv_op_xpu.cc delete mode 100644 test/book/notest_understand_sentiment.py delete mode 100644 test/book/test_recommender_system.py delete mode 100644 test/ir/inference/test_seqconv_eltadd_relu_fuse_pass.py delete mode 100644 test/legacy_test/dist_text_classification.py delete mode 100644 test/legacy_test/test_fusion_seqconv_eltadd_relu_op.py delete mode 100644 test/sequence/test_sequence_conv.py delete mode 100644 test/xpu/test_sequence_conv_op_xpu.py diff --git a/paddle/fluid/framework/ir/CMakeLists.txt b/paddle/fluid/framework/ir/CMakeLists.txt index d310d4ea5a4ee..faa4c17411401 100755 --- a/paddle/fluid/framework/ir/CMakeLists.txt +++ b/paddle/fluid/framework/ir/CMakeLists.txt @@ -81,7 +81,6 @@ pass_library(seq_concat_fc_fuse_pass inference) pass_library(multi_batch_merge_pass base) pass_library(map_op_to_another_pass inference) pass_library(conv_bn_fuse_pass inference) -pass_library(seqconv_eltadd_relu_fuse_pass inference) pass_library(seqpool_concat_fuse_pass inference) pass_library(seqpool_cvm_concat_fuse_pass inference) pass_library(repeated_fc_relu_fuse_pass inference) diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc index 3910e7586e35c..2ee2f7399cf60 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.cc +++ b/paddle/fluid/framework/ir/graph_pattern_detector.cc @@ -1057,52 +1057,6 @@ PDNode *patterns::OperatorReshape2::operator()(const std::string &operator_type, return reshape2_out; } -PDNode *patterns::SeqConvEltAddRelu::operator()( - paddle::framework::ir::PDNode *seqconv_input) { - // Create Operators - seqconv_input->assert_is_op_input("sequence_conv", "X"); - auto *seqconv_op = pattern->NewNode(seqconv_repr()) - ->assert_is_op("sequence_conv") - ->assert_has_n_inputs(2) - ->assert_op_attr("paddingTrainable", false) - ->assert_op_attr("contextStride", 1); - - auto *eltadd_op = - pattern->NewNode(eltadd_repr())->assert_is_op("elementwise_add"); - auto *relu_op = pattern->NewNode(relu_repr())->assert_is_op("relu"); - // Create variables - // Filter - auto *seqconv_weight_var = - pattern->NewNode(seqconv_weight_repr()) - ->AsInput() - ->assert_is_persistable_var() - ->assert_is_op_input("sequence_conv", "Filter"); - // Bias - auto *eltadd_bias_var = pattern->NewNode(eltadd_bias_repr()) - ->AsInput() - ->assert_is_op_input("elementwise_add"); - // intermediate variable, will be removed in the IR after fuse. - auto *seqconv_out_var = pattern->NewNode(seqconv_out_repr()) - ->AsIntermediate() - ->assert_is_only_output_of_op("sequence_conv") - ->assert_is_op_input("elementwise_add"); - auto *eltadd_out_var = pattern->NewNode(eltadd_out_repr()) - ->AsIntermediate() - ->assert_is_only_output_of_op("elementwise_add") - ->assert_is_only_input_of_op("relu"); - // output - auto *relu_out_var = pattern->NewNode(relu_out_repr()) - ->AsOutput() - ->assert_is_op_output("relu"); - - seqconv_op->LinksFrom({seqconv_input, seqconv_weight_var}) - .LinksTo({seqconv_out_var}); - eltadd_op->LinksFrom({seqconv_out_var, eltadd_bias_var}) - .LinksTo({eltadd_out_var}); - relu_op->LinksFrom({eltadd_out_var}).LinksTo({relu_out_var}); - return relu_out_var; -} - PDNode *patterns::FC::operator()(paddle::framework::ir::PDNode *x, bool with_bias, bool with_relu) { diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.h b/paddle/fluid/framework/ir/graph_pattern_detector.h index 4eac3440a4514..f7f3a2f9c3ea1 100644 --- a/paddle/fluid/framework/ir/graph_pattern_detector.h +++ b/paddle/fluid/framework/ir/graph_pattern_detector.h @@ -600,31 +600,6 @@ struct OperatorReshape2 : public PatternBase { PATTERN_DECL_NODE(reshape2_out); }; -// SEQCONV with Elementwise_Add ReLU -// op: seqconv + elementwise_add + relu -// named nodes: -// seqconv_input, seqconv_weight, -// seqconv_out, seqconv, -// elementwise_add_bias, elementwise_add_out, elementwise_add -// relu_out, relu -struct SeqConvEltAddRelu : public PatternBase { - SeqConvEltAddRelu(PDPattern* pattern, const std::string& name_scope) - : PatternBase(pattern, name_scope, "seqconv_eltadd_relu") {} - - PDNode* operator()(PDNode* seqconv_input); - - // declare operator node's name - PATTERN_DECL_NODE(seqconv); - PATTERN_DECL_NODE(eltadd); - PATTERN_DECL_NODE(relu); - // declare variable node's name - PATTERN_DECL_NODE(seqconv_weight); - PATTERN_DECL_NODE(seqconv_out); - PATTERN_DECL_NODE(eltadd_bias); - PATTERN_DECL_NODE(eltadd_out); - PATTERN_DECL_NODE(relu_out); -}; - // FC with bias // op: mul + elementwise_add // named nodes: diff --git a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc deleted file mode 100644 index c2015d6b4305c..0000000000000 --- a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.cc +++ /dev/null @@ -1,163 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h" - -#include - -#include "paddle/fluid/framework/op_version_registry.h" - -namespace paddle { -namespace framework { -class Scope; -} // namespace framework -} // namespace paddle - -namespace paddle { -namespace framework { -namespace ir { - -SeqConvEltAddReluFusePass::SeqConvEltAddReluFusePass() { - AddOpCompat(OpCompat("sequence_conv")) - .AddInput("X") - .IsTensor() - .End() - .AddInput("Filter") - .IsTensor() - .End() - .AddInput("PaddingData") - .IsOptional() - .IsTensor() - .End() - .AddOutput("Out") - .IsTensor() - .End() - .AddAttr("contextLength") - .IsNumGT(0) - .End() - .AddAttr("contextStart") // the contextStart attribute can be negative, - // unconstrained - .End() - .AddAttr("contextStride") - .IsNumEQ(1) - .End(); - - AddOpCompat(OpCompat("elementwise_add")) - .AddInput("X") - .IsTensor() - .End() - .AddInput("Y") - .IsTensor() - .End() - .AddOutput("Out") - .IsTensor() - .End() - .AddAttr("axis") - .IsNumEQ(1) - .End(); - - AddOpCompat(OpCompat("relu")) - .AddInput("X") - .IsTensor() - .End() - .AddOutput("Out") - .IsTensor() - .End(); -} - -class Node; - -void SeqConvEltAddReluFusePass::ApplyImpl(ir::Graph* graph) const { - FusePassBase::Init(name_scope_, graph); - GraphPatternDetector gpd; - auto* pattern = gpd.mutable_pattern(); - - PDNode* x = pattern->NewNode(patterns::PDNodeName(name_scope_, "X")) - ->assert_is_op_input("sequence_conv") - ->assert_var_not_persistable(); - patterns::SeqConvEltAddRelu fuse_pattern(pattern, name_scope_); - fuse_pattern(x); - - // Create New OpDesc - auto fuse_creator = [&](Node* seqconv, - Node* input, - Node* seqconv_weight, - Node* eltadd_bias, - Node* relu_out) { - OpDesc op_desc; - op_desc.SetType("fusion_seqconv_eltadd_relu"); - op_desc.SetInput("X", {input->Name()}); - op_desc.SetInput("Filter", {seqconv_weight->Name()}); - op_desc.SetInput("Bias", {eltadd_bias->Name()}); - op_desc.SetAttr("contextLength", seqconv->Op()->GetAttr("contextLength")); - op_desc.SetAttr("contextStart", seqconv->Op()->GetAttr("contextStart")); - op_desc.SetAttr("contextStride", seqconv->Op()->GetAttr("contextStride")); - const std::string ColMat = patterns::UniqueKey("SeqConvColMat"); - op_desc.SetOutput("ColMat", {ColMat}); - op_desc.SetOutput("Out", {relu_out->Name()}); - VarDesc key(ColMat); - key.SetPersistable(false); - auto* key_col_mat = graph->CreateVarNode(&key); - - auto* op = graph->CreateOpNode(&op_desc); - IR_NODE_LINK_TO(input, op); - IR_NODE_LINK_TO(seqconv_weight, op); - IR_NODE_LINK_TO(eltadd_bias, op); - IR_NODE_LINK_TO(op, relu_out); - IR_NODE_LINK_TO(op, key_col_mat); - return op; - }; - - int fusion_count{0}; - - auto handler = [&](const GraphPatternDetector::subgraph_t& subgraph, - Graph* g) { - if (!IsCompat(subgraph, g)) { - LOG(WARNING) << "Pass in op compat failed."; - return; - } - VLOG(4) << "handle SeqConv EltAdd Relu fuse"; - GET_IR_NODE_FROM_SUBGRAPH(seqconv, seqconv, fuse_pattern); - GET_IR_NODE_FROM_SUBGRAPH(seqconv_weight, seqconv_weight, fuse_pattern); - GET_IR_NODE_FROM_SUBGRAPH(seqconv_out, seqconv_out, fuse_pattern); - GET_IR_NODE_FROM_SUBGRAPH(eltadd, eltadd, fuse_pattern); - GET_IR_NODE_FROM_SUBGRAPH(eltadd_bias, eltadd_bias, fuse_pattern); - GET_IR_NODE_FROM_SUBGRAPH(eltadd_out, eltadd_out, fuse_pattern); - GET_IR_NODE_FROM_SUBGRAPH(relu, relu, fuse_pattern); - GET_IR_NODE_FROM_SUBGRAPH(relu_out, relu_out, fuse_pattern); - - fuse_creator( - seqconv, subgraph.at(x), seqconv_weight, eltadd_bias, relu_out); - std::unordered_set marked_nodes( - {seqconv, seqconv_out, eltadd, eltadd_out, relu}); - GraphSafeRemoveNodes(graph, marked_nodes); - ++fusion_count; - }; - - gpd(graph, handler); - AddStatis(fusion_count); -} - -} // namespace ir -} // namespace framework -} // namespace paddle - -REGISTER_PASS(seqconv_eltadd_relu_fuse_pass, - paddle::framework::ir::SeqConvEltAddReluFusePass); -REGISTER_PASS_CAPABILITY(seqconv_eltadd_relu_fuse_pass) - .AddCombination( - paddle::framework::compatible::OpVersionComparatorCombination() - .EQ("sequence_conv", 0) - .LE("elementwise_add", 1) - .EQ("relu", 0)); diff --git a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h b/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h deleted file mode 100644 index fe06002251ae2..0000000000000 --- a/paddle/fluid/framework/ir/seqconv_eltadd_relu_fuse_pass.h +++ /dev/null @@ -1,42 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#pragma once - -#include - -#include "paddle/fluid/framework/ir/fuse_pass_base.h" -#include "paddle/fluid/framework/ir/graph.h" -#include "paddle/fluid/framework/ir/graph_pattern_detector.h" - -namespace paddle { -namespace framework { -namespace ir { - -class Graph; - -class SeqConvEltAddReluFusePass : public FusePassBase { - public: - SeqConvEltAddReluFusePass(); - virtual ~SeqConvEltAddReluFusePass() {} - - protected: - void ApplyImpl(ir::Graph* graph) const override; - - const std::string name_scope_{"seqconv_eltadd_relu_fuse"}; -}; - -} // namespace ir -} // namespace framework -} // namespace paddle diff --git a/paddle/fluid/inference/api/paddle_pass_builder.cc b/paddle/fluid/inference/api/paddle_pass_builder.cc index 48d25ebc390b5..f53e89d9f812c 100644 --- a/paddle/fluid/inference/api/paddle_pass_builder.cc +++ b/paddle/fluid/inference/api/paddle_pass_builder.cc @@ -234,8 +234,7 @@ const std::vector kCINNCompilerPasses{ const std::vector CpuBasicPasses{ "simplify_with_basic_ops_pass", // "layer_norm_fuse_pass", - "attention_lstm_fuse_pass", // - "seqconv_eltadd_relu_fuse_pass", // + "attention_lstm_fuse_pass", // // "seqpool_concat_fuse_pass", // "seqpool_cvm_concat_fuse_pass", // // "embedding_fc_lstm_fuse_pass", // diff --git a/paddle/fluid/operators/sequence_ops/sequence_conv_op.cc b/paddle/fluid/operators/sequence_ops/sequence_conv_op.cc deleted file mode 100644 index c94f57807cd52..0000000000000 --- a/paddle/fluid/operators/sequence_ops/sequence_conv_op.cc +++ /dev/null @@ -1,279 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/fluid/operators/sequence_ops/sequence_conv_op.h" - -#include -#include -#include -#include - -namespace paddle { -namespace operators { - -class SequenceConvOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(framework::InferShapeContext *ctx) const override { - OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "SequenceConv"); - OP_INOUT_CHECK(ctx->HasInput("Filter"), "Input", "Filter", "SequenceConv"); - OP_INOUT_CHECK(ctx->HasOutput("Out"), "Output", "Out", "SequenceConv"); - - int context_length = ctx->Attrs().Get("contextLength"); - int context_start = ctx->Attrs().Get("contextStart"); - - auto in_dims = ctx->GetInputDim("X"); - auto filter_dims = ctx->GetInputDim("Filter"); - PADDLE_ENFORCE_EQ( - ctx->Attrs().Get("contextStride"), - 1, - platform::errors::InvalidArgument( - "Currently, SequenceConvOp only supports contextStride=1. But " - "received contextStride = %u.", - ctx->Attrs().Get("contextStride"))); - PADDLE_ENFORCE_EQ( - in_dims.size() == 2 && filter_dims.size() == 2, - true, - platform::errors::InvalidArgument( - "Input(X, Filter) should be 2-D tensor. But received Input(X): " - "input rank %u, input shape [%s]; received Input(Filter): " - "input rank %u, input shape [%s].", - in_dims.size(), - in_dims, - filter_dims.size(), - filter_dims)); - PADDLE_ENFORCE_EQ( - filter_dims[0], - context_length * in_dims[1], - platform::errors::InvalidArgument( - "Filter's height should be context_length * " - "input_hidden_size. But received: filter's height = %d, " - "context_length * input_hidden_size = %d.", - filter_dims[0], - context_length * in_dims[1])); - - if (ctx->Attrs().Get("paddingTrainable")) { - OP_INOUT_CHECK(ctx->HasInput("PaddingData"), - "Input", - "PaddingData", - "sequence_conv"); - framework::DDim padding_dim = ctx->GetInputDim("PaddingData"); - int up_pad = std::max(0, -context_start); - int down_pad = std::max(0, context_start + context_length - 1); - int total_pad = up_pad + down_pad; - int input_width = static_cast(in_dims[1]); - bool start_equals_zero = context_start == 0; - bool length_equals_one = context_length == 1; - bool start_length = start_equals_zero && length_equals_one; - - PADDLE_ENFORCE_EQ( - start_length, - false, - platform::errors::InvalidArgument( - "If context_start is 0 and context_length is 1, paddingTrainable " - "should be false.")); - PADDLE_ENFORCE_EQ( - padding_dim.size(), - 2, - platform::errors::InvalidArgument( - "Input(PaddingData) should be 2-D tensor. But received: " - "input rank %u, input shape [%s].", - padding_dim.size(), - padding_dim)); - PADDLE_ENFORCE_EQ( - padding_dim[0] == total_pad && padding_dim[1] == input_width, - true, - platform::errors::InvalidArgument("Input(PaddingData)'s shape is not " - "consistent with 'context_start' " - "and 'context_length'. Received " - "Input(PaddingData): input rank " - "%u, " - "input shape [%s].", - padding_dim.size(), - padding_dim)); - } - - in_dims[1] = filter_dims[1]; - ctx->SetOutputDim("Out", in_dims); - ctx->ShareLoD("X", "Out"); - } -}; - -class SequenceConvGradOp : public framework::OperatorWithKernel { - public: - using framework::OperatorWithKernel::OperatorWithKernel; - - protected: - void InferShape(framework::InferShapeContext *ctx) const override { - OP_INOUT_CHECK(ctx->HasInput(framework::GradVarName("Out")), - "Input", - framework::GradVarName("Out"), - "SequenceConvGrad"); - OP_INOUT_CHECK(ctx->HasInput("X"), "Input", "X", "SequenceConvGrad"); - - if (ctx->Attrs().Get("paddingTrainable") && - ctx->HasOutput(framework::GradVarName("PaddingData"))) { - ctx->SetOutputDim(framework::GradVarName("PaddingData"), - ctx->GetInputDim("PaddingData")); - } - if (ctx->HasOutput(framework::GradVarName("X"))) { - ctx->ShareDim("X", /*->*/ framework::GradVarName("X")); - ctx->ShareLoD("X", /*->*/ framework::GradVarName("X")); - } - if (ctx->HasOutput(framework::GradVarName("Filter"))) { - ctx->SetOutputDim(framework::GradVarName("Filter"), - ctx->GetInputDim("Filter")); - } - } -}; - -class SequenceConvOpMaker : public framework::OpProtoAndCheckerMaker { - public: - void Make() override { - AddInput( - "X", - "(phi::DenseTensor) the input(X) is a LodTensor, which supports " - "variable-time length input sequence. The underlying tensor in " - "this phi::DenseTensor is a matrix with shape (T, N), where T is the " - "total time steps in this mini-batch and N is the input_hidden_size."); - AddInput( - "PaddingData", - "(phi::DenseTensor, optional) the input(PaddingData) is an optional " - "parameter, and it is learnable. " - "This is a tensor with shape (P, N), where P is the " - "top_pad + bottom_pad, N is the input_hidden_size. In order to " - "ensure the equal length of sequence before and after " - "convolution, it is necessary to fill the top and bottom of each " - "sequence according to context_length, context_stride and " - "context_start") - .AsDispensable(); - AddInput( - "Filter", - "(phi::DenseTensor) the input(Filter) is an learnable parameter." - "This is a tensor with shape (K, M), where K is the " - "context_length * input_hidden_size, M is the output feature size."); - AddOutput( - "Out", - "(phi::DenseTensor) the output(Out) is a LodTensor, which support " - "variable-time length output sequence. The underlying tensor in " - "this phi::DenseTensor is a matrix with shape (T, M), where, T is the " - "total time steps in this mini-batch, M is the output feature size."); - - AddAttr("paddingTrainable", - "(bool, default:false) the padding data of SequenceConvOp " - "is trainable or not.") - .SetDefault(false); - AddAttr("contextLength", - "(int) the contextLength of SequenceConvOp is the " - "height of the convolution kernel.") - .GreaterThan(0); - AddAttr("contextStart", - "(int, default:0) the contextStart of SequenceConvOp " - "represents the beginning of the convolution of the number of " - "rows of sequence, which can be negative. The negative number " - "means to pad contextStart time-steps of zeros or learnable " - "parameters at the beginning of each instance. The positive " - "number means to skip contextStart time-steps of each " - "instance.") - .SetDefault(0); - AddAttr("contextStride", - "(int, default:1) the contextStride of SequenceConvOp " - "represents the stride length of convolution kernel. " - "Currently, SequenceConvOp only supports" - "contextStride=1.") - .SetDefault(1) - .GreaterThan(0); - - AddComment(R"DOC( -Sequence Conv Operator. - -SequenceConvOp performs convolution operation on features of contextLength -time-steps of each instance. The convolution operation calculates the output -based on the input, filter, strides and paddings parameters. -The size of each dimension of the parameters is checked during infer-shape. -In order to ensure the equal length of sequence before and after convolution, -it is necessary to fill the top and bottom of each sequence based on -context_length, context_stride and context_start. - - )DOC"); - } -}; - -template -class SequenceConvGradOpMaker : public framework::SingleGradOpMaker { - public: - using framework::SingleGradOpMaker::SingleGradOpMaker; - - protected: - void Apply(GradOpPtr op) const override { - op->SetType("sequence_conv_grad"); - op->SetAttrMap(this->Attrs()); - - if (op->HasAttr("paddingTrainable") && - PADDLE_GET_CONST(bool, op->GetAttr("paddingTrainable")) && - this->HasInput("PaddingData")) { - op->SetInput("PaddingData", this->Input("PaddingData")); - op->SetOutput(framework::GradVarName("PaddingData"), - this->InputGrad("PaddingData")); - } - - op->SetInput("X", this->Input("X")); - op->SetInput("Filter", this->Input("Filter")); - op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out")); - - op->SetOutput(framework::GradVarName("X"), this->InputGrad("X")); - op->SetOutput(framework::GradVarName("Filter"), this->InputGrad("Filter")); - } -}; - -class SequenceConvGradNoNeedBufferVarsInference - : public framework::NoNeedBufferVarsInference { - public: - using framework::NoNeedBufferVarsInference::NoNeedBufferVarsInference; - - const std::unordered_set &operator()( - const framework::InferNoNeedBufferVarsContext &ctx) const final { - static const std::unordered_set kPaddingData({"PaddingData"}); - if (!PADDLE_GET_CONST(bool, ctx.GetAttr("paddingTrainable"))) { - return kPaddingData; - } else { - return Empty(); - } - } -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; -REGISTER_OPERATOR(sequence_conv, - ops::SequenceConvOp, - ops::SequenceConvOpMaker, - ops::SequenceConvGradOpMaker, - ops::SequenceConvGradOpMaker); - -REGISTER_OPERATOR(sequence_conv_grad, - ops::SequenceConvGradOp, - ops::SequenceConvGradNoNeedBufferVarsInference); - -PD_REGISTER_STRUCT_KERNEL( - sequence_conv, CPU, ALL_LAYOUT, ops::SequenceConvKernel, float, double) {} -PD_REGISTER_STRUCT_KERNEL(sequence_conv_grad, - CPU, - ALL_LAYOUT, - ops::SequenceConvGradKernel, - float, - double) {} diff --git a/paddle/fluid/operators/sequence_ops/sequence_conv_op.cu.cc b/paddle/fluid/operators/sequence_ops/sequence_conv_op.cu.cc deleted file mode 100644 index a96efea48557a..0000000000000 --- a/paddle/fluid/operators/sequence_ops/sequence_conv_op.cu.cc +++ /dev/null @@ -1,25 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#include "paddle/fluid/operators/sequence_ops/sequence_conv_op.h" - -namespace ops = paddle::operators; -PD_REGISTER_STRUCT_KERNEL( - sequence_conv, GPU, ALL_LAYOUT, ops::SequenceConvKernel, float, double) {} -PD_REGISTER_STRUCT_KERNEL(sequence_conv_grad, - GPU, - ALL_LAYOUT, - ops::SequenceConvGradKernel, - float, - double) {} diff --git a/paddle/fluid/operators/sequence_ops/sequence_conv_op.h b/paddle/fluid/operators/sequence_ops/sequence_conv_op.h deleted file mode 100644 index 347db6e37db10..0000000000000 --- a/paddle/fluid/operators/sequence_ops/sequence_conv_op.h +++ /dev/null @@ -1,204 +0,0 @@ -/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#pragma once -#include - -#include "paddle/fluid/framework/op_registry.h" -#include "paddle/fluid/operators/math/context_project.h" -#include "paddle/phi/kernels/funcs/math_function.h" - -namespace paddle { -namespace operators { - -template -class SequenceConvKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* in = context.Input("X"); - auto* out = context.Output("Out"); - auto filter = *context.Input("Filter"); - - out->mutable_data(context.GetPlace()); - - int context_start = context.Attr("contextStart"); - int context_length = context.Attr("contextLength"); - int context_stride = context.Attr("contextStride"); - bool padding_trainable = context.Attr("paddingTrainable"); - - PADDLE_ENFORCE_EQ(in->lod().empty(), - false, - platform::errors::InvalidArgument( - "Input(X) phi::DenseTensor of SequenceConvOp " - "does not contain LoD information.")); - PADDLE_ENFORCE_EQ( - in->lod().size(), - 1UL, - platform::errors::InvalidArgument( - "Only support input sequence with lod level equal to 1 at " - "present. But received: lod level %u.", - in->lod().size())); - - const phi::DenseTensor* padding_data = nullptr; - if (padding_trainable) { - padding_data = context.Input("PaddingData"); - } - - int up_pad = std::max(0, -context_start); - int down_pad = std::max(0, context_start + context_length - 1); - auto sequence_width = static_cast(in->dims()[1]); - - framework::DDim col_shape = {in->dims()[0], - context_length * sequence_width}; - phi::DenseTensor col; - col.mutable_data(col_shape, context.GetPlace()); - // Because if padding_trainable is false, padding data should be zeros. - phi::funcs::SetConstant set_zero; - auto& dev_ctx = context.template device_context(); - auto blas = phi::funcs::GetBlas(dev_ctx); - set_zero(dev_ctx, &col, static_cast(0)); - math::ContextProjectFunctor seq_project_functor; - - seq_project_functor(dev_ctx, - *in, - padding_data, - padding_trainable, - context_start, - context_length, - context_stride, - up_pad, - down_pad, - &col); - - blas.MatMul(col, filter, out); - } -}; - -template -class SequenceConvGradKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* in_g = context.Output(framework::GradVarName("X")); - auto* out_g = - context.Input(framework::GradVarName("Out")); - auto* filter_g = - context.Output(framework::GradVarName("Filter")); - auto* padding_data_g = - context.Output(framework::GradVarName("PaddingData")); - auto* in = context.Input("X"); - auto* filter = context.Input("Filter"); - - int context_start = context.Attr("contextStart"); - int context_length = context.Attr("contextLength"); - int context_stride = context.Attr("contextStride"); - bool padding_trainable = context.Attr("paddingTrainable"); - - PADDLE_ENFORCE_EQ( - in->lod().size(), - 1UL, - platform::errors::InvalidArgument( - "Only support input sequence with lod level equal to 1 at " - "present. But received: lod level %u.", - in->lod().size())); - auto lod_g_level_0 = in->lod()[0]; - - int up_pad = std::max(0, -context_start); - int down_pad = std::max(0, context_start + context_length - 1); - auto sequence_width = static_cast(in->dims()[1]); - - phi::funcs::SetConstant set_zero; - auto& dev_ctx = context.template device_context(); - auto blas = phi::funcs::GetBlas(dev_ctx); - // use col_shape in the im2col calculation - framework::DDim col_shape = {in->dims()[0], - sequence_width * context_length}; - phi::DenseTensor col; - - if (in_g || filter_g || (padding_trainable && padding_data_g)) { - col.mutable_data(col_shape, context.GetPlace()); - // Because if padding_trainable is false, padding data should be zeros. - set_zero(dev_ctx, &col, static_cast(0)); - blas.MatMul(*out_g, false, *filter, true, &col); - } - math::ContextProjectFunctor seq_project_functor; - math::ContextProjectGradFunctor seq_project_grad_functor; - - if (in_g) { - in_g->mutable_data(context.GetPlace()); - in_g->set_lod(in->lod()); - set_zero(dev_ctx, in_g, static_cast(0)); - - seq_project_grad_functor(dev_ctx, - *in_g, - padding_trainable, - context_start, - context_length, - context_stride, - up_pad, - down_pad, - false, - true, - padding_data_g, - &col); - } - - if (padding_trainable && padding_data_g) { - padding_data_g->mutable_data(context.GetPlace()); - set_zero(dev_ctx, padding_data_g, static_cast(0)); - - phi::DenseTensor* input = const_cast(in); - seq_project_grad_functor(dev_ctx, - *input, - padding_trainable, - context_start, - context_length, - context_stride, - up_pad, - down_pad, - true, - false, - padding_data_g, - &col); - } - - if (filter_g) { - filter_g->mutable_data(context.GetPlace()); - set_zero(dev_ctx, filter_g, static_cast(0)); - - phi::DenseTensor filter_grad = *filter_g; - phi::DenseTensor out_grad = *out_g; - - const phi::DenseTensor* padding_data = nullptr; - if (padding_trainable) { - padding_data = context.Input("PaddingData"); - } - - seq_project_functor(dev_ctx, - *in, - padding_data, - padding_trainable, - context_start, - context_length, - context_stride, - up_pad, - down_pad, - &col); - - blas.MatMul(col, true, out_grad, false, &filter_grad); - } - } -}; - -} // namespace operators -} // namespace paddle diff --git a/paddle/fluid/operators/sequence_ops/sequence_conv_op_xpu.cc b/paddle/fluid/operators/sequence_ops/sequence_conv_op_xpu.cc deleted file mode 100644 index 53fb13180c36a..0000000000000 --- a/paddle/fluid/operators/sequence_ops/sequence_conv_op_xpu.cc +++ /dev/null @@ -1,374 +0,0 @@ -/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved. - -Licensed under the Apache License, Version 2.0 (the "License"); -you may not use this file except in compliance with the License. -You may obtain a copy of the License at - - http://www.apache.org/licenses/LICENSE-2.0 - -Unless required by applicable law or agreed to in writing, software -distributed under the License is distributed on an "AS IS" BASIS, -WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -See the License for the specific language governing permissions and -limitations under the License. */ - -#ifdef PADDLE_WITH_XPU - -#include "paddle/fluid/operators/sequence_ops/sequence_conv_op.h" -#include "paddle/fluid/platform/device/device_wrapper.h" - -namespace paddle { -namespace operators { - -template -class SequenceConvXPUKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* in = context.Input("X"); - auto* out = context.Output("Out"); - auto filter = *context.Input("Filter"); - - out->mutable_data(context.GetPlace()); - - int context_start = context.Attr("contextStart"); - int context_length = context.Attr("contextLength"); - int context_stride = context.Attr("contextStride"); - bool padding_trainable = context.Attr("paddingTrainable"); - - PADDLE_ENFORCE_EQ(in->lod().empty(), - false, - platform::errors::InvalidArgument( - "Input(X) phi::DenseTensor of SequenceConvOp " - "does not contain LoD information.")); - PADDLE_ENFORCE_EQ( - in->lod().size(), - 1UL, - platform::errors::InvalidArgument( - "Only support input sequence with lod level equal to 1 at " - "present. But received: lod level %u.", - in->lod().size())); - - PADDLE_ENFORCE_EQ( - padding_trainable, - false, - platform::errors::InvalidArgument("Only support padding_trainable " - "equal false.")); - - int up_pad = std::max(0, -context_start); - int down_pad = std::max(0, context_start + context_length - 1); - PADDLE_ENFORCE_EQ( - up_pad, - 2, - platform::errors::InvalidArgument("Only support up_pad equal 2.")); - PADDLE_ENFORCE_EQ( - down_pad, - 2, - platform::errors::InvalidArgument("Only support down_pad equal 2.")); - - auto xpu_context = - context.template device_context().x_context(); - auto sequence_width = static_cast(in->dims()[1]); - framework::DDim col_shape = {in->dims()[0], - context_length * sequence_width}; - xpu::ctx_guard RAII_GUARD(xpu_context); - int col_numel = col_shape[0] * col_shape[1]; - T* col_data = RAII_GUARD.alloc_l3_or_gm(col_numel); - PADDLE_ENFORCE_NOT_NULL( - col_data, paddle::platform::errors::Fatal("XPU memory is not enough")); - - auto lod_level_0 = in->lod()[0]; - int lod_size = lod_level_0.size(); - // If batch size set to 256, the lod is {0, batch[0] - 0, - // batch[1] - batch [0], ..., batch[255] - batch[254]}, - // so the lod_size will be 257. - PADDLE_ENFORCE_LE( - lod_size, - 257, - platform::errors::InvalidArgument("Only support batch size <= 256.")); - - std::vector cpu_lodx(lod_size); - for (int i = 0; i < lod_size; i++) { - cpu_lodx[i] = lod_level_0[i]; - } - xpu::VectorParam lodx = { - cpu_lodx.data(), static_cast(cpu_lodx.size()), nullptr}; - - int r = xpu::sequence_context_projection(xpu_context, - in->data(), - col_data, - nullptr, - lodx, - sequence_width, - context_start, - context_length, - context_stride, - {2, 2}); - PADDLE_ENFORCE_XDNN_SUCCESS(r, "sequence_context_projection"); - - bool trans_a = false; - bool trans_b = false; - int m = col_shape[0]; - int k = col_shape[1]; - int k1 = filter.dims()[0]; - int n = filter.dims()[1]; - PADDLE_ENFORCE_EQ(k, - k1, - platform::errors::InvalidArgument( - "The shape of FC in SequenceConvOp is invalid." - "The k of matrix A is %d, k1 of matrix B is %d." - "But expect k == k1", - k, - k1)); - int lda = (!trans_a) ? k : m; - int ldb = (!trans_b) ? n : k; - int ldc = n; - T alpha = static_cast(1.0); - T beta = static_cast(0.0); - const T* data_a = col_data; - const T* data_b = filter.data(); - T* data_c = out->data(); - - r = xpu::fc_fusion(xpu_context, - data_a, - data_b, - data_c, - m, - n, - k, - trans_a, - trans_b, - nullptr, - nullptr, - nullptr, - lda, - ldb, - ldc, - alpha, - beta, - nullptr, - xpu::Activation_t::LINEAR); - PADDLE_ENFORCE_XDNN_SUCCESS(r, "fc_fusion"); - if (xpu_context->xpu_stream != nullptr) { - xpu_wait(xpu_context->xpu_stream); - } - } -}; - -template -class SequenceConvGradXPUKernel : public framework::OpKernel { - public: - void Compute(const framework::ExecutionContext& context) const override { - auto* in_g = context.Output(framework::GradVarName("X")); - auto* out_g = - context.Input(framework::GradVarName("Out")); - auto* filter_g = - context.Output(framework::GradVarName("Filter")); - auto* in = context.Input("X"); - auto* filter = context.Input("Filter"); - - int context_start = context.Attr("contextStart"); - int context_length = context.Attr("contextLength"); - int context_stride = context.Attr("contextStride"); - bool padding_trainable = context.Attr("paddingTrainable"); - - PADDLE_ENFORCE_EQ(in->lod().empty(), - false, - platform::errors::InvalidArgument( - "Input(X) phi::DenseTensor of SequenceConvOp " - "does not contain LoD information.")); - PADDLE_ENFORCE_EQ( - in->lod().size(), - 1UL, - platform::errors::InvalidArgument( - "Only support input sequence with lod level equal to 1 at " - "present. But received: lod level %u.", - in->lod().size())); - - PADDLE_ENFORCE_EQ( - padding_trainable, - false, - platform::errors::InvalidArgument("Only support padding_trainable " - "equal false.")); - - int up_pad = std::max(0, -context_start); - int down_pad = std::max(0, context_start + context_length - 1); - PADDLE_ENFORCE_EQ( - up_pad, - 2, - platform::errors::InvalidArgument("Only support up_pad equal 2.")); - PADDLE_ENFORCE_EQ( - down_pad, - 2, - platform::errors::InvalidArgument("Only support down_pad equal 2.")); - - auto lod_level_0 = in->lod()[0]; - int lod_size = lod_level_0.size(); - PADDLE_ENFORCE_LE( - lod_size, - 257, - platform::errors::InvalidArgument("Only support batch size <= 256.")); - - std::vector cpu_lodx(lod_size); - for (int i = 0; i < lod_size; i++) { - cpu_lodx[i] = lod_level_0[i]; - } - xpu::VectorParam lodx = { - cpu_lodx.data(), static_cast(cpu_lodx.size()), nullptr}; - - auto xpu_context = - context.template device_context().x_context(); - auto sequence_width = static_cast(in->dims()[1]); - framework::DDim col_shape = {in->dims()[0], - context_length * sequence_width}; - xpu::ctx_guard RAII_GUARD(xpu_context); - int col_numel = col_shape[0] * col_shape[1]; - T* col_data = RAII_GUARD.alloc_l3_or_gm(col_numel); - PADDLE_ENFORCE_NOT_NULL( - col_data, paddle::platform::errors::Fatal("XPU memory is not enough")); - - if (in_g || filter_g) { - bool trans_a = false; - bool trans_b = true; - int m = out_g->dims()[0]; - int k = out_g->dims()[1]; - int n = filter->dims()[0]; - int k1 = filter->dims()[1]; - PADDLE_ENFORCE_EQ(k, - k1, - platform::errors::InvalidArgument( - "The shape of FC in SequenceConvGradOp is invalid." - "The k of matrix A is %d, k1 of matrix B is %d." - "But expect k == k1", - k, - k1)); - int lda = (!trans_a) ? k : m; - int ldb = (!trans_b) ? n : k; - int ldc = n; - T alpha = static_cast(1.0); - T beta = static_cast(0.0); - const T* data_a = out_g->data(); - const T* data_b = filter->data(); - T* data_c = col_data; - - int r = xpu::fc_fusion(xpu_context, - data_a, - data_b, - data_c, - m, - n, - k, - trans_a, - trans_b, - nullptr, - nullptr, - nullptr, - lda, - ldb, - ldc, - alpha, - beta, - nullptr, - xpu::Activation_t::LINEAR); - PADDLE_ENFORCE_XDNN_SUCCESS(r, "fc_fusion"); - } - - if (in_g) { - PADDLE_ENFORCE_LT(sequence_width, - 512, - platform::errors::InvalidArgument( - "Only support sequence_width < 512.")); - - in_g->mutable_data(context.GetPlace()); - in_g->set_lod(in->lod()); - - int r = xpu::sequence_context_projection_grad(xpu_context, - in_g->data(), - col_data, - nullptr, - lodx, - sequence_width, - context_start, - context_length, - context_stride, - {2, 2}); - PADDLE_ENFORCE_XDNN_SUCCESS(r, "sequence_context_projection_grad"); - } - - if (filter_g) { - filter_g->mutable_data(context.GetPlace()); - - int r = xpu::sequence_context_projection(xpu_context, - in->data(), - col_data, - nullptr, - lodx, - sequence_width, - context_start, - context_length, - context_stride, - {2, 2}); - PADDLE_ENFORCE_XDNN_SUCCESS(r, "sequence_context_projection"); - - bool trans_a = true; - bool trans_b = false; - int k = col_shape[0]; - int m = col_shape[1]; - int k1 = out_g->dims()[0]; - int n = out_g->dims()[1]; - PADDLE_ENFORCE_EQ(k, - k1, - platform::errors::InvalidArgument( - "The shape of FC in SequenceConvGradOp is invalid." - "The k of matrix A is %d, k1 of matrix B is %d." - "But expect k == k1", - k, - k1)); - int lda = (!trans_a) ? k : m; - int ldb = (!trans_b) ? n : k; - int ldc = n; - T alpha = static_cast(1.0); - T beta = static_cast(0.0); - const T* data_a = col_data; - const T* data_b = out_g->data(); - T* data_c = filter_g->data(); - - r = xpu::fc_fusion(xpu_context, - data_a, - data_b, - data_c, - m, - n, - k, - trans_a, - trans_b, - nullptr, - nullptr, - nullptr, - lda, - ldb, - ldc, - alpha, - beta, - nullptr, - xpu::Activation_t::LINEAR); - PADDLE_ENFORCE_XDNN_SUCCESS(r, "fc_fusion"); - if (xpu_context->xpu_stream != nullptr) { - xpu_wait(xpu_context->xpu_stream); - } - } - } -}; - -} // namespace operators -} // namespace paddle - -namespace ops = paddle::operators; -PD_REGISTER_STRUCT_KERNEL( - sequence_conv, XPU, ALL_LAYOUT, ops::SequenceConvXPUKernel, float) {} -PD_REGISTER_STRUCT_KERNEL(sequence_conv_grad, - XPU, - ALL_LAYOUT, - ops::SequenceConvGradXPUKernel, - float) {} - -#endif diff --git a/python/paddle/static/nn/__init__.py b/python/paddle/static/nn/__init__.py index 7d08a6eff11bf..d05631a7b690b 100755 --- a/python/paddle/static/nn/__init__.py +++ b/python/paddle/static/nn/__init__.py @@ -38,7 +38,6 @@ from .loss import nce from .sequence_lod import ( sequence_concat, - sequence_conv, sequence_enumerate, sequence_expand, sequence_expand_as, @@ -80,7 +79,6 @@ 'switch_case', 'while_loop', 'sparse_embedding', - 'sequence_conv', 'sequence_softmax', 'sequence_pool', 'sequence_concat', diff --git a/python/paddle/static/nn/sequence_lod.py b/python/paddle/static/nn/sequence_lod.py index 51cfd0d5307d5..f8cabc30521f3 100644 --- a/python/paddle/static/nn/sequence_lod.py +++ b/python/paddle/static/nn/sequence_lod.py @@ -22,147 +22,6 @@ __all__ = [] -@templatedoc() -def sequence_conv( - input, - num_filters, - filter_size=3, - filter_stride=1, - padding=True, - padding_start=None, - bias_attr=None, - param_attr=None, - act=None, - name=None, -): - r""" - - Note: - Only receives Tensor as input. If your input is Tensor, please use conv2d Op.(base.layers.** :ref:`api_paddle_nn_functional_conv2d` ). - - This operator receives input sequences with variable length and other convolutional - configuration parameters(num_filters, filter_size) to apply the convolution operation. - It fills all-zero padding data on both sides of the sequence by default to ensure that - the output is the same length as the input. You can customize the padding behavior by - configuring the parameter :attr:`padding\_start` . - - **Warning:** the parameter :attr:`padding` take no effect and will be deprecated in the future. - - .. code-block:: text - - Here we will illustrate the details of the padding operation: - For a mini-batch of 2 variable lengths sentences, containing 3, and 1 time-steps: - Assumed input (X) is a [4, N] float Tensor, and for the sake of simplicity, we assume N=2. - input.data = [[1, 1], - [2, 2], - [3, 3], - [4, 4]] - - This is to say that input (X) has 4 words and the dimension of each word - representation is 2. - - * Case1: - - If padding_start is -1 and filter_size is 3. - The length of padding data is calculated as follows: - up_pad_len = max(0, -padding_start) = 1 - down_pad_len = max(0, filter_size + padding_start - 1) = 1 - - The output of the input sequence after padding is: - data_after_padding = [[0, 0, 1, 1, 2, 2], - [1, 1, 2, 2, 3, 3], - [2, 2, 3, 3, 0, 0], - [0, 0, 4, 4, 0, 0]] - - It will be multiplied by the filter weight to get the final output. - Assume num_filters = 3 - output.data = [[ 0.3234, -0.2334, 0.7433], - [ 0.5646, 0.9464, -0.1223], - [-0.1343, 0.5653, 0.4555], - [ 0.9954, -0.1234, -0.1234]] - output.shape = [4, 3] # 3 = num_filters - output.lod = [[0, 3, 4]] # Remain the same - - - Args: - input (Tensor): Tensor with shape :math:`(M, K)`, where M is the total time-step of mini-batch - and K is hidden_size of input. Only lod_level of 1 is supported. The data type should be float32 or - float64. - num_filters (int): the number of filters. - filter_size (int): the height of filter. Specified filter width is not supported, the width is - hidden_size by default. Default: 3. - filter_stride (int, optional): stride of the filter. Currently only supports :attr:`stride` = 1. - padding (bool, optional): the parameter :attr:`padding` take no effect and will be discarded in the - future. Currently, it will always pad input to make sure the length of the output is - the same as input whether :attr:`padding` is set true or false. Because the length of - input sequence may be shorter than :attr:`filter\_size`, which will cause the convolution - result to not be computed correctly. These padding data will not be trainable or updated - while training. Default: True. - padding_start (int): It is used to indicate the start index for padding the input - sequence, which can be negative. The negative number means to pad - :attr:`|padding_start|` time-steps of all-zero data at the beginning of each instance. - The positive number means to skip :attr:`padding_start` time-steps of each instance, - and it will pad :math:`filter\_size + padding\_start - 1` time-steps of all-zero data - at the end of the sequence to ensure that the output is the same length as the input. - If set None, the same length :math:`\\frac{filter\_size}{2}` of data will be filled - on both sides of the sequence. If set 0, the length of :math:`filter\_size - 1` data - is padded at the end of each input sequence. Default: None. - bias_attr (ParamAttr): To specify the bias parameter property. Default: None, which means the - default bias parameter property is used. See usage for details in :ref:`api_paddle_ParamAttr` . - param_attr (ParamAttr): To specify the weight parameter property. Default: None, which means the - default weight parameter property is used. See usage for details in :ref:`api_paddle_ParamAttr` . - act (str): Activation to be applied to the output of this layer, such as tanh, softmax, - sigmoid, relu. For more information, please refer to :ref:`api_guide_activations_en` . Default: None. - name (str, optional): The default value is None. Normally there is no need for user to set this property. - For more information, please refer to :ref:`api_guide_Name` . - - Returns: - Tensor: Tensor with the same length as input. The data type is float32 or float64, which is same as input. - - Examples: - - .. code-block:: python - - >>> import paddle - >>> paddle.enable_static() - - >>> x = paddle.static.data(name='x', shape=[-1, 10], dtype='float32', lod_level=1) - >>> x_conved = paddle.static.nn.sequence_conv(input=x, num_filters=2, filter_size=3, padding_start=-1) - """ - - assert ( - not in_dygraph_mode() - ), "sequence layer is not supported in dygraph mode yet." - check_variable_and_dtype( - input, 'input', ['float32', 'float64'], 'sequence_conv' - ) - helper = LayerHelper('sequence_conv', **locals()) - dtype = helper.input_dtype() - filter_shape = [filter_size * input.shape[1], num_filters] - filter_param = helper.create_parameter( - attr=helper.param_attr, shape=filter_shape, dtype=dtype - ) - pre_bias = helper.create_variable_for_type_inference(dtype) - if padding_start is None: - padding_start = -int(filter_size // 2) - - helper.append_op( - type='sequence_conv', - inputs={ - 'X': [input], - 'Filter': [filter_param], - }, - outputs={"Out": pre_bias}, - attrs={ - 'contextStride': filter_stride, - 'contextStart': padding_start, - 'contextLength': filter_size, - }, - ) - pre_act = helper.append_bias_op(pre_bias) - return helper.append_activation(pre_act) - - def sequence_softmax(input, use_cudnn=False, name=None): r""" diff --git a/python/paddle/static/quantization/quant2_int8_onednn_pass.py b/python/paddle/static/quantization/quant2_int8_onednn_pass.py index 19fb5f928bea5..16827447dec7f 100644 --- a/python/paddle/static/quantization/quant2_int8_onednn_pass.py +++ b/python/paddle/static/quantization/quant2_int8_onednn_pass.py @@ -435,7 +435,6 @@ def _optimize_fp32_graph(self, graph): graph = self._apply_pass(graph, 'simplify_with_basic_ops_pass') graph = self._apply_pass(graph, 'layer_norm_fuse_pass') graph = self._apply_pass(graph, 'attention_lstm_fuse_pass') - graph = self._apply_pass(graph, 'seqconv_eltadd_relu_fuse_pass') graph = self._apply_pass(graph, 'fc_lstm_fuse_pass') graph = self._apply_pass(graph, 'mul_lstm_fuse_pass') graph = self._apply_pass(graph, 'fc_gru_fuse_pass') diff --git a/test/book/notest_understand_sentiment.py b/test/book/notest_understand_sentiment.py deleted file mode 100644 index 250b3f2dc679c..0000000000000 --- a/test/book/notest_understand_sentiment.py +++ /dev/null @@ -1,266 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import contextlib -import math -import os -import sys -import unittest - -import numpy as np - -# TODO: remove sys.path.append -sys.path.append("../legacy_test") -import nets - -import paddle -from paddle import base - - -def convolution_net( - data, label, input_dim, class_dim=2, emb_dim=32, hid_dim=32 -): - emb = paddle.static.nn.embedding( - input=data, size=[input_dim, emb_dim], is_sparse=True - ) - conv_3 = nets.sequence_conv_pool( - input=emb, - num_filters=hid_dim, - filter_size=3, - act="tanh", - pool_type="sqrt", - ) - conv_4 = nets.sequence_conv_pool( - input=emb, - num_filters=hid_dim, - filter_size=4, - act="tanh", - pool_type="sqrt", - ) - prediction = paddle.static.nn.fc( - x=[conv_3, conv_4], size=class_dim, activation="softmax" - ) - cost = paddle.nn.functional.cross_entropy( - input=prediction, label=label, reduction='none', use_softmax=False - ) - avg_cost = paddle.mean(cost) - accuracy = paddle.static.accuracy(input=prediction, label=label) - return avg_cost, accuracy, prediction - - -def train( - word_dict, - net_method, - use_cuda, - parallel=False, - save_dirname=None, - is_local=True, -): - BATCH_SIZE = 128 - PASS_NUM = 5 - dict_dim = len(word_dict) - class_dim = 2 - - data = paddle.static.data( - name="words", shape=[-1, 1], dtype="int64", lod_level=1 - ) - label = paddle.static.data(name="label", shape=[-1, 1], dtype="int64") - - if not parallel: - cost, acc_out, prediction = net_method( - data, label, input_dim=dict_dim, class_dim=class_dim - ) - else: - raise NotImplementedError() - - adagrad = paddle.optimizer.Adagrad(learning_rate=0.002) - adagrad.minimize(cost) - - train_data = paddle.batch( - paddle.reader.shuffle( - paddle.dataset.imdb.train(word_dict), buf_size=1000 - ), - batch_size=BATCH_SIZE, - ) - place = base.CUDAPlace(0) if use_cuda else base.CPUPlace() - exe = base.Executor(place) - feeder = base.DataFeeder(feed_list=[data, label], place=place) - - def train_loop(main_program): - exe.run(base.default_startup_program()) - - for pass_id in range(PASS_NUM): - for data in train_data(): - cost_val, acc_val = exe.run( - main_program, - feed=feeder.feed(data), - fetch_list=[cost, acc_out], - ) - print("cost=" + str(cost_val) + " acc=" + str(acc_val)) - if cost_val < 0.4 and acc_val > 0.8: - if save_dirname is not None: - paddle.static.io.save_inference_model( - save_dirname, data, prediction, exe - ) - return - if math.isnan(float(cost_val)): - sys.exit("got NaN loss, training failed.") - raise AssertionError(f"Cost is too large for {net_method.__name__}") - - if is_local: - train_loop(base.default_main_program()) - else: - port = os.getenv("PADDLE_PSERVER_PORT", "6174") - pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... - eplist = [] - for ip in pserver_ips.split(","): - eplist.append(':'.join([ip, port])) - pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("PADDLE_TRAINERS")) - current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) - training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") - t = paddle.distributed.transpiler.DistributeTranspiler() - t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) - if training_role == "PSERVER": - pserver_prog = t.get_pserver_program(current_endpoint) - pserver_startup = t.get_startup_program( - current_endpoint, pserver_prog - ) - exe.run(pserver_startup) - exe.run(pserver_prog) - elif training_role == "TRAINER": - train_loop(t.get_trainer_program()) - - -def infer(word_dict, use_cuda, save_dirname=None): - if save_dirname is None: - return - - place = base.CUDAPlace(0) if use_cuda else base.CPUPlace() - exe = base.Executor(place) - - inference_scope = base.core.Scope() - with base.scope_guard(inference_scope): - # Use paddle.static.io.load_inference_model to obtain the inference program desc, - # the feed_target_names (the names of variables that will be fed - # data using feed operators), and the fetch_targets (variables that - # we want to obtain data from using fetch operators). - [ - inference_program, - feed_target_names, - fetch_targets, - ] = paddle.static.io.load_inference_model(save_dirname, exe) - - word_dict_len = len(word_dict) - - # Setup input by creating LoDTensor to represent sequence of words. - # Here each word is the basic element of the LoDTensor and the shape of - # each word (base_shape) should be [1] since it is simply an index to - # look up for the corresponding word vector. - # Suppose the recursive_sequence_lengths info is set to [[3, 4, 2]], - # which has only one level of detail. Then the created LoDTensor will have only - # one higher level structure (sequence of words, or sentence) than the basic - # element (word). Hence the LoDTensor will hold data for three sentences of - # length 3, 4 and 2, respectively. - # Note that recursive_sequence_lengths should be a list of lists. - recursive_seq_lens = [[3, 4, 2]] - base_shape = [1] - # The range of random integers is [low, high] - tensor_words = base.create_random_int_lodtensor( - recursive_seq_lens, base_shape, place, low=0, high=word_dict_len - 1 - ) - - # Construct feed as a dictionary of {feed_target_name: feed_target_data} - # and results will contain a list of data corresponding to fetch_targets. - assert feed_target_names[0] == "words" - results = exe.run( - inference_program, - feed={feed_target_names[0]: tensor_words}, - fetch_list=fetch_targets, - return_numpy=False, - ) - print(results[0].recursive_sequence_lengths()) - np_data = np.array(results[0]) - print("Inference Shape: ", np_data.shape) - print("Inference results: ", np_data) - - -def main(word_dict, net_method, use_cuda, parallel=False, save_dirname=None): - if use_cuda and not base.core.is_compiled_with_cuda(): - return - - train( - word_dict, - net_method, - use_cuda, - parallel=parallel, - save_dirname=save_dirname, - ) - infer(word_dict, use_cuda, save_dirname) - - -class TestUnderstandSentiment(unittest.TestCase): - @classmethod - def setUpClass(cls): - cls.word_dict = paddle.dataset.imdb.word_dict() - - @contextlib.contextmanager - def new_program_scope(self): - prog = base.Program() - startup_prog = base.Program() - scope = base.core.Scope() - with base.scope_guard(scope): - with base.program_guard(prog, startup_prog): - yield - - def test_conv_cpu(self): - with self.new_program_scope(): - main( - self.word_dict, - net_method=convolution_net, - use_cuda=False, - save_dirname="understand_sentiment_conv.inference.model", - ) - - def test_conv_cpu_parallel(self): - with self.new_program_scope(): - main( - self.word_dict, - net_method=convolution_net, - use_cuda=False, - parallel=True, - ) - - def test_conv_gpu(self): - with self.new_program_scope(): - main( - self.word_dict, - net_method=convolution_net, - use_cuda=True, - save_dirname="understand_sentiment_conv.inference.model", - ) - - def test_conv_gpu_parallel(self): - with self.new_program_scope(): - main( - self.word_dict, - net_method=convolution_net, - use_cuda=True, - parallel=True, - ) - - -if __name__ == '__main__': - unittest.main() diff --git a/test/book/test_recommender_system.py b/test/book/test_recommender_system.py deleted file mode 100644 index 7a4a70be105d5..0000000000000 --- a/test/book/test_recommender_system.py +++ /dev/null @@ -1,392 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import math -import os -import sys -import tempfile - -import numpy as np - -# TODO: remove sys.path.append -sys.path.append("../legacy_test") -import nets - -import paddle -from paddle import base -from paddle.base import framework -from paddle.base.executor import Executor -from paddle.optimizer import SGD - -paddle.enable_static() - -IS_SPARSE = True -USE_GPU = False -BATCH_SIZE = 256 - - -def get_usr_combined_features(): - # FIXME(dzh) : old API integer_value(10) may has range check. - # currently we don't have user configurated check. - - USR_DICT_SIZE = paddle.dataset.movielens.max_user_id() + 1 - - uid = paddle.static.data(name='user_id', shape=[-1, 1], dtype='int64') - - usr_emb = paddle.static.nn.embedding( - input=uid, - dtype='float32', - size=[USR_DICT_SIZE, 32], - param_attr='user_table', - is_sparse=IS_SPARSE, - ) - - usr_fc = paddle.static.nn.fc(x=usr_emb, size=32) - - USR_GENDER_DICT_SIZE = 2 - - usr_gender_id = paddle.static.data( - name='gender_id', shape=[-1, 1], dtype='int64' - ) - - usr_gender_emb = paddle.static.nn.embedding( - input=usr_gender_id, - size=[USR_GENDER_DICT_SIZE, 16], - param_attr='gender_table', - is_sparse=IS_SPARSE, - ) - - usr_gender_fc = paddle.static.nn.fc(x=usr_gender_emb, size=16) - - USR_AGE_DICT_SIZE = len(paddle.dataset.movielens.age_table) - usr_age_id = paddle.static.data(name='age_id', shape=[-1, 1], dtype="int64") - - usr_age_emb = paddle.static.nn.embedding( - input=usr_age_id, - size=[USR_AGE_DICT_SIZE, 16], - is_sparse=IS_SPARSE, - param_attr='age_table', - ) - - usr_age_fc = paddle.static.nn.fc(x=usr_age_emb, size=16) - - USR_JOB_DICT_SIZE = paddle.dataset.movielens.max_job_id() + 1 - usr_job_id = paddle.static.data(name='job_id', shape=[-1, 1], dtype="int64") - - usr_job_emb = paddle.static.nn.embedding( - input=usr_job_id, - size=[USR_JOB_DICT_SIZE, 16], - param_attr='job_table', - is_sparse=IS_SPARSE, - ) - - usr_job_fc = paddle.static.nn.fc(x=usr_job_emb, size=16) - - concat_embed = paddle.concat( - [usr_fc, usr_gender_fc, usr_age_fc, usr_job_fc], axis=1 - ) - - usr_combined_features = paddle.static.nn.fc( - x=concat_embed, size=200, activation="tanh" - ) - - return usr_combined_features - - -def get_mov_combined_features(): - MOV_DICT_SIZE = paddle.dataset.movielens.max_movie_id() + 1 - - mov_id = paddle.static.data(name='movie_id', shape=[-1, 1], dtype='int64') - - mov_emb = paddle.static.nn.embedding( - input=mov_id, - dtype='float32', - size=[MOV_DICT_SIZE, 32], - param_attr='movie_table', - is_sparse=IS_SPARSE, - ) - - mov_fc = paddle.static.nn.fc(x=mov_emb, size=32) - - CATEGORY_DICT_SIZE = len(paddle.dataset.movielens.movie_categories()) - - category_id = paddle.static.data( - name='category_id', shape=[-1, 1], dtype='int64', lod_level=1 - ) - - mov_categories_emb = paddle.static.nn.embedding( - input=category_id, size=[CATEGORY_DICT_SIZE, 32], is_sparse=IS_SPARSE - ) - - mov_categories_hidden = paddle.static.nn.sequence_lod.sequence_pool( - input=mov_categories_emb.squeeze(-2), pool_type="sum" - ) - - MOV_TITLE_DICT_SIZE = len(paddle.dataset.movielens.get_movie_title_dict()) - - mov_title_id = paddle.static.data( - name='movie_title', shape=[-1, 1], dtype='int64', lod_level=1 - ) - - mov_title_emb = paddle.static.nn.embedding( - input=mov_title_id, size=[MOV_TITLE_DICT_SIZE, 32], is_sparse=IS_SPARSE - ) - - mov_title_conv = nets.sequence_conv_pool( - input=mov_title_emb.squeeze(-2), - num_filters=32, - filter_size=3, - act="tanh", - pool_type="sum", - ) - - concat_embed = paddle.concat( - [mov_fc, mov_categories_hidden, mov_title_conv], axis=1 - ) - - # FIXME(dzh) : need tanh operator - mov_combined_features = paddle.static.nn.fc( - x=concat_embed, size=200, activation="tanh" - ) - - return mov_combined_features - - -def model(): - usr_combined_features = get_usr_combined_features() - mov_combined_features = get_mov_combined_features() - - # need cos sim - inference = paddle.nn.functional.cosine_similarity( - x1=usr_combined_features, x2=mov_combined_features - ) - scale_infer = paddle.scale(x=inference, scale=5.0) - - label = paddle.static.data(name='score', shape=[-1, 1], dtype='float32') - square_cost = paddle.nn.functional.square_error_cost( - input=scale_infer, label=label - ) - avg_cost = paddle.mean(square_cost) - - return scale_infer, avg_cost - - -def train(use_cuda, save_dirname, is_local=True): - scale_infer, avg_cost = model() - - # test program - test_program = base.default_main_program().clone(for_test=True) - - sgd_optimizer = SGD(learning_rate=0.2) - sgd_optimizer.minimize(avg_cost) - - place = base.CUDAPlace(0) if use_cuda else base.CPUPlace() - - exe = Executor(place) - - train_reader = paddle.batch( - paddle.reader.shuffle(paddle.dataset.movielens.train(), buf_size=8192), - batch_size=BATCH_SIZE, - ) - test_reader = paddle.batch( - paddle.dataset.movielens.test(), batch_size=BATCH_SIZE - ) - - feed_order = [ - 'user_id', - 'gender_id', - 'age_id', - 'job_id', - 'movie_id', - 'category_id', - 'movie_title', - 'score', - ] - feed_infer_order = [ - 'user_id', - 'gender_id', - 'age_id', - 'job_id', - 'movie_id', - 'category_id', - 'movie_title', - ] - - def train_loop(main_program): - exe.run(framework.default_startup_program()) - - feed_list = [ - main_program.global_block().var(var_name) for var_name in feed_order - ] - feed_infer_list = [ - main_program.global_block().var(var_name) - for var_name in feed_infer_order - ] - feeder = base.DataFeeder(feed_list, place) - - PASS_NUM = 100 - for pass_id in range(PASS_NUM): - for batch_id, data in enumerate(train_reader()): - # train a mini-batch - outs = exe.run( - program=main_program, - feed=feeder.feed(data), - fetch_list=[avg_cost], - ) - out = np.array(outs[0]) - if (batch_id + 1) % 10 == 0: - avg_cost_set = [] - for test_data in test_reader(): - avg_cost_np = exe.run( - program=test_program, - feed=feeder.feed(test_data), - fetch_list=[avg_cost], - ) - avg_cost_set.append(avg_cost_np[0]) - break # test only 1 segment for speeding up CI - - # get test avg_cost - test_avg_cost = np.array(avg_cost_set).mean() - if test_avg_cost < 6.0: - # if avg_cost less than 6.0, we think our code is good. - if save_dirname is not None: - paddle.static.io.save_inference_model( - save_dirname, - feed_infer_list, - [scale_infer], - exe, - ) - return - - if math.isnan(float(out)): - sys.exit("got NaN loss, training failed.") - - if is_local: - train_loop(base.default_main_program()) - else: - port = os.getenv("PADDLE_PSERVER_PORT", "6174") - pserver_ips = os.getenv("PADDLE_PSERVER_IPS") # ip,ip... - eplist = [] - for ip in pserver_ips.split(","): - eplist.append(':'.join([ip, port])) - pserver_endpoints = ",".join(eplist) # ip:port,ip:port... - trainers = int(os.getenv("PADDLE_TRAINERS")) - current_endpoint = os.getenv("POD_IP") + ":" + port - trainer_id = int(os.getenv("PADDLE_TRAINER_ID")) - training_role = os.getenv("PADDLE_TRAINING_ROLE", "TRAINER") - t = paddle.distributed.transpiler.DistributeTranspiler() - t.transpile(trainer_id, pservers=pserver_endpoints, trainers=trainers) - if training_role == "PSERVER": - pserver_prog = t.get_pserver_program(current_endpoint) - pserver_startup = t.get_startup_program( - current_endpoint, pserver_prog - ) - exe.run(pserver_startup) - exe.run(pserver_prog) - elif training_role == "TRAINER": - train_loop(t.get_trainer_program()) - - -def infer(use_cuda, save_dirname=None): - if save_dirname is None: - return - - place = base.CUDAPlace(0) if use_cuda else base.CPUPlace() - exe = base.Executor(place) - - inference_scope = base.core.Scope() - with base.scope_guard(inference_scope): - # Use paddle.static.io.load_inference_model to obtain the inference program desc, - # the feed_target_names (the names of variables that will be fed - # data using feed operators), and the fetch_targets (variables that - # we want to obtain data from using fetch operators). - [ - inference_program, - feed_target_names, - fetch_targets, - ] = paddle.static.io.load_inference_model(save_dirname, exe) - - # Use the first data from paddle.dataset.movielens.test() as input - assert feed_target_names[0] == "user_id" - # Use create_lod_tensor(data, recursive_sequence_lengths, place) API - # to generate LoD Tensor where `data` is a list of sequences of index - # numbers, `recursive_sequence_lengths` is the length-based level of detail - # (lod) info associated with `data`. - # For example, data = [[10, 2, 3], [2, 3]] means that it contains - # two sequences of indexes, of length 3 and 2, respectively. - # Correspondingly, recursive_sequence_lengths = [[3, 2]] contains one - # level of detail info, indicating that `data` consists of two sequences - # of length 3 and 2, respectively. - user_id = base.create_lod_tensor([[np.int64(1)]], [[1]], place) - - assert feed_target_names[1] == "gender_id" - gender_id = base.create_lod_tensor([[np.int64(1)]], [[1]], place) - - assert feed_target_names[2] == "age_id" - age_id = base.create_lod_tensor([[np.int64(0)]], [[1]], place) - - assert feed_target_names[3] == "job_id" - job_id = base.create_lod_tensor([[np.int64(10)]], [[1]], place) - - assert feed_target_names[4] == "movie_id" - movie_id = base.create_lod_tensor([[np.int64(783)]], [[1]], place) - - assert feed_target_names[5] == "category_id" - category_id = base.create_lod_tensor( - [np.array([10, 8, 9], dtype='int64')], [[3]], place - ) - - assert feed_target_names[6] == "movie_title" - movie_title = base.create_lod_tensor( - [np.array([1069, 4140, 2923, 710, 988], dtype='int64')], - [[5]], - place, - ) - - # Construct feed as a dictionary of {feed_target_name: feed_target_data} - # and results will contain a list of data corresponding to fetch_targets. - results = exe.run( - inference_program, - feed={ - feed_target_names[0]: user_id, - feed_target_names[1]: gender_id, - feed_target_names[2]: age_id, - feed_target_names[3]: job_id, - feed_target_names[4]: movie_id, - feed_target_names[5]: category_id, - feed_target_names[6]: movie_title, - }, - fetch_list=fetch_targets, - return_numpy=False, - ) - print("inferred score: ", np.array(results[0])) - - -def main(use_cuda): - if use_cuda and not base.core.is_compiled_with_cuda(): - return - - # Directory for saving the inference model - temp_dir = tempfile.TemporaryDirectory() - save_dirname = os.path.join( - temp_dir.name, "recommender_system.inference.model" - ) - - train(use_cuda, save_dirname) - infer(use_cuda, save_dirname) - temp_dir.cleanup() - - -if __name__ == '__main__': - main(USE_GPU) diff --git a/test/ir/inference/test_seqconv_eltadd_relu_fuse_pass.py b/test/ir/inference/test_seqconv_eltadd_relu_fuse_pass.py deleted file mode 100644 index b31533ac958d0..0000000000000 --- a/test/ir/inference/test_seqconv_eltadd_relu_fuse_pass.py +++ /dev/null @@ -1,120 +0,0 @@ -# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import unittest -from functools import partial - -import hypothesis.strategies as st -import numpy as np -from auto_scan_test import PassAutoScanTest -from program_config import OpConfig, ProgramConfig, TensorConfig - - -class TestSeqconvEltaddReluFusePass(PassAutoScanTest): - def is_program_valid(self, program_config: ProgramConfig) -> bool: - return True - - def sample_program_config(self, draw): - contextLength = draw(st.sampled_from([1, 2, 3, 4])) - contextStart = draw(st.sampled_from([1, 2, 3])) - contextStride = draw(st.sampled_from([1])) - paddingTrainable = False - axis = draw(st.sampled_from([1])) - batch_size = draw(st.integers(min_value=1, max_value=4)) - - def generate_input(): - shape = [batch_size, 128, 6, 120] - return np.random.random(shape).astype(np.float32) - - def generate_weight(shape): - return np.random.random(shape).astype(np.float32) - - im2sequence_op = OpConfig( - type="im2sequence", - inputs={"X": ["input_data"]}, - outputs={"Out": ["seq_out"]}, - attrs={ - "kernels": [6, 1], - "out_stride": [1, 1], - "paddings": [0, 0, 0, 0], - "strides": [1, 1], - }, - ) - - sequence_conv_op = OpConfig( - type="sequence_conv", - inputs={"X": ["seq_out"], "Filter": ["conv_weight"]}, - outputs={"Out": ["conv_out"]}, - attrs={ - "contextLength": contextLength, - "contextStart": contextStart, - "contextStride": contextStride, - "paddingTrainable": paddingTrainable, - }, - ) - - elementwise_add_op = OpConfig( - type="elementwise_add", - inputs={"X": ["conv_out"], "Y": ["elt_weight"]}, - outputs={"Out": ["elt_output"]}, - attrs={'axis': axis}, - ) - - relu_op = OpConfig( - type="relu", - inputs={"X": ["elt_output"]}, - outputs={"Out": ["relu_output"]}, - attrs={}, - ) - - model_net = [ - im2sequence_op, - sequence_conv_op, - elementwise_add_op, - relu_op, - ] - - program_config = ProgramConfig( - ops=model_net, - weights={ - "conv_weight": TensorConfig( - data_gen=partial(generate_weight, [768 * contextLength, 16]) - ), - "elt_weight": TensorConfig( - data_gen=partial(generate_weight, [16]) - ), - }, - inputs={ - "input_data": TensorConfig(data_gen=partial(generate_input)) - }, - outputs=["relu_output"], - ) - - return program_config - - def sample_predictor_configs(self, program_config): - config = self.create_inference_config() - yield config, ["im2sequence", "fusion_seqconv_eltadd_relu"], ( - 1e-5, - 1e-5, - ) - - def test(self): - self.run_and_statis( - quant=False, passes=["seqconv_eltadd_relu_fuse_pass"] - ) - - -if __name__ == "__main__": - unittest.main() diff --git a/test/legacy_test/CMakeLists.txt b/test/legacy_test/CMakeLists.txt index 63d84ece4aa98..3c1024abc34ea 100644 --- a/test/legacy_test/CMakeLists.txt +++ b/test/legacy_test/CMakeLists.txt @@ -655,7 +655,6 @@ if(WITH_DISTRIBUTE) list(REMOVE_ITEM DIST_TEST_OPS "test_dist_mnist_lars") list(REMOVE_ITEM DIST_TEST_OPS "test_dist_mnist_train") list(REMOVE_ITEM DIST_TEST_OPS "test_dist_save_load") - list(REMOVE_ITEM DIST_TEST_OPS "test_dist_text_classification") list(REMOVE_ITEM DIST_TEST_OPS "test_dist_train") list(REMOVE_ITEM DIST_TEST_OPS "test_dist_word2vec") list(REMOVE_ITEM DIST_TEST_OPS "test_dist_fleet_gloo") diff --git a/test/legacy_test/dist_text_classification.py b/test/legacy_test/dist_text_classification.py deleted file mode 100644 index 0e3c79d758c80..0000000000000 --- a/test/legacy_test/dist_text_classification.py +++ /dev/null @@ -1,242 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import os -import re -import string -import tarfile - -import nets -from test_dist_base import TestDistRunnerBase, runtime_main - -import paddle -from paddle import base - -DTYPE = "float32" -VOCAB_URL = 'http://paddle-dist-ce-data.bj.bcebos.com/imdb.vocab' -VOCAB_MD5 = '23c86a0533c0151b6f12fa52b106dcc2' -DATA_URL = 'http://paddle-dist-ce-data.bj.bcebos.com/text_classification.tar.gz' -DATA_MD5 = '29ebfc94f11aea9362bbb7f5e9d86b8a' - - -# Load dictionary. -def load_vocab(filename): - vocab = {} - with open(filename, 'r', encoding="utf-8") as f: - for idx, line in enumerate(f): - vocab[line.strip()] = idx - return vocab - - -def get_worddict(dict_path): - word_dict = load_vocab(dict_path) - word_dict[""] = len(word_dict) - dict_dim = len(word_dict) - return word_dict, dict_dim - - -def conv_net( - input, - dict_dim, - emb_dim=128, - window_size=3, - num_filters=128, - fc0_dim=96, - class_dim=2, -): - emb = paddle.static.nn.embedding( - input=input, - size=[dict_dim, emb_dim], - is_sparse=False, - param_attr=base.ParamAttr( - initializer=paddle.nn.initializer.Constant(value=0.01) - ), - ) - - conv_3 = nets.sequence_conv_pool( - input=emb, - num_filters=num_filters, - filter_size=window_size, - act="tanh", - pool_type="max", - param_attr=base.ParamAttr( - initializer=paddle.nn.initializer.Constant(value=0.01) - ), - ) - - fc_0 = paddle.static.nn.fc( - x=[conv_3], - size=fc0_dim, - weight_attr=base.ParamAttr( - initializer=paddle.nn.initializer.Constant(value=0.01) - ), - ) - - prediction = paddle.static.nn.fc( - x=[fc_0], - size=class_dim, - activation="softmax", - weight_attr=base.ParamAttr( - initializer=paddle.nn.initializer.Constant(value=0.01) - ), - ) - - return prediction - - -def inference_network(dict_dim): - data = paddle.static.data( - name="words", shape=[-1, 1], dtype="int64", lod_level=1 - ) - out = conv_net(data, dict_dim) - return out - - -def get_reader(word_dict, batch_size): - # The training data set. - train_reader = paddle.batch(train(word_dict), batch_size=batch_size) - - # The testing data set. - test_reader = paddle.batch(test(word_dict), batch_size=batch_size) - - return train_reader, test_reader - - -def get_optimizer(learning_rate): - optimizer = paddle.optimizer.SGD(learning_rate=learning_rate) - return optimizer - - -class TestDistTextClassification2x2(TestDistRunnerBase): - def get_model(self, batch_size=2): - vocab = os.path.join( - paddle.dataset.common.DATA_HOME, "text_classification", "imdb.vocab" - ) - word_dict, dict_dim = get_worddict(vocab) - - # Input data - data = paddle.static.data( - name="words", shape=[-1, 1], dtype="int64", lod_level=1 - ) - label = paddle.static.data(name='label', shape=[-1, 1], dtype='int64') - - # Train program - predict = conv_net(data, dict_dim) - cost = paddle.nn.functional.cross_entropy( - input=predict, label=label, reduction='none', use_softmax=False - ) - avg_cost = paddle.mean(x=cost) - acc = paddle.static.accuracy(input=predict, label=label) - inference_program = base.default_main_program().clone() - - # Optimization - opt = get_optimizer(learning_rate=0.001) - opt.minimize(avg_cost) - - # Reader - train_reader, test_reader = get_reader(word_dict, batch_size) - - return ( - inference_program, - avg_cost, - train_reader, - test_reader, - acc, - predict, - ) - - -def tokenize(pattern): - """ - Read files that match the given pattern. Tokenize and yield each file. - """ - - with tarfile.open( - paddle.dataset.common.download( - DATA_URL, 'text_classification', DATA_MD5 - ) - ) as tarf: - # Note that we should use tarfile.next(), which does - # sequential access of member files, other than - # tarfile.extractfile, which does random access and might - # destroy hard disks. - tf = tarf.next() - while tf is not None: - if bool(pattern.match(tf.name)): - # newline and punctuations removal and ad-hoc tokenization. - yield tarf.extractfile(tf).read().rstrip(b'\n\r').translate( - None, string.punctuation.encode('latin-1') - ).lower().split() - tf = tarf.next() - - -def reader_creator(pos_pattern, neg_pattern, word_idx): - UNK = word_idx[''] - INS = [] - - def load(pattern, out, label): - for doc in tokenize(pattern): - out.append(([word_idx.get(w, UNK) for w in doc], label)) - - load(pos_pattern, INS, 0) - load(neg_pattern, INS, 1) - - def reader(): - yield from INS - - return reader - - -def train(word_idx): - """ - IMDB training set creator. - - It returns a reader creator, each sample in the reader is an zero-based ID - sequence and label in [0, 1]. - - :param word_idx: word dictionary - :type word_idx: dict - :return: Training reader creator - :rtype: callable - """ - return reader_creator( - re.compile(r"train/pos/.*\.txt$"), - re.compile(r"train/neg/.*\.txt$"), - word_idx, - ) - - -def test(word_idx): - """ - IMDB test set creator. - - It returns a reader creator, each sample in the reader is an zero-based ID - sequence and label in [0, 1]. - - :param word_idx: word dictionary - :type word_idx: dict - :return: Test reader creator - :rtype: callable - """ - return reader_creator( - re.compile(r"test/pos/.*\.txt$"), - re.compile(r"test/neg/.*\.txt$"), - word_idx, - ) - - -if __name__ == "__main__": - paddle.dataset.common.download(VOCAB_URL, 'text_classification', VOCAB_MD5) - paddle.dataset.common.download(DATA_URL, 'text_classification', DATA_MD5) - runtime_main(TestDistTextClassification2x2) diff --git a/test/legacy_test/nets.py b/test/legacy_test/nets.py index 035cb04a6f6d7..d63199ab9baf8 100644 --- a/test/legacy_test/nets.py +++ b/test/legacy_test/nets.py @@ -18,7 +18,6 @@ __all__ = [ "simple_img_conv_pool", - "sequence_conv_pool", "glu", "scaled_dot_product_attention", "img_conv_group", @@ -273,85 +272,6 @@ def __extend_list__(obj): return pool_out -def sequence_conv_pool( - input, - num_filters, - filter_size, - param_attr=None, - act="sigmoid", - pool_type="max", - bias_attr=None, -): - """ - :api_attr: Static Graph - - **This api takes input as an LoDTensor. If input is a Tensor, please use** - :ref:`api_base_nets_simple_img_conv_pool` **instead** - - The sequence_conv_pool is composed of :ref:`api_base_layers_sequence_conv` - and :ref:`api_base_layers_sequence_pool` . - - Args: - input (Tensor): 2-D LoDTensor, the input of sequence_conv, - which supports variable-time length input sequence. - The underlying of input is a matrix with shape - (T, N), where T is the total time steps in this mini-batch and N is - the input_hidden_size. The data type is float32 or float64. - num_filters(int): The number of filter. - filter_size (int): The filter size. - param_attr (ParamAttr): The parameters of the sequence_conv Layer. Default: None. - act (str|None): Activation type for Sequence_conv Layer. - If set to None, no activation will be applied. Default: "sigmoid". - pool_type (str): Pooling type can be :math:`max` for max-pooling, :math:`average` for - average-pooling, :math:`sum` for sum-pooling, :math:`sqrt` for sqrt-pooling. - Default :math:`max`. - bias_attr (ParamAttr|bool|None): The parameter attribute for the bias of sequence_conv. - If it is set to False, no bias will be added to the output units. - If it is set to None or one attribute of ParamAttr, sequence_conv - will create ParamAttr as bias_attr. If the Initializer of the bias_attr - is not set, the bias is initialized zero. Default: None. - - Returns: - The final result after sequence_conv and sequence_pool. - It is a 2-D Tensor, with the same data type as :attr:`input` - - Return Type: - Tensor - - Examples: - .. code-block:: python - - import paddle.base as base - import paddle - paddle.enable_static() - input_dim = 100 #len(word_dict) - emb_dim = 128 - hid_dim = 512 - data = paddle.static.data(name="words", shape=[None, 1], dtype="int64", lod_level=1) - emb = paddle.static.nn.embedding(input=data, size=[input_dim, emb_dim], is_sparse=True) - seq_conv = base.nets.sequence_conv_pool(input=emb, - num_filters=hid_dim, - filter_size=3, - act="tanh", - pool_type="sqrt") - """ - - check_variable_and_dtype(input, 'input', ['float32', 'float64'], 'input') - conv_out = paddle.static.nn.sequence_lod.sequence_conv( - input=input, - num_filters=num_filters, - filter_size=filter_size, - param_attr=param_attr, - bias_attr=bias_attr, - act=act, - ) - - pool_out = paddle.static.nn.sequence_lod.sequence_pool( - input=conv_out, pool_type=pool_type - ) - return pool_out - - @deprecated(since="2.0.0", update_to="paddle.nn.functional.glu") def glu(input, dim=-1): r""" diff --git a/test/legacy_test/test_dist_text_classification.py b/test/legacy_test/test_dist_text_classification.py index f777775724aee..46711a67bfb22 100644 --- a/test/legacy_test/test_dist_text_classification.py +++ b/test/legacy_test/test_dist_text_classification.py @@ -25,28 +25,12 @@ def _setup_config(self): self._sync_mode = True self._enforce_place = "CPU" - def test_text_classification(self): - self.check_with_place( - "dist_text_classification.py", - delta=1e-6, - check_error_log=True, - log_name=flag_name, - ) - class TestDistTextClassification2x2Async(TestDistBase): def _setup_config(self): self._sync_mode = False self._enforce_place = "CPU" - def test_se_resnext(self): - self.check_with_place( - "dist_text_classification.py", - delta=100, - check_error_log=True, - log_name=flag_name, - ) - if __name__ == "__main__": unittest.main() diff --git a/test/legacy_test/test_fusion_seqconv_eltadd_relu_op.py b/test/legacy_test/test_fusion_seqconv_eltadd_relu_op.py deleted file mode 100644 index b4b2471d95da9..0000000000000 --- a/test/legacy_test/test_fusion_seqconv_eltadd_relu_op.py +++ /dev/null @@ -1,95 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import sys -import unittest - -import numpy as np -from op_test import OpTest - -sys.path.append("../../test/sequence") -from test_sequence_conv import seqconv - - -class TestSeqConvEltAddRelu(OpTest): - def set_conf(self): - pass - - def setUp(self): - self.op_type = 'fusion_seqconv_eltadd_relu' - self.lod = [[6, 4]] - self.in_fea_size = 16 - self.out_fea_size = 8 - self.context_length = 4 - self.context_stride = 1 - self.context_start = 0 - self.set_conf() - - assert self.context_stride == 1 - - T = sum(self.lod[0]) - x = np.random.uniform(-1, 1, [T, self.in_fea_size]).astype('float32') - w = np.random.uniform( - -1, 1, [self.in_fea_size * self.context_length, self.out_fea_size] - ).astype('float32') - b = np.random.uniform(-2, 1, [1, self.out_fea_size]).astype('float32') - out = seqconv(x, self.lod, w, self.context_length, self.context_start) - out = np.maximum(out + b, 0) - - self.inputs = {'X': (x, self.lod), 'Filter': w, 'Bias': b} - self.attrs = { - 'contextStart': self.context_start, - 'contextLength': self.context_length, - 'contextStride': self.context_stride, - } - self.outputs = {'Out': out} - - def test_check_output(self): - self.check_output(check_dygraph=False) - - -class TestSeqConvEltAddReluBS1(TestSeqConvEltAddRelu): - def set_conf(self): - self.lod = [[10]] - - -class TestSeqConvEltAddReluBS1Case2(TestSeqConvEltAddRelu): - def set_conf(self): - self.lod = [[2]] - - -class TestSeqConvEltAddReluCase1(TestSeqConvEltAddRelu): - def set_conf(self): - self.lod = [[3, 5, 1, 6]] - self.context_length = 3 - self.context_start = -2 - - -class TestSeqConvEltAddReluCase2(TestSeqConvEltAddRelu): - def set_conf(self): - self.lod = [[10, 1, 2, 4, 1, 5, 6]] - self.in_fea_size = 2 - self.context_length = 4 - self.context_start = -1 - - -class TestSeqConvEltAddReluCase3(TestSeqConvEltAddRelu): - def set_conf(self): - self.lod = [[10, 1, 2, 4, 1, 5, 6]] - self.context_length = 5 - self.context_start = -4 - - -if __name__ == '__main__': - unittest.main() diff --git a/test/sequence/CMakeLists.txt b/test/sequence/CMakeLists.txt index 4f5b02114ebfd..ff11cacdf9ae0 100644 --- a/test/sequence/CMakeLists.txt +++ b/test/sequence/CMakeLists.txt @@ -7,7 +7,6 @@ string(REPLACE ".py" "" TEST_OPS "${TEST_OPS}") foreach(TEST_OP ${TEST_OPS}) py_test_modules(${TEST_OP} MODULES ${TEST_OP}) endforeach() -set_tests_properties(test_sequence_conv PROPERTIES TIMEOUT 120) set_tests_properties(test_sequence_concat PROPERTIES TIMEOUT 120) set_tests_properties(test_sequence_pool PROPERTIES TIMEOUT 120) diff --git a/test/sequence/test_sequence_conv.py b/test/sequence/test_sequence_conv.py deleted file mode 100644 index 21644be661dde..0000000000000 --- a/test/sequence/test_sequence_conv.py +++ /dev/null @@ -1,312 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import random -import unittest - -import numpy as np -from op_test import OpTest - -import paddle - - -def seqconv( - x, - lod, - filter, - context_length, - context_start, - padding_trainable=False, - padding_data=None, -): - [T, M] = x.shape - col = np.zeros((T, context_length * M)).astype('float32') - offset = [0] - for seq_len in lod[0]: - offset.append(offset[-1] + seq_len) - begin_pad = np.max([0, -context_start]) - for i in range(len(offset) - 1): - for j in range(context_length): - in_begin = offset[i] + context_start + j - in_end = offset[i + 1] + context_start + j - out_begin = offset[i] - out_end = offset[i + 1] - if in_begin < offset[i]: - pad_size = np.min( - [offset[i] - in_begin, offset[i + 1] - offset[i]] - ) - if padding_trainable: - sub_w = padding_data[j : j + pad_size, :] - col[ - offset[i] : offset[i] + pad_size, j * M : (j + 1) * M - ] = sub_w - out_begin = offset[i] + pad_size - in_begin = offset[i] - - if in_end > offset[i + 1]: - pad_size = np.min( - [in_end - offset[i + 1], offset[i + 1] - offset[i]] - ) - if padding_trainable: - sub_w = padding_data[ - begin_pad - + context_start - + j - - pad_size : begin_pad - + context_start - + j, - :, - ] - col[ - offset[i + 1] - pad_size : offset[i + 1], - j * M : (j + 1) * M, - ] = sub_w - in_end = offset[i + 1] - out_end = offset[i + 1] - pad_size - if in_end <= in_begin: - continue - in_sub = x[in_begin:in_end, :] - col[out_begin:out_end, j * M : (j + 1) * M] += in_sub - return np.dot(col, filter) - - -class TestSeqProject(OpTest): - def setUp(self): - self.init_test_case() - self.op_type = 'sequence_conv' - - if ( - self.context_length == 1 - and self.context_start == 0 - and self.padding_trainable - ): - print( - "If context_start is 0 " - "and context_length is 1," - " padding_trainable should be false." - ) - return - - # one level, batch size - x = np.random.uniform( - 0.1, 1, [self.input_size[0], self.input_size[1]] - ).astype('float32') - w = np.random.uniform( - 0.1, - 1, - [ - self.context_length * self.input_size[1], - self.output_represention, - ], - ).astype('float32') - - begin_pad = np.max([0, -self.context_start]) - end_pad = np.max([0, self.context_start + self.context_length - 1]) - total_pad = begin_pad + end_pad - padding_data = np.random.uniform( - 0.1, 1, [total_pad, self.input_size[1]] - ).astype('float32') - self.pad_data = padding_data - self.inputs = { - 'X': (x, self.lod), - 'Filter': w, - } - self.inputs_val = ['X', 'Filter'] - self.inputs_val_no_x = ['Filter'] - self.inputs_val_no_f = ['X'] - - if total_pad != 0: - self.inputs['PaddingData'] = padding_data - self.inputs_val = ['X', 'PaddingData', 'Filter'] - self.inputs_val_no_x = ['PaddingData', 'Filter'] - self.inputs_val_no_f = ['PaddingData', 'X'] - - self.attrs = { - 'contextStart': self.context_start, - 'contextLength': self.context_length, - 'paddingTrainable': self.padding_trainable, - 'contextStride': self.context_stride, - } - out = seqconv( - x, - self.lod, - w, - self.context_length, - self.context_start, - self.padding_trainable, - self.pad_data, - ) - self.outputs = {'Out': out} - - def test_check_output(self): - # NODE(yjjiang11): This op will be deprecated. - self.check_output(check_dygraph=False) - - def test_check_grad(self): - if self.padding_trainable: - self.check_grad( - set(self.inputs_val), - 'Out', - max_relative_error=0.05, - check_dygraph=False, - ) - - def test_check_grad_input(self): - self.check_grad( - ['X'], - 'Out', - max_relative_error=0.05, - no_grad_set=set(self.inputs_val_no_x), - check_dygraph=False, - ) - - def test_check_grad_padding_data(self): - if self.padding_trainable: - self.check_grad( - ['PaddingData'], - 'Out', - no_grad_set={'X', 'Filter'}, - check_dygraph=False, - ) - - def test_check_grad_Filter(self): - self.check_grad( - ['Filter'], - 'Out', - max_relative_error=0.05, - no_grad_set=set(self.inputs_val_no_f), - check_dygraph=False, - ) - - def test_check_grad_input_filter(self): - if self.padding_trainable: - self.check_grad( - ['X', 'Filter'], - 'Out', - max_relative_error=0.05, - no_grad_set={'PaddingData'}, - check_dygraph=False, - ) - - def test_check_grad_padding_input(self): - if self.padding_trainable: - self.check_grad( - self.inputs_val_no_f, - 'Out', - max_relative_error=0.05, - no_grad_set={'Filter'}, - check_dygraph=False, - ) - - def test_check_grad_padding_filter(self): - if self.padding_trainable: - self.check_grad( - self.inputs_val_no_x, - 'Out', - max_relative_error=0.05, - no_grad_set={'X'}, - check_dygraph=False, - ) - - def init_test_case(self): - self.input_row = 11 - self.context_start = 0 - self.context_length = 1 - self.padding_trainable = False - self.context_stride = 1 - - self.input_size = [self.input_row, 23] - offset_lod = [[0, 4, 5, 8, self.input_row]] - self.lod = [[]] - # convert from offset-based lod to length-based lod - for i in range(len(offset_lod[0]) - 1): - self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) - self.output_represention = 8 # output feature size - - -class TestSeqProjectCase1(TestSeqProject): - def init_test_case(self): - self.input_row = 11 - self.context_start = -1 - self.context_length = 3 - self.padding_trainable = True - self.context_stride = 1 - - self.input_size = [self.input_row, 50] - offset_lod = [[0, 4, 5, 8, self.input_row]] - self.lod = [[]] - # convert from offset-based lod to length-based lod - for i in range(len(offset_lod[0]) - 1): - self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) - self.output_represention = 8 # output feature size - - -class TestSeqProjectCase2Len0(TestSeqProject): - def init_test_case(self): - self.input_row = 11 - self.context_start = -1 - self.context_length = 3 - self.padding_trainable = True - self.context_stride = 1 - - self.input_size = [self.input_row, 50] - offset_lod = [[0, 0, 4, 5, 5, 8, self.input_row, self.input_row]] - self.lod = [[]] - # convert from offset-based lod to length-based lod - for i in range(len(offset_lod[0]) - 1): - self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) - self.output_represention = 8 # output feature size - - -class TestSeqProjectCase3(TestSeqProject): - def init_test_case(self): - self.input_row = 25 - self.context_start = 2 - self.context_length = 3 - self.padding_trainable = True - self.context_stride = 1 - - self.input_size = [self.input_row, 25] - idx = list(range(self.input_size[0])) - del idx[0] - offset_lod = [ - [0] + np.sort(random.sample(idx, 8)).tolist() + [self.input_size[0]] - ] - self.lod = [[]] - # convert from offset-based lod to length-based lod - for i in range(len(offset_lod[0]) - 1): - self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) - self.output_represention = 8 # output feature size - - -class TestSeqConvApi(unittest.TestCase): - def test_api(self): - from paddle import base - - x = paddle.static.data('x', shape=[-1, 32], lod_level=1) - y = paddle.static.nn.sequence_lod.sequence_conv( - input=x, num_filters=2, filter_size=3, padding_start=None - ) - - place = base.CPUPlace() - x_tensor = base.create_lod_tensor( - np.random.rand(10, 32).astype("float32"), [[2, 3, 1, 4]], place - ) - exe = base.Executor(place) - exe.run(base.default_startup_program()) - ret = exe.run(feed={'x': x_tensor}, fetch_list=[y], return_numpy=False) - - -if __name__ == '__main__': - unittest.main() diff --git a/test/xpu/test_sequence_conv_op_xpu.py b/test/xpu/test_sequence_conv_op_xpu.py deleted file mode 100644 index 1b0816938a248..0000000000000 --- a/test/xpu/test_sequence_conv_op_xpu.py +++ /dev/null @@ -1,448 +0,0 @@ -# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -# -# Licensed under the Apache License, Version 2.0 (the "License"); -# you may not use this file except in compliance with the License. -# You may obtain a copy of the License at -# -# http://www.apache.org/licenses/LICENSE-2.0 -# -# Unless required by applicable law or agreed to in writing, software -# distributed under the License is distributed on an "AS IS" BASIS, -# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -# See the License for the specific language governing permissions and -# limitations under the License. - -import random -import sys -import unittest - -import numpy as np - -import paddle - -sys.path.append("../") -from get_test_cover_info import ( - XPUOpTestWrapper, - create_test_class, - get_xpu_op_support_types, -) -from op_test_xpu import XPUOpTest - -paddle.enable_static() -np.set_printoptions(threshold=np.inf) - - -def seqconv( - x, - lod, - filter, - context_length, - context_start, - padding_trainable=False, - padding_data=None, -): - [T, M] = x.shape - col = np.zeros((T, context_length * M)).astype('float32') - offset = [0] - for seq_len in lod[0]: - offset.append(offset[-1] + seq_len) - begin_pad = np.max([0, -context_start]) - for i in range(len(offset) - 1): - for j in range(context_length): - in_begin = offset[i] + context_start + j - in_end = offset[i + 1] + context_start + j - out_begin = offset[i] - out_end = offset[i + 1] - if in_begin < offset[i]: - pad_size = np.min( - [offset[i] - in_begin, offset[i + 1] - offset[i]] - ) - if padding_trainable: - sub_w = padding_data[j : j + pad_size, :] - col[ - offset[i] : offset[i] + pad_size, j * M : (j + 1) * M - ] = sub_w - out_begin = offset[i] + pad_size - in_begin = offset[i] - - if in_end > offset[i + 1]: - pad_size = np.min( - [in_end - offset[i + 1], offset[i + 1] - offset[i]] - ) - if padding_trainable: - sub_w = padding_data[ - begin_pad - + context_start - + j - - pad_size : begin_pad - + context_start - + j, - :, - ] - col[ - offset[i + 1] - pad_size : offset[i + 1], - j * M : (j + 1) * M, - ] = sub_w - in_end = offset[i + 1] - out_end = offset[i + 1] - pad_size - if in_end <= in_begin: - continue - in_sub = x[in_begin:in_end, :] - col[out_begin:out_end, j * M : (j + 1) * M] += in_sub - return np.dot(col, filter) - - -class XPUTestSequenceConv(XPUOpTestWrapper): - def __init__(self): - self.op_name = 'sequence_conv' - - class TestSeqProject(XPUOpTest): - def setUp(self): - self.init_test_case() - self.op_type = 'sequence_conv' - self.dtype = self.in_type - self.use_xpu = True - - if ( - self.context_length == 1 - and self.context_start == 0 - and self.padding_trainable - ): - print( - "If context_start is 0 " - "and context_length is 1," - " padding_trainable should be false." - ) - return - - # one level, batch size - x = np.random.uniform( - -6.10907e-05, - 0.000104218, - [self.input_size[0], self.input_size[1]], - ).astype(self.dtype) - w = np.random.uniform( - -3.17068e-05, - 0.000159822, - [ - self.context_length * self.input_size[1], - self.output_represention, - ], - ).astype(self.dtype) - - begin_pad = np.max([0, -self.context_start]) - end_pad = np.max([0, self.context_start + self.context_length - 1]) - total_pad = begin_pad + end_pad - padding_data = np.random.uniform( - 0, 0, [total_pad, self.input_size[1]] - ).astype(self.dtype) - self.pad_data = padding_data - self.inputs = { - 'X': (x, self.lod), - 'Filter': w, - } - self.inputs_val = ['X', 'Filter'] - self.inputs_val_no_x = ['Filter'] - self.inputs_val_no_f = ['X'] - - if total_pad != 0: - self.inputs['PaddingData'] = padding_data - self.inputs_val = ['X', 'PaddingData', 'Filter'] - self.inputs_val_no_x = ['PaddingData', 'Filter'] - self.inputs_val_no_f = ['PaddingData', 'X'] - - self.attrs = { - 'contextStart': self.context_start, - 'contextLength': self.context_length, - 'paddingTrainable': self.padding_trainable, - 'contextStride': self.context_stride, - } - out = seqconv( - x, - self.lod, - w, - self.context_length, - self.context_start, - self.padding_trainable, - self.pad_data, - ) - self.outputs = {'Out': out} - - def test_check_output(self): - place = paddle.XPUPlace(0) - self.check_output_with_place(place) - - def test_check_grad_input(self): - self.check_grad(['X'], 'Out', no_grad_set=set(self.inputs_val_no_x)) - - def test_check_grad_padding_data(self): - if self.padding_trainable: - self.check_grad( - ['PaddingData'], 'Out', no_grad_set={'X', 'Filter'} - ) - - def test_check_grad_Filter(self): - self.check_grad( - ['Filter'], 'Out', no_grad_set=set(self.inputs_val_no_f) - ) - - def test_check_grad_input_filter(self): - if self.padding_trainable: - self.check_grad( - ['X', 'Filter'], 'Out', no_grad_set={'PaddingData'} - ) - - def test_check_grad_padding_input(self): - if self.padding_trainable: - self.check_grad( - self.inputs_val_no_f, 'Out', no_grad_set={'Filter'} - ) - - def test_check_grad_padding_filter(self): - if self.padding_trainable: - self.check_grad(self.inputs_val_no_x, 'Out', no_grad_set={'X'}) - - def init_test_case(self): - self.input_row = 7 - self.input_col = 25 - self.context_start = -2 - self.context_length = 5 - self.padding_trainable = False - self.context_stride = 1 - - self.input_size = [self.input_row, self.input_col] - offset_lod = [[0, 1, self.input_row]] - self.lod = [[]] - # convert from offset-based lod to length-based lod - for i in range(len(offset_lod[0]) - 1): - self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) - self.output_represention = 8 # output feature size - - class TestSeqProjectCase1(TestSeqProject): - def init_test_case(self): - self.input_row = 11 - self.context_start = -2 - self.context_length = 5 - self.padding_trainable = False - self.context_stride = 1 - - self.input_size = [self.input_row, 50] - offset_lod = [[0, 4, 5, 8, self.input_row]] - self.lod = [[]] - # convert from offset-based lod to length-based lod - for i in range(len(offset_lod[0]) - 1): - self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) - self.output_represention = 8 # output feature size - - class TestSeqProjectCase2Len0(TestSeqProject): - def init_test_case(self): - self.input_row = 11 - self.context_start = -2 - self.context_length = 5 - self.padding_trainable = False - self.context_stride = 1 - - self.input_size = [self.input_row, 50] - offset_lod = [[0, 0, 4, 5, 5, 8, self.input_row, self.input_row]] - self.lod = [[]] - # convert from offset-based lod to length-based lod - for i in range(len(offset_lod[0]) - 1): - self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) - self.output_represention = 8 # output feature size - - class TestSeqProjectCase3(TestSeqProject): - def init_test_case(self): - self.input_row = 25 - self.context_start = -2 - self.context_length = 5 - self.padding_trainable = False - self.context_stride = 1 - - self.input_size = [self.input_row, 25] - idx = list(range(self.input_size[0])) - del idx[0] - offset_lod = [ - [0] - + np.sort(random.sample(idx, 8)).tolist() - + [self.input_size[0]] - ] - self.lod = [[]] - # convert from offset-based lod to length-based lod - for i in range(len(offset_lod[0]) - 1): - self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) - self.output_represention = 8 # output feature size - - class TestSeqProjectCase4(TestSeqProject): - def init_test_case(self): - self.input_row = 7835 - self.input_col = 128 - self.context_start = -2 - self.context_length = 5 - self.padding_trainable = False - self.context_stride = 1 - - self.input_size = [self.input_row, self.input_col] - offset_lod = [ - [ - 0, - 1, - 2, - 3, - 131, - 241, - 242, - 263, - 264, - 265, - 266, - 267, - 268, - 387, - 515, - 516, - 644, - 645, - 772, - 794, - 922, - 923, - 924, - 944, - 945, - 1073, - 1074, - 1202, - 1330, - 1458, - 1556, - 1557, - 1558, - 1686, - 1748, - 1876, - 1912, - 1913, - 1914, - 2032, - 2066, - 2194, - 2308, - 2309, - 2347, - 2475, - 2476, - 2477, - 2478, - 2606, - 2607, - 2735, - 2736, - 2737, - 2738, - 2838, - 2966, - 2967, - 2968, - 2969, - 3097, - 3225, - 3353, - 3481, - 3482, - 3520, - 3642, - 3643, - 3754, - 3882, - 3883, - 4010, - 4011, - 4012, - 4140, - 4219, - 4228, - 4356, - 4357, - 4415, - 4475, - 4476, - 4604, - 4605, - 4606, - 4694, - 4695, - 4808, - 4936, - 4961, - 4962, - 5004, - 5132, - 5260, - 5312, - 5440, - 5441, - 5569, - 5570, - 5675, - 5676, - 5750, - 5810, - 5811, - 5939, - 6021, - 6149, - 6277, - 6278, - 6364, - 6425, - 6519, - 6647, - 6648, - 6739, - 6867, - 6995, - 6996, - 7120, - 7223, - 7244, - 7367, - 7407, - 7408, - 7467, - 7595, - 7699, - 7827, - 7835, - ] - ] - self.lod = [[]] - # convert from offset-based lod to length-based lod - for i in range(len(offset_lod[0]) - 1): - self.lod[0].append(offset_lod[0][i + 1] - offset_lod[0][i]) - self.output_represention = 8 # output feature size - - -support_types = get_xpu_op_support_types('sequence_conv') -for stype in support_types: - create_test_class(globals(), XPUTestSequenceConv, stype) - - -class TestSeqConvApi(unittest.TestCase): - def test_api(self): - from paddle import base - - x = paddle.static.data('x', shape=[-1, 32], lod_level=1) - y = paddle.static.nn.sequence_lod.sequence_conv( - input=x, num_filters=2, filter_size=3, padding_start=None - ) - place = base.CPUPlace() - x_tensor = base.create_lod_tensor( - np.random.rand(10, 32).astype("float32"), [[2, 3, 1, 4]], place - ) - exe = base.Executor(place) - exe.run(base.default_startup_program()) - ret = exe.run(feed={'x': x_tensor}, fetch_list=[y], return_numpy=False) - - -if __name__ == '__main__': - unittest.main() From 213a8ffa4e251c81bbf62a2dc0bb9a9e07b66fb0 Mon Sep 17 00:00:00 2001 From: co63oc Date: Sun, 7 Apr 2024 21:50:58 +0800 Subject: [PATCH 2/6] Fix --- .../api/analyzer_seq_conv1_tester.cc | 190 ------------------ 1 file changed, 190 deletions(-) delete mode 100644 test/cpp/inference/api/analyzer_seq_conv1_tester.cc diff --git a/test/cpp/inference/api/analyzer_seq_conv1_tester.cc b/test/cpp/inference/api/analyzer_seq_conv1_tester.cc deleted file mode 100644 index cc4cb24201b97..0000000000000 --- a/test/cpp/inference/api/analyzer_seq_conv1_tester.cc +++ /dev/null @@ -1,190 +0,0 @@ -// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -#include "test/cpp/inference/api/tester_helper.h" - -namespace paddle { -namespace inference { - -struct DataRecord { - std::vector> title1, title2, title3, l1; - std::vector lod1, lod2, lod3, l1_lod; - size_t batch_iter{0}, batch_size{1}, num_samples; // total number of samples - DataRecord() = default; - explicit DataRecord(const std::string &path, int batch_size = 1) - : batch_size(batch_size) { - Load(path); - } - DataRecord NextBatch() { - DataRecord data; - size_t batch_end = batch_iter + batch_size; - // NOTE skip the final batch, if no enough data is provided. - if (batch_end <= title1.size()) { - GetInputPerBatch(title1, &data.title1, &data.lod1, batch_iter, batch_end); - GetInputPerBatch(title2, &data.title2, &data.lod2, batch_iter, batch_end); - GetInputPerBatch(title3, &data.title3, &data.lod3, batch_iter, batch_end); - GetInputPerBatch(l1, &data.l1, &data.l1_lod, batch_iter, batch_end); - } - batch_iter += batch_size; - return data; - } - void Load(const std::string &path) { - std::ifstream file(path); - std::string line; - int num_lines = 0; - while (std::getline(file, line)) { - num_lines++; - std::vector data; - split(line, '\t', &data); - PADDLE_ENFORCE_GT( - data.size(), - 4, - paddle::platform::errors::Fatal("The size of data is invaild.")); - // load title1 data - std::vector title1_data; - split_to_int64(data[0], ' ', &title1_data); - // load title2 data - std::vector title2_data; - split_to_int64(data[1], ' ', &title2_data); - // load title3 data - std::vector title3_data; - split_to_int64(data[2], ' ', &title3_data); - // load l1 data - std::vector l1_data; - split_to_int64(data[3], ' ', &l1_data); - title1.push_back(std::move(title1_data)); - title2.push_back(std::move(title2_data)); - title3.push_back(std::move(title3_data)); - l1.push_back(std::move(l1_data)); - } - num_samples = num_lines; - } -}; - -void PrepareInputs(std::vector *input_slots, - DataRecord *data, - int batch_size) { - PaddleTensor title1_tensor, title2_tensor, title3_tensor, l1_tensor; - title1_tensor.name = "title1"; - title2_tensor.name = "title2"; - title3_tensor.name = "title3"; - l1_tensor.name = "l1"; - auto one_batch = data->NextBatch(); - // assign data - TensorAssignData(&title1_tensor, one_batch.title1, one_batch.lod1); - TensorAssignData(&title2_tensor, one_batch.title2, one_batch.lod2); - TensorAssignData(&title3_tensor, one_batch.title3, one_batch.lod3); - TensorAssignData(&l1_tensor, one_batch.l1, one_batch.l1_lod); - // Set inputs. - input_slots->assign({title1_tensor, title2_tensor, title3_tensor, l1_tensor}); - for (auto &tensor : *input_slots) { - tensor.dtype = PaddleDType::INT64; - } -} - -void SetConfig(AnalysisConfig *cfg) { - cfg->SetModel(FLAGS_infer_model); - cfg->DisableGpu(); - cfg->SwitchSpecifyInputNames(); - cfg->SwitchIrOptim(); -} - -void SetInput(std::vector> *inputs) { - DataRecord data(FLAGS_infer_data, FLAGS_batch_size); - std::vector input_slots; - int epoch = - FLAGS_test_all_data ? data.num_samples / FLAGS_batch_size : 1; // NOLINT - LOG(INFO) << "number of samples: " << epoch * FLAGS_batch_size; - for (int bid = 0; bid < epoch; ++bid) { - PrepareInputs(&input_slots, &data, FLAGS_batch_size); - (*inputs).emplace_back(input_slots); - } -} - -// Easy for profiling independently. -TEST(Analyzer_seq_conv1, profile) { - AnalysisConfig cfg; - SetConfig(&cfg); - std::vector> outputs; - - std::vector> input_slots_all; - SetInput(&input_slots_all); - TestPrediction(reinterpret_cast(&cfg), - input_slots_all, - &outputs, - FLAGS_num_threads); - - if (FLAGS_num_threads == 1 && !FLAGS_test_all_data) { - // the first inference result - PADDLE_ENFORCE_GT(outputs.size(), - 0, - paddle::platform::errors::Fatal( - "The size of output should be greater than 0.")); - auto output = outputs.back(); - PADDLE_ENFORCE_EQ(output.size(), - 1UL, - paddle::platform::errors::Fatal( - "The size of output should be equal to 0.")); - size_t size = GetSize(output[0]); - PADDLE_ENFORCE_GT(size, - 0, - paddle::platform::errors::Fatal( - "The size of output should be greater than 0.")); - float *result = static_cast(output[0].data.data()); - // output is probability, which is in (0, 1). - for (size_t i = 0; i < size; i++) { - EXPECT_GT(result[i], 0); - EXPECT_LT(result[i], 1); - } - } -} - -// Check the fuse status -TEST(Analyzer_seq_conv1, fuse_statis) { - AnalysisConfig cfg; - SetConfig(&cfg); - int num_ops; - auto predictor = CreatePaddlePredictor(cfg); - - auto fuse_statis = GetFuseStatis(predictor.get(), &num_ops); - ASSERT_TRUE(fuse_statis.count("fc_fuse")); - ASSERT_TRUE(fuse_statis.count("seqconv_eltadd_relu_fuse")); - EXPECT_EQ(fuse_statis.at("fc_fuse"), 2); - EXPECT_EQ(fuse_statis.at("seqconv_eltadd_relu_fuse"), 6); -} - -// Compare result of NativeConfig and AnalysisConfig -TEST(Analyzer_seq_conv1, compare) { - AnalysisConfig cfg; - SetConfig(&cfg); - - std::vector> input_slots_all; - SetInput(&input_slots_all); - CompareNativeAndAnalysis( - reinterpret_cast(&cfg), input_slots_all); -} - -// Compare Deterministic result -TEST(Analyzer_seq_conv1, compare_determine) { - AnalysisConfig cfg; - SetConfig(&cfg); - - std::vector> input_slots_all; - SetInput(&input_slots_all); - CompareDeterministic(reinterpret_cast(&cfg), - input_slots_all); -} - -} // namespace inference -} // namespace paddle From a6b2593d09ee3b840d0dbcfb469388d871eb94fc Mon Sep 17 00:00:00 2001 From: co63oc Date: Mon, 8 Apr 2024 06:51:31 +0800 Subject: [PATCH 3/6] ci From 4ad1fea9e46342d60c130c1741ce69daeecd0ac9 Mon Sep 17 00:00:00 2001 From: co63oc Date: Mon, 8 Apr 2024 09:10:57 +0800 Subject: [PATCH 4/6] Fix --- test/cpp/inference/api/CMakeLists.txt | 8 -------- 1 file changed, 8 deletions(-) diff --git a/test/cpp/inference/api/CMakeLists.txt b/test/cpp/inference/api/CMakeLists.txt index ffb306df0810d..63e4acea5ea2c 100644 --- a/test/cpp/inference/api/CMakeLists.txt +++ b/test/cpp/inference/api/CMakeLists.txt @@ -485,14 +485,6 @@ if(WITH_TESTING AND WITH_INFERENCE_API_TEST) test_analyzer_text_classification ${TEXT_CLASSIFICATION_INSTALL_DIR} analyzer_text_classification_tester.cc EXTRA_DEPS common) - # seq_conv1 - set(SEQ_CONV1_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/seq_conv1") - download_model_and_data_without_verify( - ${SEQ_CONV1_INSTALL_DIR} "seq_conv1_model.tar.gz" - "seq_conv1_data.txt.tar.gz") - inference_analysis_api_test(test_analyzer_seq_conv1 ${SEQ_CONV1_INSTALL_DIR} - analyzer_seq_conv1_tester.cc EXTRA_DEPS common) - # transformer, the dataset only works on batch_size=8 now set(TRANSFORMER_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/transformer") download_model_and_data_without_verify( From 19b16edd905ef320eeffb9933b5f5321215b1fb2 Mon Sep 17 00:00:00 2001 From: co63oc Date: Mon, 8 Apr 2024 16:48:46 +0800 Subject: [PATCH 5/6] Fix --- paddle/fluid/inference/api/paddle_pass_builder.cc | 1 - 1 file changed, 1 deletion(-) diff --git a/paddle/fluid/inference/api/paddle_pass_builder.cc b/paddle/fluid/inference/api/paddle_pass_builder.cc index f53e89d9f812c..db52f899698c4 100644 --- a/paddle/fluid/inference/api/paddle_pass_builder.cc +++ b/paddle/fluid/inference/api/paddle_pass_builder.cc @@ -442,7 +442,6 @@ void CpuPassStrategy::EnableMkldnnInt8() { passes_.emplace_back("squeeze2_transpose2_onednn_fuse_pass"); passes_.emplace_back("layer_norm_fuse_pass"); passes_.emplace_back("attention_lstm_fuse_pass"); - passes_.emplace_back("seqconv_eltadd_relu_fuse_pass"); passes_.emplace_back("fc_lstm_fuse_pass"); passes_.emplace_back("mul_lstm_fuse_pass"); passes_.emplace_back("fc_gru_fuse_pass"); From f486a87d561df5789600a57cb32ff555b13de69b Mon Sep 17 00:00:00 2001 From: co63oc Date: Mon, 8 Apr 2024 16:49:45 +0800 Subject: [PATCH 6/6] Fix --- tools/parallel_UT_rule.py | 2 -- tools/static_mode_white_list.py | 1 - 2 files changed, 3 deletions(-) diff --git a/tools/parallel_UT_rule.py b/tools/parallel_UT_rule.py index db192979cee60..cf6aceaa806b0 100755 --- a/tools/parallel_UT_rule.py +++ b/tools/parallel_UT_rule.py @@ -1336,7 +1336,6 @@ 'test_imperative_container_parameterlist', 'test_multiplex_op', 'test_trt_transpose_flatten_concat_fuse_pass', - 'test_seqconv_eltadd_relu_fuse_pass', 'test_assert_op', 'test_scatter_nd_op', 'test_sequence_expand', @@ -2602,7 +2601,6 @@ 'test_elementwise_div_op', 'test_gather_tree_op', 'test_imperative_named_members', - 'test_seqconv_eltadd_relu_fuse_pass', 'test_analysis_predictor', 'test_convert_operators', 'test_add_reader_dependency', diff --git a/tools/static_mode_white_list.py b/tools/static_mode_white_list.py index 48f7178fa23dc..0c9be61a7f200 100755 --- a/tools/static_mode_white_list.py +++ b/tools/static_mode_white_list.py @@ -542,7 +542,6 @@ 'test_conv_elementwise_add_fuse_pass', 'test_fc_fuse_pass', 'test_repeated_fc_relu_fuse_pass', - 'test_seqconv_eltadd_relu_fuse_pass', 'test_squared_mat_sub_fuse_pass', 'test_transpose_flatten_concat_fuse_pass', 'test_detection_map_op',