From f6e4be0de6b4a21b1229b7d47e7f92602843a874 Mon Sep 17 00:00:00 2001 From: yaoxuefeng Date: Sat, 6 Jun 2020 03:29:43 +0800 Subject: [PATCH] add FNN and support load specific vars --- core/trainers/single_trainer.py | 20 ++++- models/rank/fnn/__init__.py | 13 ++++ models/rank/fnn/config.yaml | 85 ++++++++++++++++++++ models/rank/fnn/fm_model.py | 132 +++++++++++++++++++++++++++++++ models/rank/fnn/model.py | 133 ++++++++++++++++++++++++++++++++ 5 files changed, 382 insertions(+), 1 deletion(-) create mode 100755 models/rank/fnn/__init__.py create mode 100755 models/rank/fnn/config.yaml create mode 100755 models/rank/fnn/fm_model.py create mode 100755 models/rank/fnn/model.py diff --git a/core/trainers/single_trainer.py b/core/trainers/single_trainer.py index d6461827..79fad6e7 100755 --- a/core/trainers/single_trainer.py +++ b/core/trainers/single_trainer.py @@ -310,15 +310,33 @@ class SingleTrainer(TranspileTrainer): context['is_exit'] = True def load(self, is_fleet=False): + def name_has_embedding(var): + res = "embedding_1.w_0" == var.name + return res + dirname = envs.get_global_env( "runner." + self._runner_name + ".init_model_path", None) + load_vars = envs.get_global_env( + "runner." + self._runner_name + ".load_vars", None) + + def name_has_embedding(var): + res = False + for var_name in load_vars: + if var_name == var.name: + return True + return res + if dirname is None or dirname == "": return print("going to load ", dirname) if is_fleet: fleet.load_persistables(self._exe, dirname) else: - fluid.io.load_persistables(self._exe, dirname) + if load_vars is None or dirname == "": + fluid.io.load_persistables(self._exe, dirname) + else: + fluid.io.load_vars( + self._exe, dirname, predicate=name_has_embedding) def save(self, epoch_id, is_fleet=False): def need_save(epoch_id, epoch_interval, is_last=False): diff --git a/models/rank/fnn/__init__.py b/models/rank/fnn/__init__.py new file mode 100755 index 00000000..abf198b9 --- /dev/null +++ b/models/rank/fnn/__init__.py @@ -0,0 +1,13 @@ +# Copyright (c) 2020 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. diff --git a/models/rank/fnn/config.yaml b/models/rank/fnn/config.yaml new file mode 100755 index 00000000..3ed970c7 --- /dev/null +++ b/models/rank/fnn/config.yaml @@ -0,0 +1,85 @@ +# Copyright (c) 2020 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. + + +# global settings +debug: false +workspace: "paddlerec.models.rank.fnn" + +dataset: + - name: train_sample + type: QueueDataset + batch_size: 5 + data_path: "{workspace}/../dataset/Criteo_data/sample_data/train" + sparse_slots: "label feat_idx" + dense_slots: "feat_value:39" + - name: infer_sample + type: QueueDataset + batch_size: 5 + data_path: "{workspace}/../dataset/Criteo_data/sample_data/train" + sparse_slots: "label feat_idx" + dense_slots: "feat_value:39" + +hyper_parameters: + # 用户自定义配置 + optimizer: + class: SGD + learning_rate: 0.0001 + sparse_feature_number: 1086460 + sparse_feature_dim: 9 + reg: 0.001 + num_field: 39 + fc_sizes: [512, 256, 128, 32] + +mode: train_FM_runner #for FM phase: train_FM_runner for dnn phase: train_DNN_runner +# if infer, change mode to "infer_runner" and change phase to "infer_phase" + +runner: + - name: train_FM_runner + trainer_class: single_train + epochs: 1 + device: cpu + init_model_path: "" + save_checkpoint_interval: 1 + save_inference_interval: 1 + save_checkpoint_path: "increment" + save_inference_path: "inference" + print_interval: 1 + - name: train_DNN_runner + trainer_class: single_train + epochs: 1 + device: cpu + init_model_path: "increment/0" + load_vars: "embedding_1.w_0, embedding_0.w_0" + save_checkpoint_interval: 1 + save_inference_interval: 1 + save_checkpoint_path: "increment_fnn" + save_inference_path: "inference_fnn" + print_interval: 1 + - name: infer_runner + trainer_class: single_infer + epochs: 1 + device: cpu + init_model_path: "increment/0" + print_interval: 1 + +phase: +- name: phase1 + model: "{workspace}/fm_model.py" # for FM phase: fm_model.py for dnn phase model.py + dataset_name: train_sample + thread_num: 1 +#- name: infer_phase +# model: "{workspace}/model.py" +# dataset_name: infer_sample +# thread_num: 1 diff --git a/models/rank/fnn/fm_model.py b/models/rank/fnn/fm_model.py new file mode 100755 index 00000000..3cd20108 --- /dev/null +++ b/models/rank/fnn/fm_model.py @@ -0,0 +1,132 @@ +# Copyright (c) 2020 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 +from collections import OrderedDict + +import paddle.fluid as fluid + +from paddlerec.core.utils import envs +from paddlerec.core.model import Model as ModelBase + + +class Model(ModelBase): + def __init__(self, config): + ModelBase.__init__(self, config) + + def _init_hyper_parameters(self): + self.is_distributed = True if envs.get_trainer( + ) == "CtrTrainer" else False + self.sparse_feature_number = envs.get_global_env( + "hyper_parameters.sparse_feature_number", None) + self.sparse_feature_dim = envs.get_global_env( + "hyper_parameters.sparse_feature_dim", None) + self.reg = envs.get_global_env("hyper_parameters.reg", 1e-4) + self.num_field = envs.get_global_env("hyper_parameters.num_field", + None) + + def net(self, inputs, is_infer=False): + raw_feat_idx = self._sparse_data_var[1] # (batch_size * num_field) * 1 + raw_feat_value = self._dense_data_var[0] # batch_size * num_field + self.label = self._sparse_data_var[0] # batch_size * 1 + + init_value_ = 0.1 + + feat_idx = raw_feat_idx + feat_value = fluid.layers.reshape( + raw_feat_value, + [-1, self.num_field, 1]) # batch_size * num_field * 1 + + # ------------------------- first order term -------------------------- + + first_weights_re = fluid.embedding( + input=feat_idx, + is_sparse=True, + is_distributed=self.is_distributed, + dtype='float32', + size=[self.sparse_feature_number + 1, 1], + padding_idx=0, + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.TruncatedNormalInitializer( + loc=0.0, scale=init_value_), + regularizer=fluid.regularizer.L1DecayRegularizer(self.reg)) + ) # (batch_size * num_field) * 1 * 1(embedding_size) + first_weights = fluid.layers.reshape( + first_weights_re, + shape=[-1, self.num_field, 1]) # batch_size * num_field * 1 + y_first_order = fluid.layers.reduce_sum((first_weights * feat_value), + 1) # batch_size * 1 + b_linear = fluid.layers.create_parameter( + shape=[1], + dtype='float32', + default_initializer=fluid.initializer.ConstantInitializer( + value=0)) # 1 + # ------------------------- second order term -------------------------- + + feat_embeddings_re = fluid.embedding( + input=feat_idx, + is_sparse=True, + is_distributed=self.is_distributed, + dtype='float32', + size=[self.sparse_feature_number + 1, self.sparse_feature_dim], + padding_idx=0, + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.TruncatedNormalInitializer( + loc=0.0, + scale=init_value_ / + math.sqrt(float(self.sparse_feature_dim)))) + ) # (batch_size * num_field) * 1 * embedding_size + feat_embeddings = fluid.layers.reshape( + feat_embeddings_re, + shape=[-1, self.num_field, self.sparse_feature_dim + ]) # batch_size * num_field * embedding_size + feat_embeddings = feat_embeddings * feat_value # batch_size * num_field * embedding_size + + # sum_square part + summed_features_emb = fluid.layers.reduce_sum( + feat_embeddings, 1) # batch_size * embedding_size + summed_features_emb_square = fluid.layers.square( + summed_features_emb) # batch_size * embedding_size + + # square_sum part + squared_features_emb = fluid.layers.square( + feat_embeddings) # batch_size * num_field * embedding_size + squared_sum_features_emb = fluid.layers.reduce_sum( + squared_features_emb, 1) # batch_size * embedding_size + + y_FM = 0.5 * fluid.layers.reduce_sum( + summed_features_emb_square - squared_sum_features_emb, + dim=1, + keep_dim=True) # batch_size * 1 + + # ------------------------- Predict -------------------------- + + self.predict = fluid.layers.sigmoid(b_linear + y_first_order + y_FM) + + cost = fluid.layers.log_loss( + input=self.predict, label=fluid.layers.cast(self.label, + "float32")) # log_loss + avg_cost = fluid.layers.reduce_sum(cost) + + self._cost = avg_cost + + predict_2d = fluid.layers.concat([1 - self.predict, self.predict], 1) + label_int = fluid.layers.cast(self.label, 'int64') + auc_var, batch_auc_var, _ = fluid.layers.auc(input=predict_2d, + label=label_int, + slide_steps=0) + self._metrics["AUC"] = auc_var + self._metrics["BATCH_AUC"] = batch_auc_var + if is_infer: + self._infer_results["AUC"] = auc_var diff --git a/models/rank/fnn/model.py b/models/rank/fnn/model.py new file mode 100755 index 00000000..936172f6 --- /dev/null +++ b/models/rank/fnn/model.py @@ -0,0 +1,133 @@ +# Copyright (c) 2020 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 +from collections import OrderedDict + +import paddle.fluid as fluid + +from paddlerec.core.utils import envs +from paddlerec.core.model import Model as ModelBase + + +class Model(ModelBase): + def __init__(self, config): + ModelBase.__init__(self, config) + + def _init_hyper_parameters(self): + self.is_distributed = True if envs.get_trainer( + ) == "CtrTrainer" else False + self.sparse_feature_number = envs.get_global_env( + "hyper_parameters.sparse_feature_number", None) + self.sparse_feature_dim = envs.get_global_env( + "hyper_parameters.sparse_feature_dim", None) + self.reg = envs.get_global_env("hyper_parameters.reg", 1e-4) + self.num_field = envs.get_global_env("hyper_parameters.num_field", + None) + + def net(self, inputs, is_infer=False): + raw_feat_idx = self._sparse_data_var[1] # (batch_size * num_field) * 1 + raw_feat_value = self._dense_data_var[0] # batch_size * num_field + self.label = self._sparse_data_var[0] # batch_size * 1 + + init_value_ = 0.1 + + feat_idx = raw_feat_idx + feat_value = fluid.layers.reshape( + raw_feat_value, + [-1, self.num_field, 1]) # batch_size * num_field * 1 + + # ------------------------- first order term -------------------------- + + first_weights_re = fluid.embedding( + input=feat_idx, + is_sparse=True, + is_distributed=self.is_distributed, + dtype='float32', + size=[self.sparse_feature_number + 1, 1], + padding_idx=0, + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.TruncatedNormalInitializer( + loc=0.0, scale=init_value_), + regularizer=fluid.regularizer.L1DecayRegularizer(self.reg)) + ) # (batch_size * num_field) * 1 * 1(embedding_size) + first_weights = fluid.layers.reshape( + first_weights_re, + shape=[-1, self.num_field, 1]) # batch_size * num_field * 1 + + # ------------------------- second order term -------------------------- + + feat_embeddings_re = fluid.embedding( + input=feat_idx, + is_sparse=True, + is_distributed=self.is_distributed, + dtype='float32', + size=[self.sparse_feature_number + 1, self.sparse_feature_dim], + padding_idx=0, + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.TruncatedNormalInitializer( + loc=0.0, + scale=init_value_ / + math.sqrt(float(self.sparse_feature_dim)))) + ) # (batch_size * num_field) * 1 * embedding_size + feat_embeddings = fluid.layers.reshape( + feat_embeddings_re, + shape=[-1, self.num_field, self.sparse_feature_dim + ]) # batch_size * num_field * embedding_size + feat_embeddings = feat_embeddings * feat_value # batch_size * num_field * embedding_size + + concated = fluid.layers.concat( + [feat_embeddings, first_weights], axis=2) + concated = fluid.layers.reshape( + concated, + shape=[-1, self.num_field * (self.sparse_feature_dim + 1)]) + + fcs = [concated] + hidden_layers = envs.get_global_env("hyper_parameters.fc_sizes") + + for size in hidden_layers: + output = fluid.layers.fc( + input=fcs[-1], + size=size, + act='relu', + param_attr=fluid.ParamAttr( + initializer=fluid.initializer.Normal( + scale=1.0 / math.sqrt(fcs[-1].shape[1])))) + fcs.append(output) + + predict = fluid.layers.fc( + input=fcs[-1], + size=1, + act="sigmoid", + param_attr=fluid.ParamAttr(initializer=fluid.initializer.Normal( + scale=1 / math.sqrt(fcs[-1].shape[1])))) + + self.predict = predict + + cost = fluid.layers.log_loss( + input=self.predict, label=fluid.layers.cast(self.label, + "float32")) # log_loss + avg_cost = fluid.layers.reduce_sum(cost) + + self._cost = avg_cost + + predict_2d = fluid.layers.concat([1 - self.predict, self.predict], 1) + label_int = fluid.layers.cast(self.label, 'int64') + auc_var, batch_auc_var, _ = fluid.layers.auc(input=predict_2d, + label=label_int, + slide_steps=0) + self._metrics["AUC"] = auc_var + self._metrics["BATCH_AUC"] = batch_auc_var + if is_infer: + self._infer_results["AUC"] = auc_var -- GitLab