package protos

Get desktop application:
View/edit binary Protocol Buffers messages

Used in: KerasLayer

• optional uint32 project_dim = 1

• optional MLP transfer_mlp = 2

• optional bool stop_gradient = 3

Used in: EvalMetrics

• optional uint32 num_thresholds = 1

Used in: EvalMetrics

(message has no fields)

Configuration message for the AdagradOptimizer See: https://www.tensorflow.org/api_docs/python/tf/train/AdagradOptimizer

Used in: Optimizer

• optional LearningRate learning_rate = 1

• optional float initial_accumulator_value = 2

Only available on pai-tf, which has better performance than AdamOptimizer

Used in: Optimizer

• optional LearningRate learning_rate = 1

• optional float beta1 = 3

• optional float beta2 = 4

Used in: Optimizer

• optional LearningRate learning_rate = 1

• optional float weight_decay = 2

• optional float beta1 = 3

• optional float beta2 = 4

Configuration message for the AdamOptimizer See: https://www.tensorflow.org/api_docs/python/tf/train/AdamOptimizer

Used in: Optimizer

• optional LearningRate learning_rate = 1

• optional float beta1 = 3

• optional float beta2 = 4

Used in: Optimizer

• optional LearningRate learning_rate = 1

• optional float weight_decay = 2

• optional float beta1 = 3

• optional float beta2 = 4

Used in: KerasLayer

• optional bool use_scale = 1

• optional bool scale_by_dim = 2

• optional string score_mode = 3

• optional float dropout = 4

• optional int32 seed = 5

• optional bool return_attention_scores = 6

• optional bool use_causal_mask = 7

Used in: SequenceCombiner

(message has no fields)

Used in: KerasLayer

• required uint32 embedding_dim = 1

• required uint32 num_bins = 2

• required float keep_prob = 3

• required float temperature = 4

• optional bool output_3d_tensor = 5

• optional bool output_tensor_list = 6

Used in: EasyRecModel

• required uint32 multi_head_num = 1

  The number of heads

• required uint32 multi_head_size = 2

  The dimension of heads

• required uint32 interacting_layer_num = 3

  The number of interacting layers

• required float l2_regularization = 4

Used in: EvalMetrics

• optional uint32 topk = 1

Used in: KerasLayer

• required uint32 hidden_size = 1

  Size of the encoder layers and the pooler layer

• required uint32 num_hidden_layers = 2

  Number of hidden layers in the Transformer encoder

• required uint32 num_attention_heads = 3

  Number of attention heads for each attention layer in the Transformer encoder

• required uint32 intermediate_size = 4

  The size of the "intermediate" (i.e. feed-forward) layer in the Transformer encoder

• required string hidden_act = 5

  The non-linear activation function (function or string) in the encoder and pooler.

  "gelu", "relu", "tanh" and "swish" are supported.

• required float hidden_dropout_prob = 6

  The dropout probability for all fully connected layers in the embeddings, encoder, and pooler

• required float attention_probs_dropout_prob = 7

  The dropout ratio for the attention probabilities

• required uint32 max_position_embeddings = 8

  The maximum sequence length that this model might ever be used with

• required bool use_position_embeddings = 9

  Whether to add position embeddings for the position of each token in the text sequence

• required float initializer_range = 10

  The stddev of the truncated_normal_initializer for initializing all weight matrices

• required bool output_all_token_embeddings = 11

  Whether to output all token embedding, if set to false, then only output the first token embedding

• required string target_item_position = 12

  The position of target item (i.e. head, tail, ignore)

• required bool reserve_target_position = 13

  Whether to preserve a position for target

Used in: MultiTower

• required string input = 1

• required uint32 seq_len = 2

• required uint32 multi_head_size = 3

Used in: EasyRecModel

• repeated BlockPackage packages = 1

  a few sub DAGs

• repeated Block blocks = 2

  a few blocks generating a DAG

• repeated string concat_blocks = 3

  the names of output blocks, will be merge into a tensor

• repeated string output_blocks = 4

  the names of output blocks, return as a list or single tensor

• optional MLP top_mlp = 5

  optional top mlp layer

Used in: DBMTL, ModelParams

• required string tower_name = 1

  task name for the task tower

• optional string label_name = 2

  label for the task, default is label_fields by order

• repeated EvalMetrics metrics_set = 3

  metrics for the task

• optional LossType loss_type = 4

  loss for the task

• optional uint32 num_class = 5

  num_class for multi-class classification loss

• optional DNN dnn = 6

  task specific dnn

• repeated string relation_tower_names = 7

  related tower names

• optional DNN relation_dnn = 8

  relation dnn

• optional float weight = 9

  training loss weights

• optional string task_space_indicator_label = 10

  label name for indicating the sample space for the task tower

• optional float in_task_space_weight = 11

  the loss weight for sample in the task space

• optional float out_task_space_weight = 12

  the loss weight for sample out the task space

• repeated Loss losses = 15

  level for prediction required uint32 prediction_level = 13; prediction weights optional float prediction_weight = 14 [default = 1.0]; multiple losses

• required bool use_sample_weight = 16

  whether to use sample weight in this tower

• optional string task_space_indicator_name = 17

  field name for indicating the sample space for this task

• optional string task_space_indicator_value = 18

  field value for indicating the sample space for this task

Used in: FiBiNet, KerasLayer

• required string type = 1

• required bool use_plus = 2

• required uint32 num_output_units = 3

Used in: EasyRecConfig

• repeated string category_path = 1

  support gfile.Glob

• repeated string dense_path = 2

• repeated string label_path = 3

Used in: KD, Loss

• required float gamma = 1

• optional float alpha = 2

• optional float ohem_ratio = 3

• optional float label_smoothing = 4

Used in: BackboneTower, BlockPackage

• required string name = 1

• repeated Input inputs = 2

  the input names of feature groups or other blocks

• optional int32 input_concat_axis = 3

• optional bool merge_inputs_into_list = 4

• optional string extra_input_fn = 5

• repeated Layer layers = 100

  sequential layers

• oneof layer

  only take effect when there are no layers

  • InputLayer input_layer = 101

  • Lambda lambda = 102

  • KerasLayer keras_layer = 103

  • RecurrentLayer recurrent = 104

  • RepeatLayer repeat = 105

  • RawInputLayer raw_input = 106

  • EmbeddingLayer embedding_layer = 107

a package of blocks for reuse; e.g. call in a contrastive learning manner

Used in: BackboneTower

• required string name = 1

  package name

• repeated Block blocks = 2

  a few blocks generating a DAG

• repeated string concat_blocks = 3

  the names of output blocks, will be merge into a tensor

• repeated string output_blocks = 4

  the names of output blocks, return as a list or single tensor

Used in: KerasLayer

• repeated int32 hidden_feature_sizes = 1

Used in: EasyRecModel

• required CMBFTower config = 1

• required DNN final_dnn = 2

Used in: CMBF, DBMTL

• required uint32 multi_head_num = 1

  The number of heads of cross modal fusion layer

• required uint32 image_multi_head_num = 101

  The number of heads of image feature learning layer

• required uint32 text_multi_head_num = 102

  The number of heads of text feature learning layer

• required uint32 text_head_size = 2

  The dimension of text heads

• required uint32 image_head_size = 3

  The dimension of image heads

• required uint32 image_feature_patch_num = 4

  The number of patches of image feature, take effect when there is only one image feature

• required uint32 image_feature_dim = 5

  Do dimension reduce to this size for image feature before single modal learning module

• required uint32 image_self_attention_layer_num = 6

  The number of self attention layers for image features

• required uint32 text_self_attention_layer_num = 7

  The number of self attention layers for text features

• required uint32 cross_modal_layer_num = 8

  The number of cross modal layers

• required uint32 image_cross_head_size = 9

  The dimension of image cross modal heads

• required uint32 text_cross_head_size = 10

  The dimension of text cross modal heads

• required float hidden_dropout_prob = 11

  Dropout probability for hidden layers

• required float attention_probs_dropout_prob = 12

  Dropout probability of the attention probabilities

• required bool use_token_type = 13

  Whether to add embeddings for different text sequence features

• required bool use_position_embeddings = 14

  Whether to add position embeddings for the position of each token in the text sequence

• required uint32 max_position_embeddings = 15

  Maximum sequence length that might ever be used with this model

• required float text_seq_emb_dropout_prob = 16

  Dropout probability for text sequence embeddings

• optional DNN other_feature_dnn = 17

  dnn layers for other features

Used in: MIND

• optional uint32 max_k = 1

  max number of high capsules

• required uint32 max_seq_len = 2

  max behaviour sequence length

• required uint32 high_dim = 3

  high capsule embedding vector dimension

• optional uint32 num_iters = 4

  number EM iterations

• optional float routing_logits_scale = 5

  routing logits scale

• optional float routing_logits_stddev = 6

  routing logits initial stddev

• optional float squash_pow = 7

  squash power

• optional float scale_ratio = 8

  output ratio

• optional bool const_caps_num = 9

  constant interest number in default, use log(seq_len)

Used in: CoMetricLearningI2I, KD, Loss

• required float margin = 1

• required float gamma = 2

Used in: EasyRecModel

• optional string session_id = 1

• repeated HighWayTower highway = 2

• optional string input = 4

• required DNN dnn = 5

• required float l2_regularization = 6

• required bool output_l2_normalized_emb = 7

• optional string sample_id = 8

• oneof loss

  • CircleLoss circle_loss = 101

  • MultiSimilarityLoss multi_similarity_loss = 102

• optional string item_id = 9

Used in: Initializer

• repeated float consts = 1

Configuration message for a constant learning rate.

Used in: LearningRate

• optional float learning_rate = 1

Configuration message for a cosine decaying learning rate as defined in utils/learning_schedules.py

Used in: LearningRate

• optional float learning_rate_base = 1

• optional uint32 total_steps = 2

• optional float warmup_learning_rate = 3

• optional uint32 warmup_steps = 4

• optional uint32 hold_base_rate_steps = 5

Used in: DCN

• required string input = 1

• required uint32 cross_num = 2

  The number of cross layers

Used in: EasyRecModel

• required DATTower user_tower = 1

• required DATTower item_tower = 2

• required float l2_regularization = 3

• optional Similarity simi_func = 4

• required bool ignore_in_batch_neg_sam = 5

• optional float temperature = 6

• required float amm_i_weight = 7

  loss weight for amm_i

• required float amm_u_weight = 8

  loss weight for amm_u

Used in: DAT

• required string id = 1

• required DNN dnn = 2

Used in: EasyRecModel

• optional CMBFTower bottom_cmbf = 101

  shared bottom cmbf layer

• optional UniterTower bottom_uniter = 102

  shared bottom uniter layer

• optional DNN bottom_dnn = 1

  shared bottom dnn layer

• optional DNN expert_dnn = 2

  mmoe expert dnn layer definition

• optional uint32 num_expert = 3

  number of mmoe experts

• repeated BayesTaskTower task_towers = 4

  bayes task tower

• optional float l2_regularization = 5

  l2 regularization

Used in: EasyRecModel

• required Tower deep_tower = 1

• required CrossTower cross_tower = 2

• required DNN final_dnn = 3

• required float l2_regularization = 5

Used in: KerasLayer

• required MLP attention_dnn = 1

  din attention layer

• required bool need_target_feature = 2

  whether to keep target item feature

• required string attention_normalizer = 3

  option: softmax, sigmoid

Used in: MultiTower

• required string input = 1

• required DNN dnn = 2

Used in: EasyRecModel

• required DNN top_dnn = 1

• required DNN bot_dnn = 2

• optional string arch_interaction_op = 3

  options are: dot and cat

• optional bool arch_interaction_itself = 4

  whether a feature will interact with itself

• optional bool arch_with_dense_feature = 5

  whether to include dense features after interaction

• optional float l2_regularization = 10

Used in: BayesTaskTower, CMBF, CMBFTower, CoMetricLearningI2I, DATTower, DBMTL, DCN, DINTower, DLRM, DSSMTower, DSSM_SENet_Tower, DeepFM, DropoutNet, ExpertTower, ExtractionNetwork, MIND, MMoE, MultiTower, MultiTowerRecall, PDN, RecallTower, RocketLaunching, SeqAttGroupConfig, TaskTower, Tower, Uniter, UniterTower, WideAndDeep

• repeated uint32 hidden_units = 1

  hidden units for each layer

• repeated float dropout_ratio = 2

  ratio of dropout

• optional string activation = 3

  activation function

• optional bool use_bn = 4

  use batch normalization

Used in: EasyRecModel

• required DSSMTower user_tower = 1

• required DSSMTower item_tower = 2

• required float l2_regularization = 3

• optional Similarity simi_func = 4

• optional bool scale_simi = 5

  add a layer for scaling the similarity

• optional string item_id = 9

• required bool ignore_in_batch_neg_sam = 10

• optional float temperature = 11

  normalize user_tower_embedding and item_tower_embedding

Used in: DSSM

• required string id = 1

• required DNN dnn = 2

Used in: EasyRecModel

• required DSSM_SENet_Tower user_tower = 1

• required DSSM_SENet_Tower item_tower = 2

• required float l2_regularization = 3

• optional Similarity simi_func = 4

• optional bool scale_simi = 5

  add a layer for scaling the similarity

• optional string item_id = 9

• required bool ignore_in_batch_neg_sam = 10

• optional float temperature = 11

  normalize user_tower_embedding and item_tower_embedding

Used in: DSSM_SENet

• required string id = 1

• required SENet senet = 2

• required DNN dnn = 3

Used in: EasyRecConfig

• required string akId = 1

• required string akSecret = 2

• required string endpoint = 3

• required string project = 4

• required string topic = 5

• oneof offset

  • string offset_info = 60

    in json format: {"0":{"cursor": ""}, "1":{"cursor":""}}

  • string offset_time = 62

    offset_time could be two formats: 1: %Y%m%d %H:%M:%S "20220508 12:00:00" 2: %s "1651982400"

Used in: EasyRecConfig

• optional uint32 batch_size = 1

  mini batch size to use for training and evaluation.

• optional bool auto_expand_input_fields = 3

  set auto_expand_input_fields to true to auto_expand field[1-21] to field1, field2, ..., field21

• repeated string label_fields = 4

  label fields, normally only one field is used. For multiple target models such as MMOE multiple label_fields will be set.

• repeated string label_sep = 41

  label separator

• repeated uint32 label_dim = 42

  label dimensions which need to be set when there are labels have dimension > 1

• repeated DatasetConfig.LabelFunction extra_label_func = 43

  extra transformation functions that generate new labels

• optional bool shuffle = 5

  whether to shuffle data

• optional int32 shuffle_buffer_size = 11

  shufffle buffer for better performance, even shuffle buffer is set, it is suggested to do full data shuffle before training especially when the performance of models is not good.

• optional uint32 num_epochs = 6

  The number of times a data source is read. If set to zero, the data source will be reused indefinitely.

• optional uint32 prefetch_size = 7

  Number of decoded batches to prefetch.

• optional bool shard = 801

  shard dataset to 1/num_workers in distribute mode this param is not used anymore

• optional bool file_shard = 802

  shard by file, not by sample, valid only for CSVInput

• required DatasetConfig.InputType input_type = 10

• optional string separator = 12

  separator of column features, only used for CSVInput* not used in OdpsInput* binary separators are supported: CTRL+A could be set as '\001' CTRL+B could be set as '\002' CTRL+C could be set as '\003' for RTPInput and OdpsRTPInput it is usually set to '\002'

• optional uint32 num_parallel_calls = 13

  parallel preproces of raw data, avoid using too small or too large numbers(suggested be to small than number of the cores)

• optional string selected_cols = 14

  only used for OdpsInput/OdpsInputV2/OdpsRTPInput, comma separated for RTPInput, selected_cols use indices as column names such as '1,2,4', where 1,2 are label columns, and 4 is the feature column, column 0,3 are not used,

• optional string selected_col_types = 15

  selected col types, only used for OdpsInput/OdpsInputV2 to avoid error setting of data types

• repeated DatasetConfig.Field input_fields = 16

  the input fields must be the same number and in the same order as data in csv files or odps tables

• optional string rtp_separator = 17

  for RTPInput only

• optional bool ignore_error = 18

  ignore some data errors it is not suggested to set this parameter

• optional bool pai_worker_queue = 19

  whether to use pai global shuffle queue, only for OdpsInput, OdpsInputV2, OdpsRTPInputV2

• optional int32 pai_worker_slice_num = 20

• optional bool chief_redundant = 21

  if true, one worker will duplicate the data of the chief node and undertake the gradient computation of the chief node

• optional string sample_weight = 22

  input field for sample weight

• optional string data_compression_type = 23

  the compression type of tfrecord

• optional uint32 n_data_batch_tfrecord = 24

  n data for one feature in tfrecord

• optional bool with_header = 25

  for csv files, may optionally with an header in that case, input_name must match header name, and the number and the order of input_fields may not be the same as that in csv files.

• repeated string feature_fields = 26

• oneof sampler

  • NegativeSampler negative_sampler = 101

  • NegativeSamplerV2 negative_sampler_v2 = 102

  • HardNegativeSampler hard_negative_sampler = 103

  • HardNegativeSamplerV2 hard_negative_sampler_v2 = 104

  • NegativeSamplerInMemory negative_sampler_in_memory = 105

• optional uint32 eval_batch_size = 1001

• optional bool drop_remainder = 1002

Used in: DatasetConfig

• required string input_name = 1

• required FieldType input_type = 2

• optional string default_val = 3

• optional uint32 input_dim = 4

• optional uint32 input_shape = 5

• optional string user_define_fn = 6

  user-defined function for label. eg: tf.math.log1p, remap_lbl

• optional string user_define_fn_path = 7

  user-defined function path. eg: /samples/demo_script/process_lbl.py

• optional FieldType user_define_fn_res_type = 8

  output field type of user-defined function.

• optional string ignore_val = 9

  ignore value

Used in: Field

• INT32 = 0

• INT64 = 1

• STRING = 2

• FLOAT = 4

• DOUBLE = 5

• BOOL = 6

Used in: DatasetConfig

• CSVInput = 10

  csv format input, could be used in local or hdfs support .gz compression(but not .tar.gz files)

• CSVInputV2 = 11

  @Depreciated

• CSVInputEx = 12

  extended csv format, allow quote in fields

• OdpsInput = 2

  @Depreciated, has memory leak problem

• OdpsInputV2 = 3

  odps input, used on pai

• DataHubInput = 15

• OdpsInputV3 = 9

• RTPInput = 4

• RTPInputV2 = 5

• OdpsRTPInput = 601

• OdpsRTPInputV2 = 602

• TFRecordInput = 7

• BatchTFRecordInput = 14

• DummyInput = 8

  for the purpose to debug performance bottleneck of input pipelines

• KafkaInput = 13

• HiveInput = 16

• HiveRTPInput = 17

• HiveParquetInput = 18

• ParquetInput = 19

  All features are packed into one field for fast copying to gpu, and there are no feature preprocessing step, it is assumed that features are preprocessed before training. Requirements: python3 and tf2.x due to multiprocssing spawn and RaggedTensor apis.

• ParquetInputV2 = 20

  Features are not packed, and are preprocessing separately. Requirements: python3 and tf2.x due to multiprocssing spawn and RaggedTensor apis.

• ParquetInputV3 = 21

  c++ version of parquet dataset which currently are only available with deeprec.

• CriteoInput = 1001

Used in: DatasetConfig

• required string label_name = 1

• required string label_func = 2

Used in: EasyRecModel

• required DNN dnn = 1

• optional DNN final_dnn = 2

• optional uint32 wide_output_dim = 3

• optional float wide_regularization = 4

  deprecated

• optional float dense_regularization = 5

  deprecated

• optional float l2_regularization = 6

Used in: TrainConfig

• NoStrategy = 0

  use old SyncReplicasOptimizer for ParameterServer training

• PSStrategy = 1

  PSStrategy with multiple gpus on one node could not work on pai-tf, could only work on TF >=1.15

• MirroredStrategy = 2

  could only work on PaiTF or TF >=1.15 single worker multiple gpu mode

• CollectiveAllReduceStrategy = 3

  Depreciated

• ExascaleStrategy = 4

  currently not working good

• MultiWorkerMirroredStrategy = 5

  multi worker multi gpu mode see tf.distribute.experimental.MultiWorkerMirroredStrategy

• HorovodStrategy = 6

  use horovod strategy

• SokStrategy = 7

  support kv embedding, support kv embedding shard

• EmbeddingParallelStrategy = 8

  support embedding shard, requires horovod

Used in: EasyRecModel

• required DNN user_content = 1

• required DNN user_preference = 2

• required DNN item_content = 3

• required DNN item_preference = 4

• required DNN user_tower = 5

• required DNN item_tower = 6

• required float l2_regularization = 7

• required float user_dropout_rate = 8

• required float item_dropout_rate = 9

• optional SoftmaxCrossEntropyWithNegativeMining softmax_loss = 10

for input performance test

Used in: EasyRecModel

(message has no fields)

Used in: EasyRecModel

• repeated Tower groups = 1

• required TaskTower ctr_tower = 2

• required TaskTower cvr_tower = 3

• required float l2_regularization = 4

Used in: EasyRecModel, FeatureConfig

• optional uint64 filter_freq = 1

• optional uint64 steps_to_live = 2

• optional bool use_cache = 3

  use embedding cache, only for sok hybrid embedding

• optional uint64 init_capacity = 4

  for sok hybrid key value embedding

• optional uint64 max_capacity = 5

• oneof train_path

  • string train_input_path = 1

  • KafkaServer kafka_train_input = 2

  • DatahubServer datahub_train_input = 12

  • HiveConfig hive_train_input = 101

  • BinaryDataInput binary_train_input = 102

  • string parquet_train_input = 103

• oneof eval_path

  • string eval_input_path = 3

  • KafkaServer kafka_eval_input = 4

  • DatahubServer datahub_eval_input = 13

  • HiveConfig hive_eval_input = 201

  • BinaryDataInput binary_eval_input = 202

  • string parquet_eval_input = 203

• required string model_dir = 5

• optional TrainConfig train_config = 6

  train config, including optimizer, weight decay, num_steps and so on

• optional EvalConfig eval_config = 7

• optional DatasetConfig data_config = 8

• repeated FeatureConfig feature_configs = 9

  for compatibility

• optional FeatureConfigV2 feature_config = 10

• required EasyRecModel model_config = 14

  recommendation model config

• optional ExportConfig export_config = 15

• optional string fg_json_path = 16

  Json file[RTP FG] to define input data and features: * In easy_rec.python.utils.fg_util.load_fg_json_to_config: data_config and feature_config will be generated based on fg_json. * After generation, a prefix '!' is added: fg_json_path = '!' + fg_json_path indicates config update is already done, and should not be updated anymore. In this way, we make load_fg_json_to_config function reentrant. This step is done before edit_config_json to take effect.

Used in: EasyRecConfig

• required string model_class = 1

• optional string model_name = 99

  just a name for backbone config

• repeated FeatureGroupConfig feature_groups = 2

  actually input layers, each layer produce a group of feature

• oneof model

  model parameters

  • ModelParams model_params = 100

  • DummyModel dummy = 101

  • WideAndDeep wide_and_deep = 102

  • DeepFM deepfm = 103

  • MultiTower multi_tower = 104

  • FM fm = 105

  • DCN dcn = 106

  • AutoInt autoint = 107

  • DLRM dlrm = 108

  • CMBF cmbf = 109

  • Uniter uniter = 110

  • MultiTowerRecall multi_tower_recall = 200

  • DSSM dssm = 201

  • MIND mind = 202

  • DropoutNet dropoutnet = 203

  • CoMetricLearningI2I metric_learning = 204

  • PDN pdn = 205

  • DSSM_SENet dssm_senet = 206

  • DAT dat = 207

  • MMoE mmoe = 301

  • ESMM esmm = 302

  • DBMTL dbmtl = 303

  • SimpleMultiTask simple_multi_task = 304

  • PLE ple = 305

  • RocketLaunching rocket_launching = 401

• repeated SeqAttGroupConfig seq_att_groups = 7

• optional float embedding_regularization = 8

  implemented in easy_rec/python/model/easy_rec_estimator add regularization to all variables with "embedding_weights:" in name

• optional LossType loss_type = 9

• optional uint32 num_class = 10

• optional EVParams ev_params = 11

• repeated KD kd = 12

• repeated string restore_filters = 13

  filter variables matching any pattern in restore_filters common filters are Adam, Momentum, etc.

• optional VariationalDropoutLayer variational_dropout = 14

• repeated Loss losses = 15

• required EasyRecModel.LossWeightStrategy loss_weight_strategy = 16

• optional BackboneTower backbone = 17

• optional string label_name = 18

  label name for rank_model to select one label between multiple labels

Used in: EasyRecModel

• Fixed = 0

• Uncertainty = 1

• Random = 2

Used in: Block

• required uint32 embedding_dim = 1

• optional uint32 vocab_size = 2

• optional string combiner = 3

• optional bool concat = 4

Message for configuring EasyRecModel evaluation jobs (eval.py).

Used in: EasyRecConfig

• optional uint32 num_examples = 1

  Number of examples to process of evaluation.

• optional uint32 eval_interval_secs = 2

  How often to run evaluation.

• optional uint32 max_evals = 3

  Maximum number of times to run evaluation. If set to 0, will run forever.

• optional bool save_graph = 4

  Whether the TensorFlow graph used for evaluation should be saved to disk.

• repeated EvalMetrics metrics_set = 5

  Type of metrics to use for evaluation. possible values:

• optional bool eval_online = 6

  Evaluation online with batch forward data of training

Used in: BayesTaskTower, EvalConfig, TaskTower

• oneof metric

  • AUC auc = 1

  • RecallAtTopK recall_at_topk = 2

  • MeanAbsoluteError mean_absolute_error = 3

  • MeanSquaredError mean_squared_error = 4

  • Accuracy accuracy = 5

  • Max_F1 max_f1 = 6

  • RootMeanSquaredError root_mean_squared_error = 7

  • GAUC gauc = 8

  • SessionAUC session_auc = 9

  • Recall recall = 10

  • Precision precision = 11

  • AvgPrecisionAtTopK precision_at_topk = 12

Used in: MMoE

• required string expert_name = 1

• required DNN dnn = 2

Configuration message for an exponentially decaying learning rate. See https://www.tensorflow.org/versions/master/api_docs/python/train/ \ decaying_the_learning_rate#exponential_decay

Used in: LearningRate

• optional float initial_learning_rate = 1

• optional uint32 decay_steps = 2

• optional float decay_factor = 3

• optional bool staircase = 4

• optional float burnin_learning_rate = 5

• optional uint32 burnin_steps = 6

• optional float min_learning_rate = 7

Message for configuring exporting models.

Used in: EasyRecConfig

• optional int32 batch_size = 1

  batch size used for exported model, -1 indicates batch_size is None which is only supported by classification model right now, while other models support static batch_size

• optional string exporter_type = 2

  type of exporter [final | latest | best | none] when train_and_evaluation final: performs a single export in the end of training latest: regularly exports the serving graph and checkpoints best: export the best model according to best_exporter_metric none: do not perform export

• optional string best_exporter_metric = 4

  the metric used to determine the best checkpoint

• optional bool metric_bigger = 5

  metric value the bigger the best

• optional bool enable_early_stop = 6

  enable early stop

• optional string early_stop_func = 601

  custom early stop function, format: early_stop_func(eval_results, early_stop_params) return True if should stop

• optional string early_stop_params = 602

  custom early stop parameters

• optional int32 max_check_steps = 7

  early stop max check steps

• optional bool multi_placeholder = 8

  each feature has a placeholder

• optional int32 exports_to_keep = 9

  export to keep, only for exporter_type in [best, latest]

• optional MultiValueFields multi_value_fields = 10

  multi value field list

• optional bool auto_multi_value = 16

  auto analyze multi value fields

• optional bool placeholder_named_by_input = 11

  is placeholder named by input

• optional bool filter_inputs = 12

  filter out inputs, only keep effective ones

• optional bool export_features = 13

  export the original feature values as string

• optional bool export_rtp_outputs = 14

  export the outputs required by RTP

• repeated string asset_files = 15

  export asset files

Used in: PLE

• required string network_name = 1

• required uint32 expert_num_per_task = 2

  number of experts per task

• optional uint32 share_num = 3

  number of experts for share For the last extraction_network, no need to configure this

• required DNN task_expert_net = 4

  dnn network of experts per task

• optional DNN share_expert_net = 5

  dnn network of experts for share For the last extraction_network, no need to configure this

Used in: KD, Loss

• required float f1_beta_square = 1

• required float label_smoothing = 2

Used in: EasyRecModel, KerasLayer

• optional bool use_variant = 1

• optional float l2_regularization = 5

Used in: EasyRecConfig, FeatureConfigV2

• optional string feature_name = 1

• repeated string input_names = 2

  input field names: must be included in DatasetConfig.input_fields

• required FeatureConfig.FeatureType feature_type = 3

• optional string embedding_name = 4

• optional uint32 embedding_dim = 5

• optional uint64 hash_bucket_size = 6

• optional uint64 num_buckets = 7

  for categorical_column_with_identity

• repeated double boundaries = 8

  only for raw features

• optional string separator = 9

  separator with in features

• optional string kv_separator = 10

  delimeter to separator key from value

• optional string seq_multi_sep = 101

  delimeter to separate sequence multi-values

• optional uint32 max_seq_len = 102

  truncate sequence data to max_seq_len

• optional string vocab_file = 11

• repeated string vocab_list = 12

• repeated string shared_names = 16

  many other field share this config

• optional int32 lookup_max_sel_elem_num = 17

  lookup max select element number, default 10

• optional int32 max_partitions = 18

  max_partitions

• optional string combiner = 19

  combiner

• optional Initializer initializer = 20

  embedding initializer

• optional int32 precision = 21

  number of digits kept after dot in format float/double to string scientific format is not used. in default it is not allowed to convert float/double to string

• optional double min_val = 212

  normalize raw feature to [0-1]

• optional double max_val = 213

• optional string normalizer_fn = 214

  normalization function for raw features: such as: tf.math.log1p

• optional uint32 raw_input_dim = 24

  raw feature of multiple dimensions

• optional SequenceCombiner sequence_combiner = 25

  sequence feature combiner

• optional FeatureConfig.FeatureType sub_feature_type = 26

  sub feature type for sequence feature

• optional uint32 sequence_length = 27

  sequence length

• optional string expression = 30

  for expr feature

• optional EVParams ev_params = 31

  embedding variable params

• optional string combo_join_sep = 401

  for combo feature: if not set, use cross_column otherwise, the input features are first joined and then passed to categorical_column

• repeated string combo_input_seps = 402

  separator for each inputs if not set, combo inputs will not be split

Used in: FeatureConfig

• IdFeature = 0

• RawFeature = 1

• TagFeature = 2

• ComboFeature = 3

• LookupFeature = 4

• SequenceFeature = 5

• ExprFeature = 6

• PassThroughFeature = 7

• INT32 = 0

• INT64 = 1

• STRING = 2

• FLOAT = 4

• DOUBLE = 5

• BOOL = 6

Used in: EasyRecConfig

• repeated FeatureConfig features = 1

• optional bool embedding_on_cpu = 2

  force place embedding lookup ops on cpu to improve training and inference efficiency.

Used in: EasyRecModel

• optional string group_name = 1

• repeated string feature_names = 2

• optional WideOrDeep wide_deep = 3

• repeated SeqAttGroupConfig sequence_features = 4

• optional bool negative_sampler = 5

Used in: KerasLayer

• optional Bilinear bilinear = 1

• required SENet senet = 2

• optional MLP mlp = 8

Used in: Optimizer

• optional LearningRate learning_rate = 1

  optional float learning_rate = 1 [default=1e-4];

• optional float learning_rate_power = 2

• optional float initial_accumulator_value = 3

• optional float l1_reg = 4

• optional float l2_reg = 5

• optional float l2_shrinkage_reg = 6

Used in: EvalMetrics

• required string uid_field = 1

  uid field name

• optional string reduction = 2

  reduction method for auc of different users * "mean": simple mean of different users * "mean_by_sample_num": weighted mean with sample num of different users * "mean_by_positive_num": weighted mean with positive sample num of different users

used in PPNet

Used in: PPNet

• optional uint32 output_dim = 1

• optional uint32 hidden_dim = 2

• optional string activation = 3

  activation function

• optional bool use_bn = 4

  use batch normalization

• optional float dropout_rate = 5

Used in: Initializer

(message has no fields)

Weighted Random Sampling ItemID not in Batch and Sampling Hard Edge

Used in: DatasetConfig

• required string user_input_path = 1

  user data path userid weight

• required string item_input_path = 2

  item data path itemid weight attrs

• required string hard_neg_edge_input_path = 3

  hard negative edge path userid itemid weight

• required uint32 num_sample = 4

  number of negative sample

• required uint32 num_hard_sample = 5

  max number of hard negative sample

• repeated string attr_fields = 6

  field names of attrs in train data or eval data

• required string item_id_field = 7

  field name of item_id in train data or eval data

• required string user_id_field = 8

  field name of user_id in train data or eval data

• optional string attr_delimiter = 9

• optional uint32 num_eval_sample = 10

• optional string field_delimiter = 11

  only works on DataScience/Local

Weighted Random Sampling ItemID not with Edge and Sampling Hard Edge

Used in: DatasetConfig

• required string user_input_path = 1

  user data path userid weight

• required string item_input_path = 2

  item data path itemid weight attrs

• required string pos_edge_input_path = 3

  positive edge path userid itemid weight

• required string hard_neg_edge_input_path = 4

  hard negative edge path userid itemid weight

• required uint32 num_sample = 5

  number of negative sample

• required uint32 num_hard_sample = 6

  max number of hard negative sample

• repeated string attr_fields = 7

  field names of attrs in train data or eval data

• required string item_id_field = 8

  field name of item_id in train data or eval data

• required string user_id_field = 9

  field name of user_id in train data or eval data

• optional string attr_delimiter = 10

• optional uint32 num_eval_sample = 11

• optional string field_delimiter = 12

  only works on DataScience/Local

Used in: CoMetricLearningI2I, KerasLayer

• optional string input = 1

• required uint32 emb_size = 2

• required string activation = 3

• optional float dropout_rate = 4

• optional float init_gate_bias = 5

• optional uint32 num_layers = 6

Used in: EasyRecConfig

• required string host = 1

  hive master's ip

• required uint32 port = 2

  hive port

• required string username = 3

  hive username

• required string database = 4

  hive database

• required string table_name = 5

Used in: TrainConfig

• optional int32 sparse_save_secs = 1

• optional int32 dense_save_secs = 2

• optional int32 sparse_save_steps = 3

• optional int32 dense_save_steps = 4

• optional bool debug_save_update = 5

  if open, will save increment updates to model_dir/incr_save/

• oneof incr_update

  • IncrementSaveConfig.Kafka kafka = 501

  • IncrementSaveConfig.Datahub datahub = 502

  • IncrementSaveConfig.File fs = 503

Used in: IncrementSaveConfig

• required string akId = 1

• required string akSecret = 2

• required string region = 3

• required string project = 4

• required string topic = 5

• required Datahub.Consumer consumer = 6

Used in: Datahub

• optional int64 offset = 1

• optional int32 timeout = 2

Used in: IncrementSaveConfig

• optional string incr_save_dir = 1

• optional bool relative = 2

  relative to model_dir

• optional string mount_path = 3

  for online inference, please set the storage.mount_path to mount_path online service will fail

Used in: IncrementSaveConfig

• required string server = 1

• required string topic = 2

• required Kafka.Consumer consumer = 3

Used in: Kafka

• optional string config_topic = 1

• optional string config_global = 2

• optional int64 offset = 3

• optional int32 timeout = 4

Proto with one-of field for initializers.

Used in: FeatureConfig

• oneof initializer_oneof

  • TruncatedNormalInitializer truncated_normal_initializer = 1

  • RandomNormalInitializer random_normal_initializer = 2

  • GlorotNormalInitializer glorot_normal_initializer = 3

  • ConstantInitializer constant_initializer = 4

Used in: Block

• oneof name

  • string feature_group_name = 1

  • string block_name = 2

  • string package_name = 3

  • bool use_package_input = 4

• optional string input_fn = 11

• optional string input_slice = 12

• optional bool ignore_input = 13

• optional InputLayer reset_input = 14

• optional string package_input = 15

• optional string package_input_fn = 16

Used in: Block, Input

• optional bool do_batch_norm = 1

• optional bool do_layer_norm = 2

• optional float dropout_rate = 3

• optional float feature_dropout_rate = 4

• optional bool only_output_feature_list = 5

• optional bool only_output_3d_tensor = 6

• optional bool output_2d_tensor_and_feature_list = 7

• optional bool output_seq_and_normal_feature = 8

• optional uint32 wide_output_dim = 9

• optional bool concat_seq_feature = 10

Used in: KD, Loss

• required string session_name = 1

• optional float alpha = 2

• optional bool same_label_loss = 3

• required string loss_weight_strategy = 4

for knowledge distillation

Used in: EasyRecModel

• optional string loss_name = 10

• required string pred_name = 11

• optional bool pred_is_logits = 12

  default to be logits

• required string soft_label_name = 21

  for CROSS_ENTROPY_LOSS, soft_label must be logits instead of probs

• optional bool label_is_logits = 22

  default to be logits

• required LossType loss_type = 3

• optional float loss_weight = 4

• optional float temperature = 5

  only for loss_type == CROSS_ENTROPY_LOSS or BINARY_CROSS_ENTROPY_LOSS or KL_DIVERGENCE_LOSS

• optional string task_space_indicator_name = 6

  field name for indicating the sample space for this task

• optional string task_space_indicator_value = 7

  field value for indicating the sample space for this task

• optional float in_task_space_weight = 8

  the loss weight for sample in the task space

• optional float out_task_space_weight = 9

  the loss weight for sample out the task space

• oneof loss_param

  • F1ReweighedLoss f1_reweighted_loss = 101

  • SoftmaxCrossEntropyWithNegativeMining softmax_loss = 102

  • CircleLoss circle_loss = 103

  • MultiSimilarityLoss multi_simi_loss = 104

  • BinaryFocalLoss binary_focal_loss = 105

  • PairwiseLoss pairwise_loss = 106

  • PairwiseFocalLoss pairwise_focal_loss = 107

  • PairwiseLogisticLoss pairwise_logistic_loss = 108

  • JRCLoss jrc_loss = 109

  • PairwiseHingeLoss pairwise_hinge_loss = 110

  • ListwiseRankLoss listwise_rank_loss = 111

  • ListwiseDistillLoss listwise_distill_loss = 112

Used in: EasyRecConfig

• required string server = 1

• required string topic = 2

• required string group = 3

• oneof offset

  • string offset_info = 40

    in json format: {'0':10, '1':20}

  • string offset_time = 42

    offset_time could be two formats: 1: %Y%m%d %H:%M:%S '20220508 12:00:00' 2: %s '1651982400'

• repeated string config_global = 5

  kafka global config, such as: fetch.max.bytes=1024

• repeated string config_topic = 6

  kafka topic config, such as: max.partition.fetch.bytes=1024

Used in: Block, Layer, RecurrentLayer, RepeatLayer

• required string class_name = 1

• oneof params

  • google.protobuf.Struct st_params = 2

  • PeriodicEmbedding periodic_embedding = 3

  • AutoDisEmbedding auto_dis_embedding = 4

  • NaryDisEmbedding nary_dis_embedding = 21

  • FM fm = 5

  • MaskBlock mask_block = 6

  • MaskNet masknet = 7

  • SENet senet = 8

  • Bilinear bilinear = 9

  • FiBiNet fibinet = 10

  • MLP mlp = 11

  • DINEncoder din = 12

  • BSTEncoder bst = 13

  • MMoELayer mmoe = 14

  • SequenceAugment seq_aug = 15

  • PPNet ppnet = 16

  • TextCNN text_cnn = 17

  • HighWayTower highway = 18

  • OverlapFeature overlap = 19

  • MappedDotProduct dot_product = 20

  • Attention attention = 22

  • MultiHeadAttention multi_head_attention = 23

  • Transformer transformer = 24

  • TextEncoder text_encoder = 25

  • WeightedGate gate = 26

  • AITMTower aitm = 27

  • CIN cin = 28

Configuration proto for L2 Regularizer.

Used in: Regularizer

• optional float scale_l1 = 1

• optional float scale_l2 = 2

Configuration proto for L1 Regularizer.

Used in: Regularizer

• optional float scale = 1

Configuration proto for L2 Regularizer.

Used in: Regularizer

• optional float scale = 1

Used in: Block, Layer

• required string expression = 1

Used in: Block

• oneof layer

  • Lambda lambda = 1

  • KerasLayer keras_layer = 2

  • RecurrentLayer recurrent = 3

  • RepeatLayer repeat = 4

Used in: Optimizer

• optional LearningRate learning_rate = 1

• optional float beta1 = 3

• optional float beta2 = 4

Configuration message for optimizer learning rate.

Used in: AdagradOptimizer, AdamAsyncOptimizer, AdamAsyncWOptimizer, AdamOptimizer, AdamWOptimizer, FtrlOptimizer, LazyAdamOptimizer, MomentumOptimizer, MomentumWOptimizer, RMSPropOptimizer

• oneof learning_rate

  • ConstantLearningRate constant_learning_rate = 1

  • ExponentialDecayLearningRate exponential_decay_learning_rate = 2

  • ManualStepLearningRate manual_step_learning_rate = 3

  • CosineDecayLearningRate cosine_decay_learning_rate = 4

  • PolyDecayLearningRate poly_decay_learning_rate = 5

  • TransformerLearningRate transformer_learning_rate = 6

Used in: KD, Loss

• required float temperature = 1

• optional string session_name = 2

• optional string transform_fn = 3

• optional float label_clip_max_value = 4

• optional bool scale_logits = 5

Used in: KD, Loss

• required float temperature = 1

• optional string session_name = 2

• optional string transform_fn = 3

• optional bool label_is_logits = 4

• optional bool scale_logits = 5

Used in: BayesTaskTower, EasyRecModel, TaskTower

• required LossType loss_type = 1

• optional float weight = 2

• optional string loss_name = 3

• optional bool learn_loss_weight = 4

• oneof loss_param

  • F1ReweighedLoss f1_reweighted_loss = 101

  • SoftmaxCrossEntropyWithNegativeMining softmax_loss = 102

  • CircleLoss circle_loss = 103

  • MultiSimilarityLoss multi_simi_loss = 104

  • BinaryFocalLoss binary_focal_loss = 105

  • PairwiseLoss pairwise_loss = 106

  • PairwiseFocalLoss pairwise_focal_loss = 107

  • PairwiseLogisticLoss pairwise_logistic_loss = 108

  • JRCLoss jrc_loss = 109

  • PairwiseHingeLoss pairwise_hinge_loss = 110

  • ListwiseRankLoss listwise_rank_loss = 111

  • ListwiseDistillLoss listwise_distill_loss = 112

Used in: BayesTaskTower, EasyRecModel, KD, Loss, TaskTower

• CLASSIFICATION = 0

• L2_LOSS = 1

• SIGMOID_L2_LOSS = 2

• CROSS_ENTROPY_LOSS = 3

  crossentropy loss/log loss

• SOFTMAX_CROSS_ENTROPY = 4

• CIRCLE_LOSS = 5

• MULTI_SIMILARITY_LOSS = 6

• SOFTMAX_CROSS_ENTROPY_WITH_NEGATIVE_MINING = 7

• PAIR_WISE_LOSS = 8

• F1_REWEIGHTED_LOSS = 9

• BINARY_FOCAL_LOSS = 10

• PAIRWISE_FOCAL_LOSS = 11

• PAIRWISE_LOGISTIC_LOSS = 12

• PAIRWISE_HINGE_LOSS = 17

• JRC_LOSS = 13

• ORDER_CALIBRATE_LOSS = 14

• BINARY_CROSS_ENTROPY_LOSS = 15

• KL_DIVERGENCE_LOSS = 16

• LISTWISE_RANK_LOSS = 18

• LISTWISE_DISTILL_LOSS = 19

• ZILN_LOSS = 20

Used in: EasyRecModel

• optional DNN pre_capsule_dnn = 101

  preprocessing dnn before entering capsule layer

• required DNN user_dnn = 102

  dnn layers applied on user_context(none sequence features)

• required DNN concat_dnn = 103

  concat user and capsule dnn

• optional MIND.UserSeqCombineMethod user_seq_combine = 104

  method to combine several user sequences such as item_ids, category_ids

• required DNN item_dnn = 2

  dnn layers applied on item features

• required Capsule capsule_config = 3

• optional float simi_pow = 4

  similarity power, the paper says that the big the better

• optional Similarity simi_func = 5

• optional bool scale_simi = 6

  add a layer for scaling the similarity

• required float l2_regularization = 7

• optional string time_id_fea = 8

• optional string item_id = 9

• optional bool ignore_in_batch_neg_sam = 10

• optional float max_interests_simi = 11

  if small than 1.0, then a loss will be added to limit the maximal interest similarities, but in experiments, setup such a loss leads to low hitrate.

Used in: MIND

• CONCAT = 0

• SUM = 1

Used in: AITMTower, BackboneTower, DINEncoder, FiBiNet, KerasLayer, MMoELayer, MaskNet, PPNet, TextCNN, WeightedGate

• repeated uint32 hidden_units = 1

  hidden units for each layer

• repeated float dropout_ratio = 2

  ratio of dropout

• optional string activation = 3

  activation function

• optional bool use_bn = 4

  use batch normalization

• optional bool use_final_bn = 5

• optional string final_activation = 6

• optional bool use_bias = 7

• optional string initializer = 8

  kernel_initializer

• optional bool use_bn_after_activation = 9

• optional bool use_final_bias = 10

• optional bool add_to_outputs = 11

Used in: EasyRecModel

• repeated ExpertTower experts = 1

  deprecated: original mmoe experts config

• optional DNN expert_dnn = 2

  mmoe expert dnn layer definition

• optional uint32 num_expert = 3

  number of mmoe experts

• repeated TaskTower task_towers = 4

  task tower

• required float l2_regularization = 5

  l2 regularization

Used in: KerasLayer

• required uint32 num_task = 1

  number of tasks

• optional MLP expert_mlp = 2

  mmoe expert mlp layer definition

• optional uint32 num_expert = 3

  number of mmoe experts

Configuration message for a manually defined learning rate schedule.

Used in: LearningRate

• optional float initial_learning_rate = 1

• repeated ManualStepLearningRate.LearningRateSchedule schedule = 2

• optional bool warmup = 3

  Whether to linearly interpolate learning rates for steps in [0, schedule[0].step].

Used in: ManualStepLearningRate

• optional uint32 step = 1

• optional float learning_rate = 2

Used in: KerasLayer

• optional string separator = 1

• optional float default_value = 2

• optional string normalize_fn = 3

• repeated float boundaries = 4

• optional int32 embedding_dim = 5

• optional int32 print_first_n = 6

• optional int32 summarize = 7

Used in: KerasLayer, MaskNet

• optional float reduction_factor = 1

• optional uint32 output_size = 2

• optional uint32 aggregation_size = 3

• optional bool input_layer_norm = 4

• optional uint32 projection_dim = 5

Used in: KerasLayer

• repeated MaskBlock mask_blocks = 1

• required bool use_parallel = 2

• optional MLP mlp = 3

• optional bool input_layer_norm = 4

Used in: EvalMetrics

(message has no fields)

Used in: EvalMetrics

(message has no fields)

Used in: EvalMetrics

(message has no fields)

configure backbone network common parameters

Used in: EasyRecModel

• optional float l2_regularization = 1

• repeated string outputs = 2

• repeated BayesTaskTower task_towers = 3

• optional int32 user_tower_idx_in_output = 4

• optional int32 item_tower_idx_in_output = 5

• optional Similarity simi_func = 6

• optional float temperature = 7

• optional bool scale_simi = 8

Configuration message for the MomentumOptimizer See: https://www.tensorflow.org/api_docs/python/tf/train/MomentumOptimizer

Used in: Optimizer

• optional LearningRate learning_rate = 1

• optional float momentum_optimizer_value = 2

Used in: Optimizer

• optional LearningRate learning_rate = 1

• optional float weight_decay = 2

• optional float momentum_optimizer_value = 3

Used in: KerasLayer

• required uint32 num_heads = 1

• required uint32 key_dim = 2

• optional uint32 value_dim = 3

• optional float dropout = 4

• optional bool use_bias = 5

• optional bool return_attention_scores = 6

• optional bool use_causal_mask = 7

• optional uint32 output_shape = 8

  The expected shape of an output tensor, besides the batch and sequence dims. If not specified, projects back to the query feature dim (the query input's last dimension).

• repeated int32 attention_axes = 9

  axes over which the attention is applied.

• optional string kernel_initializer = 10

• optional string bias_initializer = 11

Used in: SequenceCombiner

(message has no fields)

Used in: CoMetricLearningI2I, KD, Loss

• required float alpha = 1

• required float beta = 2

• required float lamb = 3

• required float eps = 4

Used in: EasyRecModel

• repeated Tower towers = 1

• required DNN final_dnn = 4

• required float l2_regularization = 5

• repeated DINTower din_towers = 6

• repeated BSTTower bst_towers = 7

Used in: EasyRecModel

• required RecallTower user_tower = 1

• required RecallTower item_tower = 2

• required float l2_regularization = 3

• required DNN final_dnn = 4

• required bool ignore_in_batch_neg_sam = 10

Used in: ExportConfig

• repeated string input_name = 1

Used in: KerasLayer

• required uint32 embedding_dim = 1

• repeated uint32 carries = 2

• optional float multiplier = 3

• optional string intra_ary_pooling = 4

• optional string inter_ary_pooling = 5

  for now, inter_ary_pooling not support yet

• optional bool output_3d_tensor = 6

• optional bool output_tensor_list = 7

• optional uint32 num_replicas = 8

Weighted Random Sampling ItemID not in Batch

Used in: DatasetConfig

• required string input_path = 1

  sample data path itemid weight attrs

• required uint32 num_sample = 2

  number of negative sample

• repeated string attr_fields = 3

  field names of attrs in train data or eval data

• required string item_id_field = 4

  field name of item_id in train data or eval data

• optional string attr_delimiter = 5

• optional uint32 num_eval_sample = 6

• optional string field_delimiter = 7

  only works on DataScience/Local

Used in: DatasetConfig

• required string input_path = 1

  sample data path itemid weight attrs

• required uint32 num_sample = 2

  number of negative sample

• repeated string attr_fields = 3

  field names of attrs in train data or eval data

• required string item_id_field = 4

  field name of item_id in train data or eval data

• optional string attr_delimiter = 5

• optional uint32 num_eval_sample = 6

• optional string field_delimiter = 7

  only works on DataScience/Local

Weighted Random Sampling ItemID not with Edge

Used in: DatasetConfig

• required string user_input_path = 1

  user data path userid weight

• required string item_input_path = 2

  item data path itemid weight attrs

• required string pos_edge_input_path = 3

  positive edge path userid itemid weight

• required uint32 num_sample = 4

  number of negative sample

• repeated string attr_fields = 5

  field names of attrs in train data or eval data

• required string item_id_field = 6

  field name of item_id in train data or eval data

• required string user_id_field = 7

  field name of user_id in train data or eval data

• optional string attr_delimiter = 8

• optional uint32 num_eval_sample = 9

• optional string field_delimiter = 10

  only works on DataScience/Local

Top level optimizer message.

Used in: TrainConfig

• oneof optimizer

  • RMSPropOptimizer rms_prop_optimizer = 101

  • MomentumOptimizer momentum_optimizer = 102

  • AdamOptimizer adam_optimizer = 103

  • MomentumWOptimizer momentumw_optimizer = 104

  • AdamWOptimizer adamw_optimizer = 105

  • AdamAsyncOptimizer adam_async_optimizer = 106

  • AdagradOptimizer adagrad_optimizer = 107

  • FtrlOptimizer ftrl_optimizer = 108

  • AdamAsyncWOptimizer adam_asyncw_optimizer = 109

  • LazyAdamOptimizer lazy_adam_optimizer = 110

• optional bool use_moving_average = 5

• optional float moving_average_decay = 6

• optional float embedding_learning_rate_multiplier = 7

Used in: KerasLayer

• optional string separator = 1

• optional string default_value = 2

• repeated string methods = 3

• optional string normalize_fn = 4

• repeated float boundaries = 5

• optional int32 embedding_dim = 6

• optional int32 print_first_n = 7

• optional int32 summarize = 8

Used in: EasyRecModel

• required DNN user_dnn = 1

  encode user info

• required DNN item_dnn = 2

  encode target item info

• required DNN u2i_dnn = 3

  encode u2i seq info

• required DNN trigger_dnn = 4

  produce trigger score

• required DNN i2i_dnn = 5

  encode trigger item seqs to target item co-occurance info

• required DNN sim_dnn = 6

  produce sim score

• optional DNN direct_user_dnn = 7

  direct net user_dnn

• optional DNN direct_item_dnn = 8

  direct net item_dnn

• optional Similarity simi_func = 9

  for direct net, similar to DSSM

• optional bool scale_simi = 10

  for direct net

• optional DNN bias_dnn = 11

  bias net dnn

• optional string item_id = 12

• optional float l2_regularization = 13

Used in: EasyRecModel

• repeated ExtractionNetwork extraction_networks = 1

  extraction network

• repeated TaskTower task_towers = 2

  task tower

• optional float l2_regularization = 3

  l2 regularization

Used in: KerasLayer

• required MLP mlp = 1

• required GateNN gate_params = 2

• required string mode = 3

  run mode: eager, lazy

• optional bool full_gate_input = 4

Used in: KD, Loss

• required float gamma = 1

• optional float alpha = 2

• optional float hinge_margin = 3

• optional string session_name = 4

• optional float ohem_ratio = 5

• optional float temperature = 6

Used in: KD, Loss

• required float temperature = 1

• optional string session_name = 2

• optional float margin = 3

• optional float ohem_ratio = 4

• optional bool label_is_logits = 5

• optional bool use_label_margin = 6

• optional bool use_exponent = 7

Used in: KD, Loss

• required float temperature = 1

• optional string session_name = 2

• optional float hinge_margin = 3

• optional float ohem_ratio = 4

• optional bool use_label_margin = 5

Used in: KD, Loss

• required float margin = 1

• optional string session_name = 2

• optional float temperature = 3

Used in: KerasLayer

• required uint32 embedding_dim = 1

• required float sigma = 2

• optional bool add_linear_layer = 3

• optional string linear_activation = 4

• optional bool output_3d_tensor = 5

• optional bool output_tensor_list = 6

Configuration message for a poly decaying learning rate. See https://www.tensorflow.org/api_docs/python/tf/train/polynomial_decay.

Used in: LearningRate

• required float learning_rate_base = 1

• required int64 total_steps = 2

• required float power = 3

• optional float end_learning_rate = 4

Used in: EvalMetrics

(message has no fields)

Configuration message for the RMSPropOptimizer See: https://www.tensorflow.org/api_docs/python/tf/train/RMSPropOptimizer

Used in: Optimizer

• optional LearningRate learning_rate = 1

• optional float momentum_optimizer_value = 2

• optional float decay = 3

• optional float epsilon = 4

Configuration proto for random normal initializer. See https://www.tensorflow.org/api_docs/python/tf/random_normal_initializer

Used in: Initializer

• optional float mean = 1

• optional float stddev = 2

Used in: Block

(message has no fields)

Used in: EvalMetrics

(message has no fields)

Used in: EvalMetrics

• optional uint32 topk = 1

Used in: MultiTowerRecall

• required DNN dnn = 1

Used in: Block, Layer

• required uint32 num_steps = 1

• optional uint32 fixed_input_index = 2

• required KerasLayer keras_layer = 3

Proto with one-of field for regularizers.

• oneof regularizer_oneof

  • L1Regularizer l1_regularizer = 1

  • L2Regularizer l2_regularizer = 2

  • L1L2Regularizer l1_l2_regularizer = 3

Used in: Block, Layer

• required uint32 num_repeat = 1

• optional int32 output_concat_axis = 2

  default output the list of multiple outputs

• required KerasLayer keras_layer = 3

• optional string input_slice = 4

• optional string input_fn = 5

Used in: EasyRecModel

• required DNN share_dnn = 1

• required DNN booster_dnn = 2

• required DNN light_dnn = 3

• optional float l2_regularization = 4

• optional bool feature_based_distillation = 5

• optional Similarity feature_distillation_function = 6

  COSINE = 0; EUCLID = 1;

Used in: EvalMetrics

(message has no fields)

Used in: DSSM_SENet_Tower, FiBiNet, KerasLayer

• required uint32 reduction_ratio = 1

• optional uint32 num_squeeze_group = 2

• optional bool use_skip_connection = 3

• optional bool use_output_layer_norm = 4

Used in: EasyRecModel, FeatureGroupConfig

• optional string group_name = 1

• repeated SeqAttMap seq_att_map = 2

• optional bool tf_summary = 3

• optional DNN seq_dnn = 4

• optional bool allow_key_search = 5

• optional bool need_key_feature = 6

• optional bool allow_key_transform = 7

• optional bool transform_dnn = 8

Used in: SeqAttGroupConfig

• repeated string key = 1

• repeated string hist_seq = 2

• repeated string aux_hist_seq = 3

Used in: KerasLayer

• required float mask_rate = 1

  Percentage length of mask original sequence

• required float crop_rate = 2

  Percentage left of crop original sequence

• required float reorder_rate = 3

  Percentage length of reorder original sequence

Used in: FeatureConfig

• oneof combiner

  • AttentionCombiner attention = 1

  • MultiHeadAttentionCombiner multi_head_attention = 2

  • TextCNN text_cnn = 3

Used in: EvalMetrics

• required string session_id_field = 1

  session id field name

• optional string reduction = 2

  reduction: reduction method for auc of different sessions * "mean": simple mean of different sessions * "mean_by_sample_num": weighted mean with sample num of different sessions * "mean_by_positive_num": weighted mean with positive sample num of different sessions

Used in: DAT, DSSM, DSSM_SENet, MIND, ModelParams, PDN, RocketLaunching

• COSINE = 0

• INNER_PRODUCT = 1

• EUCLID = 2

Used in: EasyRecModel

• repeated TaskTower task_towers = 2

• required float l2_regularization = 3

Used in: DropoutNet, KD, Loss

• required uint32 num_negative_samples = 1

• required float margin = 2

• required float gamma = 3

• required float coefficient_of_support_vector = 4

Used in: ESMM, MMoE, PLE, SimpleMultiTask

• required string tower_name = 1

  task name for the task tower

• optional string label_name = 2

  label for the task, default is label_fields by order

• repeated EvalMetrics metrics_set = 3

  metrics for the task

• optional LossType loss_type = 4

  loss for the task

• optional uint32 num_class = 5

  num_class for multi-class classification loss

• optional DNN dnn = 6

  task specific dnn

• optional float weight = 7

  training loss weights

• optional string task_space_indicator_label = 10

  label name for indicating the sample space for the task tower

• optional float in_task_space_weight = 11

  the loss weight for sample in the task space

• optional float out_task_space_weight = 12

  the loss weight for sample out the task space

• repeated Loss losses = 13

  multiple losses

• required bool use_sample_weight = 14

  whether to use sample weight in this tower

• optional string task_space_indicator_name = 15

  field name for indicating the sample space for this task

• optional string task_space_indicator_value = 16

  field value for indicating the sample space for this task

Used in: KerasLayer, SequenceCombiner

• repeated uint32 filter_sizes = 1

• repeated uint32 num_filters = 2

• required uint32 pad_sequence_length = 3

• optional string activation = 4

• optional MLP mlp = 5

Used in: KerasLayer

• required Transformer transformer = 1

• required string separator = 2

• optional string vocab_file = 3

• optional int32 default_token_id = 4

Used in: DCN, ESMM, MultiTower

• required string input = 1

• required DNN dnn = 2

Message for configuring EasyRecModel training jobs (train.py). Next id: 25

Used in: EasyRecConfig

• repeated Optimizer optimizer_config = 1

  optimizer options

• optional float gradient_clipping_by_norm = 2

  If greater than 0, clips gradients by this value.

• optional uint32 num_steps = 5

  Number of steps to train the models: if 0, will train the model indefinitely.

• optional string fine_tune_checkpoint = 6

  Checkpoint to restore variables from.

• optional string fine_tune_ckpt_var_map = 8

• optional bool sync_replicas = 9

  Whether to synchronize replicas during training. In case so, build a SyncReplicateOptimizer

• optional string sparse_accumulator_type = 901

  only take effect on pai-tf when sync_replicas is set, options are: raw, hash, multi_map, list, parallel in general, multi_map runs faster than other options.

• optional float startup_delay_steps = 10

  Number of training steps between replica startup. This flag must be set to 0 if sync_replicas is set to true.

• optional uint32 save_checkpoints_steps = 141

  Step interval for saving checkpoint

• optional uint32 save_checkpoints_secs = 142

  Seconds interval for saving checkpoint

• optional uint32 keep_checkpoint_max = 143

  Max checkpoints to keep

• optional uint32 save_summary_steps = 16

  Save summaries every this many steps.

• optional uint32 log_step_count_steps = 17

  The frequency global step/sec and the loss will be logged during training.

• optional bool is_profiling = 18

  profiling or not

• optional bool force_restore_shape_compatible = 19

  if variable shape is incompatible, clip or pad variables in checkpoint

• optional DistributionStrategy train_distribute = 20

  DistributionStrategy, available values are 'mirrored' and 'collective' and 'ess' - mirrored: MirroredStrategy, single machine and multiple devices; - collective: CollectiveAllReduceStrategy, multiple machines and multiple devices.

• optional int32 num_gpus_per_worker = 21

  Number of gpus per machine

• optional bool summary_model_vars = 23

  summary model variables or not

• optional string protocol = 25

  distribute training protocol [grpc++ | star_server] grpc++: https://help.aliyun.com/document_detail/173157.html?spm=5176.10695662.1996646101.searchclickresult.3ebf450evuaPT3 star_server: https://help.aliyun.com/document_detail/173154.html?spm=a2c4g.11186623.6.627.39ad7e3342KOX4

• optional int32 inter_op_parallelism_threads = 26

  inter_op_parallelism_threads

• optional int32 intra_op_parallelism_threads = 27

  intra_op_parallelism_threads

• optional bool tensor_fuse = 28

  tensor fusion on PAI-TF

• optional bool write_graph = 29

  write graph into graph.pbtxt and summary or not

• repeated string freeze_gradient = 30

  match variable patterns to freeze

• optional IncrementSaveConfig incr_save_config = 31

  increment save config

• optional bool enable_oss_stop_signal = 32

  enable oss stop signal stop by create OSS_STOP_SIGNAL under model_dir

• optional string dead_line = 33

  stop training after dead_line time, format: 20220508 23:59:59

Used in: KerasLayer, TextEncoder

• required uint32 hidden_size = 1

  Size of the encoder layers and the pooler layer

• required uint32 num_hidden_layers = 2

  Number of hidden layers in the Transformer encoder

• required uint32 num_attention_heads = 3

  Number of attention heads for each attention layer in the Transformer encoder

• required uint32 intermediate_size = 4

  The size of the "intermediate" (i.e. feed-forward) layer in the Transformer encoder

• required string hidden_act = 5

  The non-linear activation function (function or string) in the encoder and pooler.

• required float hidden_dropout_prob = 6

  The dropout probability for all fully connected layers in the embeddings, encoder, and pooler

• required uint32 vocab_size = 7

• required uint32 max_position_embeddings = 8

  The maximum sequence length that this model might ever be used with

• required bool use_position_embeddings = 9

  Whether to add position embeddings for the position of each token in the text sequence

• required bool output_all_token_embeddings = 10

  Whether to output all token embedding, if set to false, then only output the first token embedding

• optional float attention_probs_dropout_prob = 11

  The dropout ratio for the attention probabilities

Used in: LearningRate

• required float learning_rate_base = 1

• required int32 hidden_size = 2

• required int32 warmup_steps = 3

• optional float step_scaling_rate = 4

Configuration proto for truncated normal initializer. See https://www.tensorflow.org/api_docs/python/tf/truncated_normal_initializer

Used in: Initializer

• optional float mean = 1

• optional float stddev = 2

Used in: EasyRecModel

• required UniterTower config = 1

• required DNN final_dnn = 2

Used in: DBMTL, Uniter

• required uint32 hidden_size = 1

  Size of the encoder layers and the pooler layer

• required uint32 num_hidden_layers = 2

  Number of hidden layers in the Transformer encoder

• required uint32 num_attention_heads = 3

  Number of attention heads for each attention layer in the Transformer encoder

• required uint32 intermediate_size = 4

  The size of the "intermediate" (i.e. feed-forward) layer in the Transformer encoder

• required string hidden_act = 5

  The non-linear activation function (function or string) in the encoder and pooler.

  "gelu", "relu", "tanh" and "swish" are supported.

• required float hidden_dropout_prob = 6

  The dropout probability for all fully connected layers in the embeddings, encoder, and pooler

• required float attention_probs_dropout_prob = 7

  The dropout ratio for the attention probabilities

• required uint32 max_position_embeddings = 8

  The maximum sequence length that this model might ever be used with

• required bool use_position_embeddings = 9

  Whether to add position embeddings for the position of each token in the text sequence

• required float initializer_range = 10

  The stddev of the truncated_normal_initializer for initializing all weight matrices

• optional DNN other_feature_dnn = 11

  dnn layers for other features

Used in: EasyRecModel

• optional float regularization_lambda = 1

  regularization coefficient lambda

• optional bool embedding_wise_variational_dropout = 2

  variational_dropout dimension

used in CDN model

Used in: KerasLayer

• optional uint32 weight_index = 1

• optional MLP mlp = 2

Used in: EasyRecModel

• required uint32 wide_output_dim = 1

• required DNN dnn = 2

• optional DNN final_dnn = 3

  if set, the output of dnn and wide part are concatenated and passed to the final_dnn; otherwise, they are summarized

• optional float l2_regularization = 4

Used in: FeatureGroupConfig

• DEEP = 0

• WIDE = 1

• WIDE_AND_DEEP = 2