Protocol Documentation

The number of heads Default: 1

The dimension of heads 

The number of interacting layers Default: 1

 Default: 0.0001

 Default: 0.0001

 Default: true

support gfile.Glob 

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

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

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

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

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

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

mini batch size to use for training and evaluation. Default: 32

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

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

label separator 

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

whether to shuffle data Default: true

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. Default: 32

The number of times a data source is read. If set to zero, the data source
will be reused indefinitely. Default: 0

Number of decoded batches to prefetch. Default: 32

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

shard by file, not by sample, valid only for CSVInput Default: false

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' Default: ,

parallel preproces of raw data, avoid using too small
or too large numbers(suggested be to small than
number of the cores) Default: 8

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, 

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

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

for RTPInput only Default: ;

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

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

 Default: 100

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

input field for sample weight 

the compression type of tfrecord 

n data for one feature in tfrecord 

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. Default: false

 Default: 4096

 Default: STRING

 Default: 1

 Default: 1

user data path
userid weight 

item data path
itemid weight attrs 

hard negative edge path
userid itemid weight 

number of negative sample 

max number of hard negative sample 

field names of attrs in train data or eval data 

field name of item_id in train data or eval data 

field name of user_id in train data or eval data 

 Default: :

 Default: 0

only works on DataScience/Local Default: 

user data path
userid weight 

item data path
itemid weight attrs 

positive edge path
userid itemid weight 

hard negative edge path
userid itemid weight 

number of negative sample 

max number of hard negative sample 

field names of attrs in train data or eval data 

field name of item_id in train data or eval data 

field name of user_id in train data or eval data 

 Default: :

 Default: 0

only works on DataScience/Local Default: 

sample data path
itemid weight attrs 

number of negative sample 

field names of attrs in train data or eval data 

field name of item_id in train data or eval data 

 Default: :

 Default: 0

only works on DataScience/Local Default: 

sample data path
itemid weight attrs 

number of negative sample 

field names of attrs in train data or eval data 

field name of item_id in train data or eval data 

 Default: :

 Default: 0

only works on DataScience/Local Default: 

user data path
userid weight 

item data path
itemid weight attrs 

positive edge path
userid itemid weight 

number of negative sample 

field names of attrs in train data or eval data 

field name of item_id in train data or eval data 

field name of user_id in train data or eval data 

 Default: :

 Default: 0

only works on DataScience/Local Default: 

shared bottom cmbf layer 

shared bottom uniter layer 

shared bottom dnn layer 

mmoe expert dnn layer definition 

number of mmoe experts Default: 0

bayes task tower 

l2 regularization Default: 0.0001

The number of cross layers Default: 3

 Default: 0.0001

 Default: 1

deprecated Default: 0.0001

deprecated Default: 0.0001

 Default: 0.0001

options are: dot and cat Default: dot

whether a feature will interact with itself Default: false

whether to include dense features after interaction Default: false

 Default: 1e-05

hidden units for each layer 

ratio of dropout 

activation function Default: tf.nn.relu

use batch normalization Default: true

 Default: 0

 Default: 0

 Default: 0.5

 Default: 0.0001

 Default: COSINE

add a layer for scaling the similarity Default: true

 Default: false

actually input layers, each layer produce a group of feature 

implemented in easy_rec/python/model/easy_rec_estimator
add regularization to all variables with "embedding_weights:"
in name Default: 0

 Default: CLASSIFICATION

 Default: 1

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

default to be logits Default: true

for CROSS_ENTROPY_LOSS, soft_label must be logits instead of probs 

default to be logits Default: true

currently only support CROSS_ENTROPY_LOSS and L2_LOSS 

 Default: 1

only for loss_type == CROSS_ENTROPY_LOSS Default: 1

 Default: 0.0001

 Default: 200

 Default: 5

Number of examples to process of evaluation. Default: 0

How often to run evaluation. Default: 300

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

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

Type of metrics to use for evaluation.
possible values: 

Evaluation online with batch forward data of training Default: false

uid field name 

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 Default: mean

 Default: 5

session id field name 

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 Default: mean

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 Default: -1

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 Default: final

the metric used to determine the best checkpoint Default: auc

metric value the bigger the best Default: true

enable early stop Default: false

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

custom early stop parameters 

early stop max check steps Default: 10000

each feature has a placeholder Default: true

export to keep, only for exporter_type in [best, latest] Default: 1

multi value field list 

is placeholder named by input Default: false

filter out inputs, only keep effective ones Default: true

export the original feature values as string Default: false

export the outputs required by RTP Default: false

export asset files 

 Default: 0

 Default: 0

input field names: must be included in DatasetConfig.input_fields 

 Default: IdFeature

 Default: 0

 Default: 0

for categorical_column_with_identity Default: 0

only for raw features 

separator with in features Default: |

delimeter to separator key from value 

delimeter to separate sequence multi-values 

truncate sequence data to max_seq_len 

many other field share this config 

lookup max select element number, default 10 Default: 10

max_partitions Default: 1

combiner Default: sum

embedding initializer 

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 Default: -1

normalize raw feature to [0-1] Default: 0

 Default: 0

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

raw feature of multiple dimensions Default: 1

sequence feature combiner 

sub feature type for sequence feature Default: IdFeature

sequence length Default: 1

for expr feature 

embedding variable params 

 Default: DEEP

 Default: false

 Default: false

 Default: false

 Default: true

 Default: false

 Default: 0.0001

hive master's ip 

hive port Default: 10000

hive username Default: admin

hive database Default: default

 Default: 1

 Default: 1

 Default: 1

 Default: 1

 Default: 0

 Default: 1

 Default: 0

 Default: 1

The number of heads of cross modal fusion layer Default: 1

The number of heads of image feature learning layer Default: 1

The number of heads of text feature learning layer Default: 1

The dimension of text heads 

The dimension of image heads Default: 64

The number of patches of image feature, take effect when there is only one image feature Default: 1

Do dimension reduce to this size for image feature before single modal learning module Default: 0

The number of self attention layers for image features Default: 0

The number of self attention layers for text features Default: 1

The number of cross modal layers Default: 1

The dimension of image cross modal heads 

The dimension of text cross modal heads 

Dropout probability for hidden layers Default: 0

Dropout probability of the attention probabilities Default: 0

Whether to add embeddings for different text sequence features Default: false

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

Maximum sequence length that might ever be used with this model Default: 0

Dropout probability for text sequence embeddings Default: 0.1

dnn layers for other features 

Size of the encoder layers and the pooler layer 

Number of hidden layers in the Transformer encoder 

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

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

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

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

The dropout probability for all fully connected layers in the embeddings, encoder, and pooler Default: 0.1

The dropout ratio for the attention probabilities Default: 0.1

The maximum sequence length that this model might ever be used with Default: 512

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

The stddev of the truncated_normal_initializer for initializing all weight matrices Default: 0.02

dnn layers for other features 

 Default: 0.25

 Default: 32

 Default: 1

 Default: 0

 Default: 1

 Default: 2

 Default: 50

 Default: 1

 Default: 0.1

 Default: 0

 Default: 1

 Default: 1

max number of high capsules Default: 5

max behaviour sequence length 

high capsule embedding vector dimension 

number EM iterations Default: 3

routing logits scale Default: 20

routing logits initial stddev Default: 1

squash power Default: 1

output ratio Default: 1

constant interest number
in default, use log(seq_len) Default: false

preprocessing dnn before entering capsule layer 

dnn layers applied on user_context(none sequence features) 

concat user and capsule dnn 

method to combine several user sequences
such as item_ids, category_ids Default: SUM

dnn layers applied on item features 

similarity power, the paper says that the big
the better Default: 10

 Default: COSINE

add a layer for scaling the similarity Default: true

 Default: 0.0001

 Default: false

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. Default: 1

deprecated: original mmoe experts config 

mmoe expert dnn layer definition 

number of mmoe experts Default: 0

task tower 

l2 regularization Default: 0.0001

 Default: 5

 Default: 4

 Default: 0.0001

 Default: 0.0001

 Default: false

 Default: 0.9

 Default: 0.999

 Default: 1e-06

 Default: 0.9

 Default: 0.999

 Default: 0.9

 Default: 0.999

 Default: 1e-06

 Default: 0.9

 Default: 0.999

 Default: 0.002

 Default: 0.002

 Default: 4000000

 Default: 0.0002

 Default: 10000

 Default: 0

 Default: 0.002

 Default: 4000000

 Default: 0.95

 Default: true

 Default: 0

 Default: 0

 Default: 0

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

 Default: -0.5

 Default: 0.1

 Default: 0

 Default: 0

 Default: 0

 Default: 0.002

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

 Default: 0.002

 Default: 0.9

 Default: 1e-06

 Default: 0.9

 Default: false

 Default: 0.9999

 Default: 0

 Default: 0.9

 Default: 0.9

 Default: 1

 Default: 1

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

for compatibility 

recommendation model config 

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. 

number of experts per task 

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

dnn network of experts per task 

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

extraction network 

task tower 

l2 regularization Default: 0.0001

 Default: 0.0001

 Default: false

COSINE = 0; EUCLID = 1; Default: COSINE

 Default: 0.0001

Data Type. 

Shape of the array. 

DT_FLOAT. 

DT_DOUBLE. 

DT_INT32, DT_INT16, DT_INT8, DT_UINT8. 

DT_STRING. 

DT_INT64. 

DT_BOOL. 

A named signature to evaluate. If unspecified, the default signature
will be used 

Input tensors.
Names of input tensor are alias names. The mapping from aliases to real
input tensor names is expected to be stored as named generic signature
under the key "inputs" in the model export.
Each alias listed in a generic signature named "inputs" should be provided
exactly once in order to run the prediction. 

Output filter.
Names specified are alias names. The mapping from aliases to real output
tensor names is expected to be stored as named generic signature under
the key "outputs" in the model export.
Only tensors specified here will be run/fetched and returned, with the
exception that when none is specified, all tensors specified in the
named signature will be run/fetched and returned. 

Output tensors. 

task name for the task tower 

label for the task, default is label_fields by order 

metrics for the task 

loss for the task Default: CLASSIFICATION

num_class for multi-class classification loss Default: 1

task specific dnn 

related tower names 

relation dnn 

training loss weights Default: 1

label name for indcating the sample space for the task tower 

the loss weight for sample in the task space Default: 1

the loss weight for sample out the task space Default: 1

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

task name for the task tower 

label for the task, default is label_fields by order 

metrics for the task 

loss for the task Default: CLASSIFICATION

num_class for multi-class classification loss Default: 1

task specific dnn 

training loss weights Default: 1

label name for indcating the sample space for the task tower 

the loss weight for sample in the task space Default: 1

the loss weight for sample out the task space Default: 1

multiple losses 

 Default: 0

 Default: 0

 Default: 0

 Default: 0

if open, will save increment updates to model_dir/incr_save/ Default: false

 Default: 0

 Default: 600

 Default: incr_save

relative to model_dir Default: true

for online inference, please set the storage.mount_path to mount_path
online service will fail Default: /home/admin/docker_ml/workspace/incr_save/

 Default: 0

 Default: 600

optimizer options 

If greater than 0, clips gradients by this value. Default: 0

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

Checkpoint to restore variables from. 

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

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. Default: multi_map

Number of training steps between replica startup.
This flag must be set to 0 if sync_replicas is set to true. Default: 15

Step interval for saving checkpoint Default: 1000

Seconds interval for saving checkpoint 

Max checkpoints to keep Default: 10

Save summaries every this many steps. Default: 1000

The frequency global step/sec and the loss will be logged during training. Default: 10

profiling or not Default: false

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

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

Number of gpus per machine Default: 1

summary model variables or not Default: false

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 

inter_op_parallelism_threads Default: 0

intra_op_parallelism_threads Default: 0

tensor fusion on PAI-TF Default: false

write graph into graph.pbtxt and summary or not Default: true

match variable patterns to freeze 

increment save config 

enable oss stop signal
stop by create OSS_STOP_SIGNAL under model_dir Default: false

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

regularization coefficient lambda Default: 0.01

variational_dropout dimension Default: false

 Default: 1

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

 Default: 0.0001