sparse table output

[kiminh]

i have run the wide and deep model on industrial dataset and got the sparse table output. @thundercatli

sparse_opt = tn.core.AdaGrad(learning_rate=0.01, initial_g2sum=0.1, initial_scale=0.1)

opt_name:AdaGrad
dim:8
5550086314013414980     8       0.0506565       0.0165137       -0.000290567    -0.0096354      -0.00974899     0.0368803       -0.017581  -0.0189105       0.1     1.96    0       0
3064317571301576340     8       -0.0575744      -0.0326009      0.0180476       -0.0123906      0.0242864       -0.0342714      0.0521148  0.0441031        0.1     0.98    0       0
1756588472783864528     8       0.0126326       0.0403345       -0.0335226      0.00557136      0.0392844       -0.0347473      -0.0147189 0.00544515       0.1     3.92    0       0
598407693234333660      8       0.0161784       -0.0303006      -0.0684115      0.02354 0.00463372      -0.0147034      0.0116275       -0.0255773  0.1     0.98    0       0
2807026422257546872     8       0.0213118       -0.000866071    0.012316        0.0103005       -0.0127499      0.00723172      -0.0150036 -0.0837113       0.1     1.96    0       0
6118954042852362792     8       0.00998451      -0.0084689      0.0401853       0.00317581      -0.0709747      -0.0203387      -0.000485013 -0.00477354     0.1     1.96    0       0
4096748474065003872     8       0.00899766      -0.0275911      -0.0161629      -0.00443562     -0.0283328      -0.0747014      -0.00938712 -0.0270782       0.1     16.66   0       9.8
4174304074430002404     8       0.0113422       0.0576163       -0.0221065      -0.0159259      -0.00519418     0.0400519       -0.00232559 0.0979451        0.1     1.96    0       0

I'm curious about the meaning of every column, here use_cvm_ is false.

the above has one more column which meaning is not clear, as it's conflict with the example, and i'm not able to figure out its meaning

tensornet/core/ps/table/sparse_table.cc

Line 93 in 8950fe4

op_kernel_->Serialized(file, mode);

https://github.com/Qihoo360/tensornet/blob/master/core/ps/optimizer/ada_grad_kernel.cc#L131

void SparseAdaGradValue::SerializeTxt_(std::ostream& os, int dim) {
    os << dim << "\t";
    for (int i = 0; i < dim; i++) {
        os << Weight()[i] << "\t";
    }

    os << g2sum_ << "\t";
    os << show_ << "\t";
    os << no_show_days_ << "\t";
    if (use_cvm_) {
        os << click_;
    }
}

tensornet/tensornet_tools/python/utils.py

Line 30 in 8950fe4

sparse_ada_grad_schema = ["sign", "dim", "weights", "g2sum", "show", "no_show_days"]

[thundercatli]

tensornet/examples/common/feature_column.py

Line 49 in 8950fe4

columns_group[column_group_name], sparse_opt, name=column_group_name + suffix, target_columns=["label"]

If you use target_columns option, use_cvm_ will be set to true. There will be one more column in sparse output to calculate label count. sparse_ada_grad_schema in tensornet-tools is not update-to-date. Or you don't need to enable target_columns.

[kiminh]

tensornet/examples/common/feature_column.py

Line 49 in 8950fe4

columns_group[column_group_name], sparse_opt, name=column_group_name + suffix, target_columns=["label"]

If you use target_columns option, use_cvm_ will be set to true. There will be one more column in sparse output to calculate label count. sparse_ada_grad_schema in tensornet-tools is not update-to-date. Or you don't need to enable target_columns.

thanks for your reply.
then for single task prediction task, the sparse output will has one more label column. for multi task (ctr prediction and cvr prediction), not support yet.

tensornet/tensornet/layers/embedding_features.py

Line 219 in 8950fe4

if target_columns and len(target_columns) > 1:

        if target_columns and len(target_columns) > 1:
            raise ValueError("For now cvm plugin only support one column, Given: {}".format(target_columns))

[kiminh]

tensornet/examples/common/feature_column.py

Line 49 in 8950fe4

columns_group[column_group_name], sparse_opt, name=column_group_name + suffix, target_columns=["label"]

If you use target_columns option, use_cvm_ will be set to true. There will be one more column in sparse output to calculate label count. sparse_ada_grad_schema in tensornet-tools is not update-to-date. Or you don't need to enable target_columns.

normally, enable use_cvm_ is helpful to improve the model efficiency?