Tencent_Ads_Algo_2020_TOP12

: jackhuntcnPandasCute LogicJake

• P40 24G
• 114G
• 300G

run.sh

• gensim
• torch
• transformers
• keras == 2.3.1
• keras_self_attention
• keras_multi_head
• keras_position_wise_feed_forward
• keras_layer_normalization
• lightgbm

.
 data
  keras
  lgb
  torch
 models
 probs
 raw_data
  test
  train_preliminary
  train_semi_final
 run.sh
 src
  blending
   f1_blend_and_submit.py
  keras
   f10_AGE_m11_transformer_lstm_5inputs_train_fold.py
   f11_GENDER_m1_transformer_3inputs_train_fold.py
   f12_GENDER_m2_transformer_lstm_3inputs_train_fold.py
   f13_merge_fold_results.py
   f1_save_data.py
   f2_save_sequence.py
   f3_save_embeddings.py
   f4_AGE_m3_lstm_4inputs_train_fold.py
   f5_AGE_m4_transformer_4inputs_train_fold.py
   f6_AGE_m5_transformer_3inputs_train_fold.py
   f7_AGE_m6_transformer_lstm_2inputs_train_fold.py
   f8_AGE_m8_transformer_lstm_3inputs_2r_train_fold.py
   f9_AGE_m10_transformer_lstm_5inputs_train_fold.py
  lgb
   f1_save_tfidf_countvec.py
   f2_save_target_encoding.py
   f3_save_AGE_tf_idf_stacking_feats.py
   f4_save_GENDER_tf_idf_stacking_feats.py
   f5_run_fold_training.py
  stacking
   f1_merge_stacking_feats.py
   f2_save_embeddings.py
   f3_stacking_DNN_120probs_train_fold.py
   f4_stacking_transformer_2inputs_90probs_train_fold.py
   f5_merge_fold_results.py
  torch
      f1_save_data.py
      f2_save_embedding_w2v.py
      f3_AGE_m7_lstm_3inputs_train_5fold.py
      f4_AGE_m1_lstm_6inputs_train_5fold.py
      f5_AGE_m9_transformer_3inputs_train_5fold.py
      f6_AGE_m2_transformer_6inputs_train_5fold.py
      f7_save_data.py
      f8_AGE_GENDER_m13_transformer_4inputs_train_5fold.py
      lookahead.py
      m13_transformer_4inputs.py
      m1_lstm_6inputs_age.py
      m2_transformer_6inputs_age.py
      m7_lstm_3inputs_age.py
      m9_transformer_3inputs_age.py
      pytorchtools.py
 w2v_models

17 directories, 40 files

• src , torch/keras/lgb
• data
• models
• probs
• raw_data ,
• run.sh
• w2v_models w2v embedding

keras

age gender

id : ( A )

AGE

• LSTM + Attention , 0.512
• transformer , 0.516
• transformer , 0.515
• transformer + LSTM 0.515
• transformer + LSTM 0.515
• transformer + LSTM 0.517
• transformer + LSTM 0.517

GENDER

• transformer , 0.9500
• transformer + LSTM , 0.9501

torch

transformers age gender

huggingface transformers

AGE

• LSTM + Attention , 0.513
• transformer + LSTM , 0.516
• transformer + LSTM , 0.514

AGE & GENDER

• AGE: transformer + LSTM , 0.519
• GENDER: transformer + LSTM , 0.9468,

LGB

LGB TF-IDF COUNTVEC

, AGE 0.48

NN , AGE stacking

stacking

stacking :

• stacking: DNN ( keras keras torch torch , keras , torch keras concat, ), 0.525
• stacking: id transformers concat, , 0.523

blending

AGE

huggingface transformers transformer + LSTM , 0.519, keras torch , 0.93,0.94 ( keras 0.98, torch 0.96), stacking 0.52780

0.50 * DNN_stacking + 0.15 * transformer_stacking + 0.35 * age_m13

GENDER

Gender transformer 0.95048

0.35 * gender_m1 + 0.35 * gender_m2 + 0.30 * gender_m3