# 0525 Meeting

## model

> https://hackmd.io/1epIPSFwQJmkRXmAD44sSA?view

### input
**Impression feature**

==impression內各item的feature==
1. imp_n1：price 做標準化
2. imp_n1b：price 除以 price 中的 max
3. imp_n1c：price 取 log 做跟 imp_n1 一樣的事情
1. imp_n2 = impf_stars：一個 hotel 只評一次(1-5)
2. imp_n3 = impf_from_stars：來自多個使用者的評星等(1-4)
3. imp_n4 = impf_rating：來自多個使用者的評分(4種)
4. imp_n5 = impf_n_prop (properties 長度)
5. imp_n6 = impf_is_from_stars_nan：有無評分
6. imp_n7 = impf_is_rating_nan
7. imp_n8 = impf_is_stars_nan
8. imp_n10 (16 種) (出現過的次數)
1. impression (user/session)
2. target ( 6 kinds of action + all )* (user/session)
9. imp_n11 (imp_n10 的 2./1.)
1. CTR user_id
2. CTR session_id
10. imp_n12 = time_since_last_any_action 標準化
11. imp_n14 = dwell_time

**6個block的output shape都是(?, 25, 128)**

Attention


Self-attention

## Reproduce
50% users/ sessions

讀近來大概20G, 進train前會長到55G

| Code | Time |
| -------- | -------- |
| 001 | 213s |
| 011 | 48s |
| 012 | 11477s |
| 013 | 1289s |
| 014 | 4748s |
| 015 | 12448s |
| train (cross validation) | 4x10x22mins |
| validate | 230s |



---
Whole dataset with default code




---