Alina Kolesnikova commited on
Commit
94536ec
1 Parent(s): da6f3a9

upd. README with performance test

Browse files
Files changed (1) hide show
  1. README.md +12 -1
README.md CHANGED
@@ -3,7 +3,7 @@ language:
3
  - ru
4
  ---
5
  # distilrubert-tiny-cased-conversational
6
- Conversational DistilRuBERT-tiny \(Russian, cased, 2‑layer, 768‑hidden, 12‑heads, 107M parameters\) was trained on OpenSubtitles\[1\], [Dirty](https://d3.ru/), [Pikabu](https://pikabu.ru/), and a Social Media segment of Taiga corpus\[2\] (as [Conversational RuBERT](https://huggingface.co/DeepPavlov/rubert-base-cased-conversational)). It can be considered as tiny copy of our [Conversational DistilRuBERT-base](https://huggingface.co/DeepPavlov/distilrubert-base-cased-conversational)
7
 
8
  Our DistilRuBERT-tiny was highly inspired by \[3\], \[4\]. Namely, we used
9
  * KL loss (between teacher and student output logits)
@@ -11,6 +11,17 @@ Our DistilRuBERT-tiny was highly inspired by \[3\], \[4\]. Namely, we used
11
  * Cosine embedding loss (between mean of six consecutive hidden states from teacher's encoder and one hidden state of the student)
12
  * MSE loss (between six consecutive attention maps from teacher's encoder and one attention map of the student)
13
 
 
 
 
 
 
 
 
 
 
 
 
14
  \[1\]: P. Lison and J. Tiedemann, 2016, OpenSubtitles2016: Extracting Large Parallel Corpora from Movie and TV Subtitles. In Proceedings of the 10th International Conference on Language Resources and Evaluation \(LREC 2016\)
15
 
16
  \[2\]: Shavrina T., Shapovalova O. \(2017\) TO THE METHODOLOGY OF CORPUS CONSTRUCTION FOR MACHINE LEARNING: «TAIGA» SYNTAX TREE CORPUS AND PARSER. in proc. of “CORPORA2017”, international conference , Saint-Petersbourg, 2017.
 
3
  - ru
4
  ---
5
  # distilrubert-tiny-cased-conversational
6
+ Conversational DistilRuBERT-tiny \(Russian, cased, 2‑layer, 768‑hidden, 12‑heads, 107M parameters\) was trained on OpenSubtitles\[1\], [Dirty](https://d3.ru/), [Pikabu](https://pikabu.ru/), and a Social Media segment of Taiga corpus\[2\] (as [Conversational RuBERT](https://huggingface.co/DeepPavlov/rubert-base-cased-conversational)). It can be considered as tiny copy of [Conversational DistilRuBERT-base](https://huggingface.co/DeepPavlov/distilrubert-base-cased-conversational).
7
 
8
  Our DistilRuBERT-tiny was highly inspired by \[3\], \[4\]. Namely, we used
9
  * KL loss (between teacher and student output logits)
 
11
  * Cosine embedding loss (between mean of six consecutive hidden states from teacher's encoder and one hidden state of the student)
12
  * MSE loss (between six consecutive attention maps from teacher's encoder and one attention map of the student)
13
 
14
+ The model was trained for about 30 hrs. on 8 nVIDIA Tesla P100-SXM2.0 16Gb.
15
+
16
+ To evaluate improvements in the inference speed, we ran teacher and student models on random sequences with seq_len=512, batch_size = 16 (for throughput) and batch_size=1 (for latency).
17
+ All tests were performed on Intel(R) Xeon(R) CPU E5-2698 v4 @ 2.20GHz and nVIDIA Tesla P100-SXM2.0 16Gb.
18
+
19
+ | Model | Size, Mb. | CPU latency, sec.| GPU latency, sec. | CPU throughput, samples/sec. | GPU throughput, samples/sec. |
20
+ |-------------------------------------------------|------------|------------------|-------------------|------------------------------|------------------------------|
21
+ | Teacher (RuBERT-base-cased-conversational) | 679 | 0.655 | 0.031 | 0.3754 | 36.4902 |
22
+ | Student (DistilRuBERT-tiny-cased-conversational)| 409 | 0.1656 | 0.015 | 0.9692 | 71.3553 |
23
+
24
+
25
  \[1\]: P. Lison and J. Tiedemann, 2016, OpenSubtitles2016: Extracting Large Parallel Corpora from Movie and TV Subtitles. In Proceedings of the 10th International Conference on Language Resources and Evaluation \(LREC 2016\)
26
 
27
  \[2\]: Shavrina T., Shapovalova O. \(2017\) TO THE METHODOLOGY OF CORPUS CONSTRUCTION FOR MACHINE LEARNING: «TAIGA» SYNTAX TREE CORPUS AND PARSER. in proc. of “CORPORA2017”, international conference , Saint-Petersbourg, 2017.