Commit
•
d66695c
1
Parent(s):
cebfbf0
Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,22 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
---
|
| 4 |
+
|
| 5 |
+
This is the model for my Course Project at 3 Year of HSE FCS. Python notebooks for my
|
| 6 |
+
research can be found here: https://github.com/mastavtsev/PM_NLP/tree/main.
|
| 7 |
+
|
| 8 |
+
Model is specified for the task of unsupervised anomaly detection using SqueezeBERT architecture.
|
| 9 |
+
The model is trained using masked language modeling using traces of normal program execution. Thus, the model produces a representation
|
| 10 |
+
of trace tokens corresponding to normal program execution.
|
| 11 |
+
|
| 12 |
+
Meta info:
|
| 13 |
+
- Tokenizer LOA 13 - 20000 dictionary size, 300 max. token length
|
| 14 |
+
- 512 context window size
|
| 15 |
+
- 2.5e-3 learning rate
|
| 16 |
+
- LAMB optimizer
|
| 17 |
+
- 300 epochs
|
| 18 |
+
- 43.6 million parameters
|
| 19 |
+
- 1.5 hours in Google Colab with GPU A100
|
| 20 |
+
|
| 21 |
+
Results of the model on the test data:
|
| 22 |
+
|