NavodPeiris/GraphSense-Research

★ 0Forks 0PythonGitHub ↗Compare

README

Training Architecture

architecture

Inference Architecture

architecture

gpt2-medium model (Fine-tuned on Python Algorithms dataset)

artifacts size: 1.44 GB   
avg inference time (CPU): 8 seconds 
avg inference time (GPU): 2.2662 seconds
avg memory usage: 1800 MB 

GraphSense (trained on Python Algorithms dataset)

artifacts size: 13.9 MB
avg inference time (CPU): 0.0079 seconds 
avg memory usage: 277.8194 MB 

accuracy of GraphSense (vector size: 128)

Epochs Walks Length Window Top-1 Accuracy Top-3 Accuracy Top-10 Accuracy
1 10 10 5 0.3694 0.6660 0.7575
10 10 10 5 0.4182 0.7706 0.8894
100 10 10 5 0.4718 0.8012 0.8958
100 100 10 5 0.4371 0.7299 0.8279
1000 10 10 5 0.4847 0.7960 0.8926
100 20 20 3 0.4334 0.7288 0.8482
100 20 20 10 0.4521 0.7607 0.8717

scalability of GraphSense (CPU) (vector size: 128)

vocabulary = 100,000
average memory usage: 273.777 MB
average execution time: 0.0113 seconds
artifacts size: 61.3 MB

vocabulary = 200,000
average memory usage: 325.8949 MB
average execution time: 0.0155 seconds
artifacts size: 122 MB

vocabulary = 300,000
average memory usage: 377.1085 MB
average execution time: 0.0185 seconds
artifacts size: 168 MB

vocabulary = 400,000
average memory usage: 428.3011 MB
average execution time: 0.0227 seconds
artifacts size: 224 MB

vocabulary = 500,000
average memory usage: 478.8532 MB
average execution time: 0.0273 seconds
artifacts size: 280 MB

vocabulary = 600,000
average memory usage: 531.0189 MB
average execution time: 0.0301 seconds
artifacts size: 368 MB

vocabulary = 700,000
average memory usage: 581.3494 MB
average execution time: 0.0333 seconds
artifacts size: 429 MB

vocabulary = 800,000
average memory usage: 633.226 MB
average execution time: 0.038 seconds
artifacts size: 448 MB

vocabulary = 900,000
average memory usage: 685.1932 MB
average execution time: 0.0439 seconds
artifacts size: 552 MB

vocabulary = 1,000,000
average memory usage: 734.5819 MB
average execution time: 0.0444 seconds
artifacts size: 561 MB

Linear Scaling

scaling
scaling
scaling

Contributors

NavodPeiris

Issues