note_chi_giang.md

Data link:

https://bitbucket.org/nguyenthieu2102/cro_mlnn_jcss/src/bio_inspired/data/

• column: num_of_conn

Conclude

LSTM - GPU: 5.5h GRU - GPU: 2h BiLSTM - CPU : 31h

rows: 30 000 LSTM - CPU: 3h

LSTM, GRU => SMAPE: 3-4% MLP ==> SMAPE: 8%

Giang's test on CPU

1. FL-GANN

• SMAPE: 5% => 8%
• Time: 4000s => 7000s (1h < x < 2h)
• The differents are
  • % split dataset (train, valid, test) = (33%, 33%, 33%)
  • Epoch: 700, Pop size: 250
  • Normalize the whole data before split it (This is a problem need to think about it)

2. LSTM

• SMAPE:
• Time:
• 1 Hidden Layer
• Depend on programing languagues and server architecture (even setup code tensorflow to run on 2 CPU, but its still use 8 CPU and not all 100% CPU on server) ==> Effect Time Consuming