Advances, Systems and Applications
From: VTGAN: hybrid generative adversarial networks for cloud workload prediction
Model | Layers | Configuration |
---|---|---|
Stacked LSTM | Bidirectional cuDNNLSTM | 256 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | cuDNNLSTM | 128 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | cuDNNLSTM | 128 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | FC | Dense, output units (1 or p for one or multiple-step-ahead) |
 |  | L1 kernel, and bias regularization = 0.00001 |
Stacked GRU | Bidirectional cuDNNGRU | 256 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | cuDNNGRU | 128 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | cuDNNGRU | 128 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | FC | Dense, output units (1 or p for one or multiple-step-ahead) |
 |  | L1 kernel, and bias regularization = 0.00001 |
VTGAN (LSTM-based) | - Generator parameters | Â |
 | Bidirectional cuDNNLSTM | 256 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | cuDNNLSTM | 128 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | cuDNNLSTM | 128 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | FC | Dense, output units (1 or p for one or multiple-step-ahead) |
 |  | L1 kernel, and bias regularization = 0.00001 |
 | - Discriminator parameters |  |
 | Conv1D | flter=64, kernel size=5, strides=2, padding=same |
 |  | LeakyReLU activation (alpha=0.001) |
 | Conv1D | flter=128, kernel size=5, strides=2, padding=same |
 |  | LeakyReLU activation (alpha=0.001) |
 | Conv1D | flter=128, kernel size=5, strides=2, padding=same |
 |  | LeakyReLU activation (alpha=0.001) |
 | Flatten |  |
 | FC 1 | Dense, units=64, LeakyReLU activation |
 | FC 2 | Dense, output units (1 or p for one or multiple-step-ahead), |
 |  | sigmoid activation |
VTGAN (GRU-based) | - Generator parameters | Â |
 | Bidirectional cuDNNGRU | 256 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | cuDNNGRU | 128 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | cuDNNGRU | 128 units, dropout= 0.2 |
 |  | L1 kernel, recurrent, and bias regularization = 0.00001 |
 | FC | Dense, output units (1 or p for one or multiple-step-ahead) |
 |  | L1 kernel, and bias regularization = 0.00001 |
 | - Discriminator parameters: |  |
 | as VTGAN (LSTM-based) |  |