Skip to main content

Table 5 Architectures, parameters, and hyperparameters explored for DNNs

From: Multi-channel PINN: investigating scalable and transferable neural networks for drug discovery

Base model   Value Description
PINN Separated layers 1, 2, 3, 4 The number of separated layers for PINN
Concatenated layers 1, 2 The number of concatenated layers for PINN
Number of nodes 256, 512, 1024, 2048 The number of nodes for layers
Dilated CNN Filters 4, 8, 16, 32 The number of filters for Dilated CNN
Kernel size 6, 8, 12, 22 The length of the convolution window for Dilated CNN
Embedding 16, 32 Dimension of dense embedding for low level representations
LSTM, BLSTM Units 128, 256 The units to represent hidden layers for RNN
DNN Lr 0.0005 Initial learning rate
Initializer \([ - \sqrt{ 3 / fan_{in}}, \sqrt{ 3 / fan_{in}}]\) Initial weight value called Lecun uniform distribution
Optimizer Adam Optimizer for stochastic gradient descent
Weight decay 0.0, 0.00001 Learning rate decay over each update
Activation function ReLU, ELU Neuron activation function
Drop out 0.25, 0.5 The rate of drop out
Batch 1024 Batch size for training
Epochs_training 400 Training epochs on a training task
Epochs_finetune 200 Finetuning epochs for a pretrained model on a test task