-
Notifications
You must be signed in to change notification settings - Fork 1
/
ConfigFileDescription.txt
41 lines (32 loc) · 3.29 KB
/
ConfigFileDescription.txt
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
The configuration JSON file contains the following objects:
A. reaction_network_config: Contains parameters for the reaction network example.
1. reaction_network_name: The name of the reaction network example class (must be defined in ReactionNetworkExamples.py).
2. output_folder: The output folder where results are stored.
3. final_time: This is the time T at which properties of the CME solution need to be estimated. These properties are specified by the output functions defined in the appropriate example class in ReactionNetworkExamples.py.
4. num_species: Number of species in the network.
5. num_time_interval: Specifies the time-discretisation for computation of the loss function.
6. func_names: Names for the function names for the plots.
7. exact_values_computable: "True" if exact computation for the outputs and their sensitivities is possible (the computations are defined in the appropriate example class in ReactionNetworkExamples.py). These exact values are only used for comparison.
B. net_config: Contains parameters describing the feedforward deep neural network (DNN) and the training process.
1. dnn_training_needed: "True" if DNN training is needed.
2. training_samples_needed: "True" if fresh training trajectories need to be generated. Otherwise previously saved trajectories are used.
3. validation_samples_needed: "True" if fresh validation trajectories need to be generated. Otherwise previously saved trajectories are used.
4. activation_function: Specifies the nonlinear activation function for the hidden layer nodes (e.g. "relu")
5. num_exponential_features: Specifies the number of exponential features (i.e.\ eigenvalues) for the temporal dynamics. See the manuscript for more details.
6. num_temporal_dnns: Specifies the number of temporal DNNs. See the manuscript for more details.
7. use_previous_training_weights: "True" if previous DNN training weights are to be used.
8. num_nodes_per_layer: Number of nodes per DNN layer.
9. num_hidden_layers: Number of hidden layers (not including input/output layers).
8. lr_values: The learning rate values for Adam optimiser. Specifies the learning schedule.
9. lr_boundaries: The learning rate boundaries for Adam optimiser. Specifies the learning schedule.
10. num_iterations: Number of iterations of stochastic gradient descent (SGD) for training the DNN.
11. batch_size: Number of training trajectories.
12. valid_size: Number of validation trajectories.
13. logging_frequency: Number of iterations between evaluation of validation loss function.
C. simulation_validation: Contains parameters describing estimation of outputs with Monte Carlo simulations (with the stochastic simulation algorithm (SSA)) and the parameter sensitivities (with the Bernoulli path algorithm (BPA)).
1. simulation_based_validation_needed: "True" if simulation-based estimations are required.
2. num_trajectories: Species the number of trajectories for simulation-based estimates.
3. number_of_auxiliary_paths_for_BPA: Specifies the number of auxiliary paths required by BPA for each sensitivity sample.
D. plotting: Contains parameters for sensitivity bar charts.
1. sensitivity_parameters: Only sensitivities w.r.t. these parameters is plotted.
2. parameter_labels: Axis labels for these sensitivity parameters. Latex encoding is allowed.