training pipelines

states_training.py

@@ -0,0 +1,118 @@
+import torch
+import numpy as np
+import csv
+import sys
+
+from src.dataset import PandemicDataset, Norms
+from src.problem import ReducedSIRProblem
+from src.dinn import DINN, Scheduler, Activation
+
+ALPHA = [1/14, 1/5]
+DO_STATES = True
+DO_SYNTHETIC = False
+
+ITERATIONS = 13
+
+state_starting_index = 0
+
+if "1" in sys.argv:
+     state_starting_index = 8
+
+STATE_LOOKUP = {'Schleswig_Holstein' : 2897000,
+                'Hamburg' : 1841000, 
+                'Niedersachsen' : 7982000, 
+                'Bremen' : 569352,
+                'Nordrhein_Westfalen' : 17930000,
+                'Hessen' : 6266000,
+                'Rheinland_Pfalz' : 4085000,
+                'Baden_Wuerttemberg' : 11070000,
+                'Bayern' : 13080000,
+                'Saarland' : 990509,
+                'Berlin' : 3645000,
+                'Brandenburg' : 2641000,
+                'Mecklenburg_Vorpommern' : 1610000,
+                'Sachsen' : 4078000,
+                'Sachsen_Anhalt' : 2208000,
+                'Thueringen' : 2143000}
+
+if DO_SYNTHETIC:
+    alpha = 1/3
+    covid_data = np.genfromtxt(f'./datasets/I_data.csv', delimiter=',')
+    for i in range(ITERATIONS):
+        dataset = PandemicDataset('Synthetic I', 
+                                    ['I'], 
+                                    7.6e6, 
+                                    *covid_data, 
+                                    norm_name=Norms.CONSTANT, 
+                                    use_scaled_time=True)
+
+        problem = ReducedSIRProblem(dataset, alpha)
+        dinn = DINN(2, 
+                    dataset, 
+                    [], 
+                    problem, 
+                    None, 
+                    state_variables=['R_t'], 
+                    hidden_size=100, 
+                    hidden_layers=4, 
+                    activation_layer=torch.nn.Tanh(),
+                    activation_output=Activation.POWER)
+
+        dinn.configure_training(1e-3, 
+                                20000, 
+                                scheduler_class=Scheduler.POLYNOMIAL, 
+                                lambda_physics=1e-6,
+                                verbose=True)
+        dinn.train(verbose=True, do_split_training=True)
+
+        dinn.save_training_process(f'synthetic_{i}')
+        #r_t = dinn.get_output(1).detach().cpu().numpy()
+
+        #with open(f'./results/synthetic_{i}.csv', 'w', newline='') as csvfile:
+                #writer = csv.writer(csvfile, delimiter=',')
+                #writer.writerow(r_t)
+
+for iteration in range(ITERATIONS):
+    if iteration <= 2:
+         print('skip first three iteration, as it was already done')
+         continue
+    if DO_STATES:
+        for state_idx in range(state_starting_index, state_starting_index + 8):
+            state = list(STATE_LOOKUP.keys())[state_idx]
+            exclude = ['Schleswig_Holstein', 'Hamburg', 'Niedersachsen']
+            if iteration == 3 and state in exclude:
+                print(f'skip in {state} third iteration, as it was already done')
+                continue
+            for i, alpha in enumerate(ALPHA):
+                print(f'training for {state} ({state_idx}), alpha: {alpha}, iter: {iteration}')
+
+                covid_data = np.genfromtxt(f'./datasets/I_RKI_{state}_1_{int(1/alpha)}.csv', delimiter=',')
+                dataset = PandemicDataset(state, ['I'], STATE_LOOKUP[state], *covid_data, norm_name=Norms.CONSTANT, use_scaled_time=True)
+
+                problem = ReducedSIRProblem(dataset, alpha)
+
+                dinn = DINN(2, 
+                            dataset, 
+                            [], 
+                            problem, 
+                            None, 
+                            state_variables=['R_t'], 
+                            hidden_size=100, 
+                            hidden_layers=4, 
+                            activation_layer=torch.nn.Tanh(),
+                            activation_output=Activation.POWER)
+
+                dinn.configure_training(1e-3, 
+                                        25000, 
+                                        scheduler_class=Scheduler.POLYNOMIAL,
+                                        lambda_obs=1e2, 
+                                        lambda_physics=1e-6, 
+                                        verbose=True)
+                dinn.train(verbose=True, do_split_training=True)
+
+                dinn.save_training_process(f'{state}_{i}_{iteration}')
+
+                r_t = dinn.get_output(1).detach().cpu().numpy()
+                with open(f'./results/{state}_{i}_{iteration}.csv', 'w', newline='') as csvfile:
+                    writer = csv.writer(csvfile, delimiter=',')
+                    writer.writerow(r_t)