Fixed the input matrix values for the Graph Neural Network. So now we…

… can train the GNN model using the input matrix.

modules/gnn_model.py

@@ -1,6 +1,4 @@
-import dgl
 import dgl.function as fn
-import torch as th
 import torch.nn as nn
 import torch.nn.functional as F
 
@@ -47,10 +45,10 @@ class Net(nn.Module):
   def __init__(self):
     super(Net, self).__init__()
 
-    self.layer1 = GCNLSTM(1, 10)
+    self.layer1 = GCNLSTM(2, 10)
     self.dropout1 = nn.Dropout(0.3)
     self.layer2 = GCNLinear(100, 50)
-    self.layer3 = GCNLinear(50, 1)
+    self.layer3 = GCNLinear(50, 2)
 
   def forward(self, g, features):
     batch_size, seq_len, n_feats = features.size()

scripts/create_grc_graph.py

@@ -42,6 +42,7 @@
 
 # normalize values to [0,1]
 dfc = dfc / dfc.values.max(axis=1, keepdims=True)
+dfc.fillna(0.0, inplace=True) # some GRCs have 0 cases
 
 
 # make C a directed graph
@@ -76,10 +77,16 @@
     confirmed = np.reshape(dfc[lag_dates].values, (-1, n_lags, 1))
 
     #deaths = np.reshape(dfd[lag_dates].values, (-1, n_lags, 1))
-    #features = np.concatenate((confirmed, deaths), axis=2)
 
-    features = confirmed
-    targets = np.stack((dfc[date].values)).T
+    deaths = np.zeros(shape=(confirmed.shape[0], n_lags, 1))
+
+    features = np.concatenate((confirmed, deaths), axis=2)
+
+    #features = confirmed
+
+    #targets = np.stack((dfc[date].values)).T
+
+    targets = np.stack( (dfc[date].values, np.zeros(shape=confirmed.shape[0])) ).T
 
     # save feature and response arrays with original graph
     with open(f'../graph files/{str(date)[0:10]}.pkl', 'wb') as f:

scripts/train_model.py

@@ -4,7 +4,8 @@
 import pickle
 import os
 
-
+import torch as th
+import dgl
 
 
 from modules.gnn_model import *
@@ -131,4 +132,4 @@ def several_day_forecast(g, features, n_days):
 plt.show()
 
 
-sqrt(test_loss[499])
+#sqrt(test_loss[499])