bug fixing for eval_latent for multi berno

algorithms/offline/pbrl.py

@@ -217,16 +217,17 @@ def evaluate_latent_model(model, dataset, training_data, num_t, preferred_indice
     with torch.no_grad():
         # training eval
         X_train, mu_train = training_data
-        latent_rewards_train = model(X_train).view(num_t, 2, len_t, -1)
-        latent_r_sum_train = torch.sum(latent_rewards_train, dim=2)
-        latent_p_train = torch.nn.functional.softmax(latent_r_sum_train, dim=1)[:,0]
-        latent_mu_train = torch.bernoulli(latent_p_train).long()
-
-        mu_train_flat = mu_train.view(-1)
-        latent_mu_train_flat = latent_mu_train.view(-1)
-        assert(mu_train_flat.shape == latent_mu_train_flat.shape)
-        train_accuracy = accuracy_score(mu_train_flat.cpu().numpy(), latent_mu_train_flat.cpu().numpy())
-        print(f'Train Accuracy: {train_accuracy:.3f}')
+        if torch.all(mu_train.type(torch.int64) == mu_train):
+            latent_rewards_train = model(X_train).view(num_t, 2, len_t, -1)
+            latent_r_sum_train = torch.sum(latent_rewards_train, dim=2)
+            latent_p_train = torch.nn.functional.softmax(latent_r_sum_train, dim=1)[:,0]
+            latent_mu_train = torch.bernoulli(latent_p_train).long()
+
+            mu_train_flat = mu_train.view(-1)
+            latent_mu_train_flat = latent_mu_train.view(-1)
+            assert(mu_train_flat.shape == latent_mu_train_flat.shape)
+            train_accuracy = accuracy_score(mu_train_flat.cpu().numpy(), latent_mu_train_flat.cpu().numpy())
+            print(f'Train Accuracy: {train_accuracy:.3f}')
 
         # testing eval
         t1s, t2s, ps = generate_pbrl_dataset_no_overlap(dataset, num_t=testing_num_t, len_t=len_t, save=False)