Add files via upload

DialogueRNN/train_MELD.py

@@ -70,8 +70,12 @@ def train_or_eval_model(model, loss_function, dataloader, epoch, optimizer=None,
         # import ipdb;ipdb.set_trace()
         textf, acouf, qmask, umask, label =\
                 [d.cuda() for d in data[:-1]] if cuda else data[:-1]
-#         log_prob, alpha, alpha_f, alpha_b = model(torch.cat((textf,acouf),dim=-1), qmask,umask) # seq_len, batch, n_classes
-        log_prob, alpha, alpha_f, alpha_b = model(textf, qmask,umask) # seq_len, batch, n_classes
+        if feature_type == "audio":
+            log_prob, alpha, alpha_f, alpha_b = model(acouf, qmask,umask) # seq_len, batch, n_classes
+        elif feature_type == "text":
+            log_prob, alpha, alpha_f, alpha_b = model(textf, qmask,umask) # seq_len, batch, n_classes
+        else:
+            log_prob, alpha, alpha_f, alpha_b = model(torch.cat((textf,acouf),dim=-1), qmask,umask) # seq_len, batch, n_classes
         lp_ = log_prob.transpose(0,1).contiguous().view(-1,log_prob.size()[2]) # batch*seq_len, n_classes
         labels_ = label.view(-1) # batch*seq_len
         loss = loss_function(lp_, labels_, umask)
@@ -119,7 +123,8 @@ def train_or_eval_model(model, loss_function, dataloader, epoch, optimizer=None,
 writer = SummaryWriter()
 
 # choose between 'sentiment' or 'emotion'
-classification_type = 'sentiment'
+classification_type = 'emotion'
+feature_type = 'multimodal'
 
 data_path = 'DialogueRNN_features/MELD_features/'
 batch_size = 30
@@ -133,7 +138,15 @@ def train_or_eval_model(model, loss_function, dataloader, epoch, optimizer=None,
 l2 = 0.00001
 lr = 0.0005
 
-D_m = 600
+if feature_type == 'text':
+    print("Running on the text features........")
+    D_m = 600
+elif feature_type == 'audio':
+    print("Running on the audio features........")
+    D_m = 300
+else:
+    print("Running on the multimodal features........")
+    D_m = 900
 D_g = 150
 D_p = 150
 D_e = 100