diff --git a/config.json b/config.json
index aaf36669..d3f5e5ed 100644
--- a/config.json
+++ b/config.json
@@ -16,9 +16,10 @@
   "lr_patience": 5,
   "lr_decay": 0.5,
   "batch_size": 256,
+  "r": 5,
+
   "griffin_lim_iters": 60,
   "power": 1.5,
-  "r": 5,
 
   "num_loader_workers": 32,
 
diff --git a/layers/.attention.py.swp b/layers/.attention.py.swp
deleted file mode 100644
index 8fd631d7..00000000
Binary files a/layers/.attention.py.swp and /dev/null differ
diff --git a/layers/.tacotron.py.swp b/layers/.tacotron.py.swp
deleted file mode 100644
index 631a4a89..00000000
Binary files a/layers/.tacotron.py.swp and /dev/null differ
diff --git a/models/.tacotron.py.swp b/models/.tacotron.py.swp
deleted file mode 100644
index 5c740d21..00000000
Binary files a/models/.tacotron.py.swp and /dev/null differ
diff --git a/models/tacotron.py b/models/tacotron.py
index 2201c8c4..f8e234d4 100644
--- a/models/tacotron.py
+++ b/models/tacotron.py
@@ -15,6 +15,8 @@ class Tacotron(nn.Module):
         self.use_memory_mask = use_memory_mask
         self.embedding = nn.Embedding(len(symbols), embedding_dim,
                                       padding_idx=padding_idx)
+        print(" | > Embedding dim : {}".format(len(symbols)))
+
         # Trying smaller std
         self.embedding.weight.data.normal_(0, 0.3)
         self.encoder = Encoder(embedding_dim)
diff --git a/train.py b/train.py
index 2776c1cf..72d58d19 100644
--- a/train.py
+++ b/train.py
@@ -19,7 +19,7 @@ from tensorboardX import SummaryWriter
 
 from utils.generic_utils import (Progbar, remove_experiment_folder,
                                  create_experiment_folder, save_checkpoint,
-                                 load_config)
+                                 load_config, lr_decay)
 from utils.model import get_param_size
 from datasets.LJSpeech import LJSpeechDataset
 from models.tacotron import Tacotron
@@ -97,10 +97,11 @@ def main(args):
     else:
         criterion = nn.L1Loss()
 
-    n_priority_freq = int(3000 / (c.sample_rate * 0.5) * c.num_freq)
+n_priority_freq = int(3000 / (c.sample_rate * 0.5) * c.num_freq)
 
-    lr_scheduler = ReduceLROnPlateau(optimizer, factor=c.lr_decay,
-                                   patience=c.lr_patience, verbose=True)
+    #lr_scheduler = ReduceLROnPlateau(optimizer, factor=c.lr_decay,
+    #                               patience=c.lr_patience, verbose=True)
+    
     epoch_time = 0
     for epoch in range(c.epochs):
 
@@ -119,14 +120,19 @@ def main(args):
 
             current_step = i + args.restore_step + epoch * len(dataloader) + 1
 
+            # setup lr
+            current_lr = lr_decay(init_lr, current_step)
+            for params_group in optimizer.param_groups:
+                param_group['lr'] = current_lr
+
             optimizer.zero_grad()
 
-            try:
-                mel_input = np.concatenate((np.zeros(
-                    [c.batch_size, 1, c.num_mels], dtype=np.float32),
-                    mel_input[:, 1:, :]), axis=1)
-            except:
-                raise TypeError("not same dimension")
+            #try:
+            #    mel_input = np.concatenate((np.zeros(
+            #        [c.batch_size, 1, c.num_mels], dtype=np.float32),
+            #        mel_input[:, 1:, :]), axis=1)
+            #except:
+            #    raise TypeError("not same dimension")
 
             if use_cuda:
                 text_input_var = Variable(torch.from_numpy(text_input).type(
@@ -204,7 +210,7 @@ def main(args):
                 tb.add_image('Spec/Reconstruction', const_spec, current_step)
                 tb.add_image('Spec/GroundTruth', gt_spec, current_step)
 
-        lr_scheduler.step(loss.data[0])
+        #lr_scheduler.step(loss.data[0])
         tb.add_scalar('Time/EpochTime', epoch_time, epoch)
         epoch_time = 0
 
diff --git a/utils/generic_utils.py b/utils/generic_utils.py
index aa329fb7..06e884ce 100644
--- a/utils/generic_utils.py
+++ b/utils/generic_utils.py
@@ -52,6 +52,13 @@ def save_checkpoint(state, filename='checkpoint.pth.tar'):
     torch.save(state, filename)
 
 
+def lr_decay(init_lr, global_step):
+    warmup_steps = 4000.0
+    step = global_step + 1.
+    lr = init_lr * warmup_steps**0.5 * np.minimum(step * warmup_steps**-1.5,
+                                                  step**-0.5)
+    return lr
+
 class Progbar(object):
     """Displays a progress bar.
     # Arguments