diff --git a/model.py b/model.py
index 1fef0c2..92ba98d 100644
--- a/model.py
+++ b/model.py
@@ -1,10 +1,10 @@
-# import tensorflow as tf
-# ### Only needed for me, not to block the whole GPU, you don't need this stuff
-# from keras.backend.tensorflow_backend import set_session
-# config = tf.ConfigProto()
-# config.gpu_options.per_process_gpu_memory_fraction = 0.3
-# set_session(tf.Session(config=config))
-# ### ---end of weird stuff
+import tensorflow as tf
+### Only needed for me, not to block the whole GPU, you don't need this stuff
+from keras.backend.tensorflow_backend import set_session
+config = tf.ConfigProto()
+config.gpu_options.per_process_gpu_memory_fraction = 0.2
+set_session(tf.Session(config=config))
+### ---end of weird stuff
 
 
 import sys
@@ -16,8 +16,8 @@
 from keras.layers import Bidirectional, Concatenate,Flatten,Reshape
 from keras.optimizers import SGD, Adam
 from keras.initializers import Constant
-#from keras.layers import CuDNNLSTM as LSTM  #massive speedup on graphics cards
-from keras.layers import LSTM
+from keras.layers import CuDNNLSTM as LSTM  #massive speedup on graphics cards
+#from keras.layers import LSTM
 from keras.callbacks import EarlyStopping, ModelCheckpoint
 from keras.preprocessing.sequence import pad_sequences
 
@@ -26,6 +26,8 @@
 import re
 import h5py
 
+import random
+
 def normname(n):
     return re.sub("[^a-zA-Z0-9]","_",n)
 
@@ -50,7 +52,7 @@ def build_model(self,dicts_filename,word_seq_len):
         inp_pos=Input(shape=(1,)) #one POS
         inp_deprel=Input(shape=(1,)) #one DEPREL
 
-        chars_emb=Embedding(len(self.char_dict),char_emb_dim,mask_zero=True,embeddings_initializer=Constant(value=0.01))(inp_chars)
+        chars_emb=Embedding(len(self.char_dict),char_emb_dim,mask_zero=False,embeddings_initializer=Constant(value=0.01))(inp_chars)
         pos_emb=Flatten()(Embedding(len(self.pos_dict),pos_emb_dim,embeddings_initializer=Constant(value=0.01))(inp_pos))
         drel_emb=Flatten()(Embedding(len(self.deprel_dict),deprel_emb_dim,embeddings_initializer=Constant(value=0.01))(inp_deprel))
 
@@ -71,27 +73,17 @@ def save_model(self,file_name):
             print(model_json,file=f)
         
 
-if __name__=="__main__":
-    data_train=data.vectorize_data(sys.stdin,"dicts_fi.json")
-    import random
-    random.shuffle(data_train)
-    #data_train=data_train[:3000]
-    inputs=[item[0] for item in data_train]
-    outputs=[item[1] for item in data_train]
-
-    inputs=numpy.array(inputs)
-    inputs_chars,inputs_pos,inputs_deprel=pad_sequences(inputs[:,0],padding="post"),inputs[:,1],inputs[:,2]
-
-    outputs=numpy.array(outputs)
-    print("Inp shape",inputs_chars.shape)
-    print("Out shape",outputs.shape)
-    m=Predictor()
-    _,word_seq_len=inputs_chars.shape
-    m.build_model("dicts_fi.json",word_seq_len)
-    m.save_model("first_try")
-    save_cb=ModelCheckpoint(filepath="first_try.weights.h5", monitor='val_loss', verbose=1, save_best_only=True, mode='auto')
-    hist=m.model.fit([inputs_chars,inputs_pos,inputs_deprel],[outputs[:,i] for i in range(outputs.shape[1])],verbose=1,batch_size=200,epochs=15,validation_split=0.1,callbacks=[save_cb])
-    with open("first_try.history.json","w") as f:
-        json.dump((hist.epoch,hist.history),f)
-        
+# from sklearn.metrics import accuracy_score
+# def accuracy(predictions, gold, lengths):
+#     pred_tags = numpy.concatenate([labels[:lengths[i]] for i, labels in enumerate(predictions)]).ravel()
+    
+#     gold_tags = numpy.concatenate([labels[:lengths[i], 0] for i, labels in enumerate(gold)]).ravel()
     
+#     print('Accuracy:', accuracy_score(gold_tags, pred_tags))
+
+# class EvaluateFeats(keras.callbacks.Callback):
+#     def on_epoch_end(self, epoch, logs={}):
+#         pred = numpy.argmax(self.model.predict(validation_vectorized_data_padded), axis=-1)
+#         accuracy(pred, validation_vectorized_labels_padded, validation_lengths) # FIXME: Using global variables here, not good!
+
+            
diff --git a/train.py b/train.py
new file mode 100644
index 0000000..d777120
--- /dev/null
+++ b/train.py
@@ -0,0 +1,42 @@
+import model as model
+import data
+import random
+import numpy
+import json
+from keras.callbacks import EarlyStopping, ModelCheckpoint
+from keras.preprocessing.sequence import pad_sequences
+
+
+if __name__=="__main__":
+    import argparse
+    parser = argparse.ArgumentParser(description='Train')
+    parser.add_argument('--train-file', help='.conllu')
+    parser.add_argument('--devel-file', help='.conllu')
+    args = parser.parse_args()
+
+    data_train=data.vectorize_data(open(args.train_file),"dicts_fi.json")
+    data_devel=data.vectorize_data(open(args.devel_file),"dicts_fi.json")
+
+    random.shuffle(data_train)
+    inputs_train=numpy.array([item[0] for item in data_train])
+    inputs_train_lst=[pad_sequences(inputs_train[:,0],padding="pre"),inputs_train[:,1],inputs_train[:,2]]
+    word_seq_len=inputs_train_lst[0].shape[1]
+    
+    outputs_train=numpy.array([item[1] for item in data_train])
+    outputs_train_lst=[outputs_train[:,i] for i in range(outputs_train.shape[1])]
+
+    inputs_devel=numpy.array([item[0] for item in data_devel])
+    inputs_devel_lst=[pad_sequences(inputs_devel[:,0],padding="pre",maxlen=word_seq_len),inputs_devel[:,1],inputs_devel[:,2]]
+    outputs_devel=numpy.array([item[1] for item in data_devel])
+    outputs_devel_lst=[outputs_devel[:,i] for i in range(outputs_devel.shape[1])]
+
+    m=model.Predictor()
+    m.build_model("dicts_fi.json",word_seq_len)
+
+    model_name="second_try"
+    m.save_model(model_name)
+    save_cb=ModelCheckpoint(filepath=model_name+".weights.h5", monitor='val_loss', verbose=1, save_best_only=True, mode='auto')
+    hist=m.model.fit(x=inputs_train_lst, y=outputs_train_lst, validation_data=(inputs_devel_lst,outputs_devel_lst), verbose=1, batch_size=200, epochs=15, callbacks=[save_cb])
+    with open(model_name+".history.json","w") as f:
+        json.dump((hist.epoch,hist.history),f)
+