diff --git a/model_inplement/code/__pycache__/model.cpython-36.pyc b/model_inplement/code/__pycache__/model.cpython-36.pyc
new file mode 100644
index 00000000..2e18ae7f
Binary files /dev/null and b/model_inplement/code/__pycache__/model.cpython-36.pyc differ
diff --git a/model_inplement/code/model.py b/model_inplement/code/model.py
index f73cabe3..32ebdbf9 100644
--- a/model_inplement/code/model.py
+++ b/model_inplement/code/model.py
@@ -22,19 +22,16 @@ class HAN(nn.Module):
         self.output_layer = nn.Linear(2* sent_hidden_size, output_size)
         self.softmax = nn.Softmax()
 
-    def forward(self, x, level='w'):
+    def forward(self, doc):
         # input is a sequence of vector
         # if level == w, a seq of words (a sent); level == s, a seq of sents (a doc)
-        if level == 's':
-            v = self.sent_layer(x)
-            output = self.softmax(self.output_layer(v))
-            return output
-        elif level == 'w':
-            s = self.word_layer(x)
-            return s
-        else:
-            print('unknow level in Parameter!')
-
+        s_list = []
+        for sent in doc:
+            s_list.append(self.word_layer(sent))
+        s_vec = torch.cat(s_list, dim=1).t()
+        doc_vec = self.sent_layer(s_vec)
+        output = self.softmax(self.output_layer(doc_vec))
+        return output
 
 class AttentionNet(nn.Module):
     def __init__(self, input_size, gru_hidden_size, gru_num_layers, context_vec_size):
@@ -60,11 +57,53 @@ class AttentionNet(nn.Module):
         self.context_vec.data.uniform_(-0.1, 0.1)
 
     def forward(self, inputs):
-        # inputs's dim seq_len*word_dim
+        # inputs's dim (seq_len, word_dim)
         inputs = torch.unsqueeze(inputs, 1)
         h_t, hidden = self.gru(inputs)
         h_t = torch.squeeze(h_t, 1)
         u = self.tanh(self.fc(h_t))
         alpha = self.softmax(torch.mm(u, self.context_vec))
         output = torch.mm(h_t.t(), alpha)
+        # output's dim (2*hidden_size, 1)
         return output
+
+
+'''
+Train process
+'''
+import math
+import os
+import copy
+import pickle
+
+import matplotlib.pyplot as plt
+import matplotlib.ticker as ticker
+import numpy as np
+import json
+import nltk
+
+optimizer = torch.optim.SGD(lr=0.01)
+criterion = nn.NLLLoss()
+epoch = 1
+batch_size = 10
+
+net = HAN(input_size=100, output_size=5, 
+        word_hidden_size=50, word_num_layers=1, word_context_size=100,
+        sent_hidden_size=50, sent_num_layers=1, sent_context_size=100)
+
+def dataloader(filename):
+    samples = pickle.load(open(filename, 'rb'))
+    return samples
+
+def gen_doc(text):
+    pass
+
+class SampleDoc:
+    def __init__(self, doc, label):
+        self.doc = doc
+        self.label = label
+
+    def __iter__(self):
+        for sent in self.doc:
+            for word in sent:
+
diff --git a/model_inplement/code/preprocess.py b/model_inplement/code/preprocess.py
new file mode 100644
index 00000000..37f6eb25
--- /dev/null
+++ b/model_inplement/code/preprocess.py
@@ -0,0 +1,42 @@
+import pickle
+import json
+import nltk
+from nltk.tokenize import stanford
+
+# f = open('dataset/review.json', encoding='utf-8')
+# samples = []
+# j = 0
+# for i, line in enumerate(f.readlines()):
+#     review = json.loads(line)
+#     samples.append((review['stars'], review['text']))
+#     if (i+1) % 5000 == 0:
+#         print(i)
+#         pickle.dump(samples, open('review/samples%d.pkl'%j, 'wb'))
+#         j += 1
+#         samples = []
+# pickle.dump(samples, open('review/samples%d.pkl'%j, 'wb'))
+samples = pickle.load(open('review/samples0.pkl', 'rb'))
+# print(samples[0])
+
+import os
+os.environ['JAVAHOME'] = 'D:\\java\\bin\\java.exe'
+path_to_jar = 'E:\\College\\fudanNLP\\stanford-corenlp-full-2018-02-27\\stanford-corenlp-3.9.1.jar'
+tokenizer = stanford.CoreNLPTokenizer()
+
+dirname = 'review'
+dirname1 = 'reviews'
+
+for fn in os.listdir(dirname):
+    print(fn)
+    precessed = []
+    for stars, text in pickle.load(open(os.path.join(dirname, fn), 'rb')):
+        tokens = []
+        sents = nltk.tokenize.sent_tokenize(text)
+        for s in sents:
+            tokens.append(tokenizer.tokenize(s))
+        precessed.append((stars, tokens))
+        # print(tokens)
+        if len(precessed) % 100 == 0:
+            print(len(precessed))
+    pickle.dump(precessed, open(os.path.join(dirname1, fn), 'wb'))
+    
diff --git a/model_inplement/code/train.py b/model_inplement/code/train.py
new file mode 100644
index 00000000..e69de29b