From d41cdace7580a0c8ce923e7894420fb842cf04fb Mon Sep 17 00:00:00 2001
From: Alan-Lee123 <liqiangll123@qq.com>
Date: Sat, 17 Jun 2017 08:05:52 +0800
Subject: [PATCH] v3

---
 README.md |  2 +-
 test.py   | 34 ++++++++++++++++++++++++++++++++++
 train.py  | 49 +++++++++++++++++++++++++++----------------------
 3 files changed, 62 insertions(+), 23 deletions(-)
 create mode 100644 test.py

diff --git a/README.md b/README.md
index fc583f4..02099e6 100644
--- a/README.md
+++ b/README.md
@@ -3,5 +3,5 @@ keras implementation of  [A simple neural network module for relational reasonin
 
 Relation network is a noval neural network introduced by deepmind in [A simple neural network module for relational reasoning](https://arxiv.org/pdf/1706.01427.pdf). It can achieve super-human performance in challenging visual question answering datasets such as CLEVR.
 
-I implement Relation network using keras and train it on a challenging visual question answering dataset called [Cornell NLVR](https://github.com/cornell-lic/nlvr). The training is in progress. The temporal test accuracy is 89.10%, which is much higher than the previous state of the art (61.99%).
+I implement Relation network using keras and train it on a challenging visual question answering dataset called [Cornell NLVR](https://github.com/cornell-lic/nlvr).
 
diff --git a/test.py b/test.py
new file mode 100644
index 0000000..338c504
--- /dev/null
+++ b/test.py
@@ -0,0 +1,34 @@
+import prepare
+import keras
+import pickle
+import numpy as np
+
+mxlen = 32
+
+token_file = open('tokenizer', 'rb')
+prepare.tokenizer = pickle.load(token_file)
+
+train_json = 'nlvr\\train\\train.json'
+train_img_folder = 'nlvr\\train\\images'
+test_json = 'nlvr\\test\\test.json'
+test_img_folder = 'nlvr\\test\\images'
+
+data = prepare.load_data(train_json)
+data = prepare.tokenize_data(data, mxlen)
+imgs, ws, labels = prepare.load_images(train_img_folder, data)
+data.clear()
+
+model = keras.models.load_model('model')
+
+
+test_data = prepare.load_data(test_json)
+test_data = prepare.tokenize_data(test_data, mxlen)
+test_imgs, test_ws, test_labels = prepare.load_images(test_img_folder, test_data)
+test_data.clear()
+
+imgs_mean = np.mean(imgs)
+imgs_std = np.std(imgs - imgs_mean)
+
+test_imgs = (test_imgs - imgs_mean) / imgs_std
+
+print(model.evaluate([test_imgs, test_ws], test_labels, batch_size=128))
\ No newline at end of file
diff --git a/train.py b/train.py
index 053ee70..182a2ac 100644
--- a/train.py
+++ b/train.py
@@ -14,8 +14,8 @@ import pickle
 mxlen = 32
 embedding_dim = 50
 lstm_unit = 128
-MLP_unit = 256
-epochs = 100
+MLP_unit = 128
+epochs = 50
 batch_size = 128
 
 train_json = 'nlvr\\train\\train.json'
@@ -51,7 +51,7 @@ def bn_layer(x, conv_unit):
 
 def conv_net(inputs):
     model = bn_layer(32, 3)(inputs)
-    model = MaxPooling2D((2, 2), 2)(model)
+    model = MaxPooling2D((3, 3), 3)(model)
     model = bn_layer(32, 3)(model)
     model = MaxPooling2D((2, 2), 2)(model)
     model = bn_layer(32, 3)(model)
@@ -65,9 +65,9 @@ def conv_net(inputs):
 input1 = Input((50, 200, 3))
 input2 = Input((mxlen,))
 cnn_features = conv_net(input1)
-# embedding_layer = prepare.embedding_layer(prepare.tokenizer.word_index, prepare.get_embeddings_index(), mxlen)
-# embedding = embedding_layer(input2)
-embedding = Embedding(mxlen, embedding_dim)(input2)
+embedding_layer = prepare.embedding_layer(prepare.tokenizer.word_index, prepare.get_embeddings_index(), mxlen)
+embedding = embedding_layer(input2)
+# embedding = Embedding(mxlen, embedding_dim)(input2)
 bi_lstm = Bidirectional(LSTM(lstm_unit, implementation=2, return_sequences=False))
 lstm_encode = bi_lstm(embedding)
 shapes = cnn_features.shape
@@ -111,35 +111,40 @@ for feature1 in features:
         relations.append(concat([feature1, feature2, lstm_encode]))
 
 
-def get_dense(n):
+def stack_layer(layers):
+    def f(x):
+        for k in range(len(layers)):
+            x = layers[k](x)
+        return x
+    return f
+
+
+def get_MLP(n):
     r = []
     for k in range(n):
-        r.append(Dense(MLP_unit, activation='relu'))
-    return r
-
-
-def get_MLP(n, denses):
-    def g(x):
-        d = x
-        for k in range(n):
-            d = denses[k](d)
-        return d
-
-    return g
+        s = stack_layer([
+            Dense(MLP_unit),
+            BatchNormalization(),
+            Activation('relu')
+        ])
+        r.append(s)
+    return stack_layer(r)
 
 
 def bn_dense(x):
     y = Dense(MLP_unit)(x)
     y = BatchNormalization()(y)
     y = Activation('relu')(y)
+    y = Dropout(0.5)(y)
     return y
 
 
-g_MLP = get_MLP(4, get_dense(4))
+g_MLP = get_MLP(3)
 
 mid_relations = []
 for r in relations:
     mid_relations.append(g_MLP(r))
+    print(len(mid_relations))
 combined_relation = Add()(mid_relations)
 
 rn = bn_dense(combined_relation)
@@ -147,7 +152,7 @@ rn = bn_dense(rn)
 pred = Dense(1, activation='sigmoid')(rn)
 
 model = Model(inputs=[input1, input2], outputs=pred)
-optimizer = Adam(lr=3e-5)
+optimizer = Adam(lr=3e-4)
 model.compile(optimizer=optimizer, loss='binary_crossentropy', metrics=['accuracy'])
 for epoch in range(epochs):
     model.fit([imgs, ws], labels, epochs=1, batch_size=batch_size)
@@ -163,4 +168,4 @@ tokenizer_file = open('tokenizer', 'wb')
 pickle.dump(prepare.tokenizer, tokenizer_file)
 tokenizer_file.close()
 gc.collect()
-subprocess.Popen("rundll32.exe powrprof.dll,SetSuspendState 0,1,0")
+# subprocess.Popen("rundll32.exe powrprof.dll,SetSuspendState 0,1,0")