训练神经元

stock/dnn_train_dmi.py

@@ -19,7 +19,7 @@ def read_data(path):
     random.shuffle(lines)
     print('读取数据完毕')
 
-    d=int(0.89*len(lines))
+    d=int(0.81*len(lines))
 
     size = len(lines[0])
     train_x=[s[:size - 2] for s in lines[0:d]]
@@ -43,7 +43,7 @@ def resample(path, suffix='test'):
     lines = []
     with open(data_dir + path + '.log') as f:
         i = 0
-        for x in range(84000): # 63万 13万 8.4万
+        for x in range(64000): # 56万 11万 6.4万
             # print(i)
             lines.append(eval(f.readline().strip()))
             i = i + 1
@@ -59,7 +59,7 @@ def resample(path, suffix='test'):
         v = line[1:x*k + 1]
         v = np.array(v)
         v = v.reshape(k, x)
-        v = v[:,4:8]
+        v = v[:,6:10]
         v = v.reshape(1, 4*k)
         # print(v)
         r = estimator.predict(v)
@@ -67,9 +67,9 @@ def resample(path, suffix='test'):
 
 
 def mul_train(name="10_18d"):
-    for x in range(0, 12):
-    # for x in [3,6,10]:
-        score = train(input_dim=440, result_class=5, file_path=data_dir + "kmeans\\stock"+ name + "_train_" + str(x) + ".log",
+    # for x in range(8, 12):
+    for x in [2, 4, 6, 8]:
+        score = train(input_dim=439, result_class=5, file_path=data_dir + "kmeans\\stock"+ name + "_train_" + str(x) + ".log",
               model_name=name + '_dnn_seq_' + str(x) + '.h5')
 
         with open(data_dir + 'stock' + name + '_dmi.log', 'a') as f:
@@ -80,24 +80,24 @@ def train(input_dim=400, result_class=3, file_path="D:\\data\\quantization\\stoc
     train_x,train_y,test_x,test_y=read_data(file_path)
 
     model = Sequential()
-    model.add(Dense(units=120+input_dim, input_dim=input_dim,  activation='relu', kernel_regularizer=regularizers.l1(0.003)))
-    model.add(Dense(units=120+input_dim, activation='relu'))
+    model.add(Dense(units=220+input_dim, input_dim=input_dim,  activation='relu', kernel_regularizer=regularizers.l1(0.003)))
+    model.add(Dense(units=220+input_dim, activation='relu'))
     model.add(Dropout(0.2))
-    model.add(Dense(units=120+input_dim, activation='relu'))
-    model.add(Dense(units=120+input_dim, activation='relu'))
-    model.add(Dense(units=120+input_dim, activation='relu',kernel_regularizer=regularizers.l1(0.002)))
+    model.add(Dense(units=220+input_dim, activation='relu'))
+    model.add(Dense(units=220+input_dim, activation='relu'))
+    model.add(Dense(units=220+input_dim, activation='relu',kernel_regularizer=regularizers.l1(0.002)))
     model.add(Dropout(0.2))
     model.add(Dense(units=120 + input_dim, activation='relu'))
-    model.add(Dropout(0.2))
-    model.add(Dense(units=120+input_dim, activation='selu'))
+    # model.add(Dropout(0.2))
+    # model.add(Dense(units=120+input_dim, activation='selu'))
     # model.add(Dense(units=120+input_dim, activation='selu'))
-    model.add(Dense(units=512, activation='relu'))
+    model.add(Dense(units=1024, activation='relu'))
 
     model.add(Dense(units=result_class, activation='softmax'))
     model.compile(loss='categorical_crossentropy', optimizer="adam",metrics=['accuracy'])
 
     print("Starting training ")
-    model.fit(train_x, train_y, batch_size=4096*4, epochs=600 + 4*int(len(train_x)/1200), shuffle=True)
+    model.fit(train_x, train_y, batch_size=4096*4, epochs=550 + 4*int(len(train_x)/1000), shuffle=True)
     score = model.evaluate(test_x, test_y)
     print(score)
     print('Test score:', score[0])
@@ -115,5 +115,5 @@ def train(input_dim=400, result_class=3, file_path="D:\\data\\quantization\\stoc
 
 
 if __name__ == '__main__':
-    # resample('stock14_18d_test', suffix='test')
-    mul_train('14_18d')
+    resample('stock15_18d_test', suffix='test')
+    # mul_train('15_18d')