处理图像、搭建模型

train_model.py

@@ -1,4 +1,4 @@
-import pandas as pd # data analysis toolkit - create, read, update, delete datasets
+#import pandas as pd # data analysis toolkit - create, read, update, delete datasets
 import numpy as np #matrix math
 from sklearn.model_selection import train_test_split #to split out training and testing data 
 #keras is a high level wrapper on top of tensorflow (machine learning library)
@@ -26,7 +26,7 @@
 
 IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNELS = 120, 160, 3
 INPUT_SHAPE = (IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNELS)
-
+(
 def load_data():
     """
     Load training data and split it into training and validation set

zth_drive.py

@@ -0,0 +1 @@
+# 利用训练好的模型操作小车，实现自动驾驶

zth_process_img.py

@@ -0,0 +1,81 @@
+# 将图片处理为npz格式
+
+import os
+import numpy as np
+import matplotlib.image as mpimg
+from time import time
+import math
+from PIL import Image
+
+CHUNK_SIZE = 256
+
+
+
+
+def process_img(img_path, key):
+
+    print(img_path, key)
+    image = Image.open(img_path)
+    image_array = np.array(image)
+    image_array = np.expand_dims(image_array, axis=0)
+
+
+
+
+    #image_array = mpimg.imread(img_path)
+    #image_array = np.expand_dims(image_array, axis=0)
+
+    print(image_array.shape)
+
+    if key == 2:
+        label_array = [0., 0., 1., 0., 0.]
+    elif key == 3:
+        label_array = [0., 0., 0., 1., 0.]
+    elif key == 0:
+        label_array = [1., 0., 0., 0., 0.]
+    elif key == 1:
+        label_array = [0., 1., 0., 0., 0.]
+    elif key == 4:
+        label_array = [0., 0., 0., 0., 1.]
+
+    return (image_array, label_array)
+    # 返回图片的数据（矩阵），和对应的标签值
+
+
+if __name__ == '__main__':
+    path = "training_data"
+    files = os.listdir(path)                           # 将该路径下的文件名都存入列表
+    turns = int(math.ceil(len(files) / CHUNK_SIZE))
+    print("number of files: {}".format(len(files)))
+    print("turns: {}".format(turns))
+
+    for turn in range(0, turns):
+        train_labels = np.zeros((1, 5), 'float')
+        train_imgs = np.zeros([1, 120, 160, 3])
+
+        CHUNK_files = files[turn * CHUNK_SIZE: (turn + 1) * CHUNK_SIZE]
+        print("number of CHUNK files: {}".format(len(CHUNK_files)))
+        for file in CHUNK_files:
+            if not os.path.isdir(file) and file[len(file) - 3:len(file)] == 'jpg':
+                try:
+                    key = int(file[0])
+                    image_array, label_array = process_img(path + "/" + file, key)
+                    train_imgs = np.vstack((train_imgs, image_array))
+                    train_labels = np.vstack((train_labels, label_array))
+                except:
+                    print('prcess error')
+
+        # 去掉第0位的全零图像数组，全零图像数组是 train_imgs = np.zeros([1,120,160,3]) 初始化生成的
+        train_imgs = train_imgs[1:, :]
+        train_labels = train_labels[1:, :]
+        file_name = str(int(time()))
+        directory = "/Volumes/Seagate Expansion Drive/tianhangz/project/selfdrive/training_data_npz"
+
+        if not os.path.exists(directory):
+            os.makedirs(directory)
+        try:
+            np.savez(directory + '/' + file_name + '.npz', train_imgs=train_imgs, train_labels=train_labels)
+        except IOError as e:
+            print(e)
+
+

zth_train.py

@@ -0,0 +1,184 @@
+# 搭建深度学习模型
+# 导入库
+import keras
+import tensorflow
+import sys
+import os
+import h5py
+import numpy as np
+import glob
+from sklearn.model_selection import train_test_split
+from keras.models import Sequential
+from keras.layers import Lambda, Conv2D, MaxPooling2D, Dropout, Dense, Flatten
+from keras.models import load_model, Model, Input
+from keras.callbacks import ModelCheckpoint, EarlyStopping, TensorBoard
+from keras.optimizers import Adam, SGD
+
+
+# 全局变量
+IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNELS = 120, 160, 3
+INPUT_SHAPE = (IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNELS)
+
+
+# step1,载入数据，并且分割为训练和验证集
+def load_data():
+
+    # load
+    image_array = np.zeros((1, 120, 160, 3))
+    label_array = np.zeros((1, 5), 'float')
+    training_data = glob.glob('training_data_npz/*.npz')   # 匹配所有的符合条件的文件，并将其以list的形式返回。
+
+    # 为空
+    if not training_data:
+        print('文件内没有训练数据了，退出')
+        sys.exit()
+    # 遍历训练文件
+    for single_nzp in training_data:
+        with np.load(single_nzp) as data:
+            print(data.keys())
+            train_temp = data['train_imgs']
+            train_labels_temp = data['train_labels']
+        image_array = np.vstack((image_array, train_temp))
+        label_array = np.vstack((label_array, train_labels_temp))
+    X = image_array[1:, :]
+    y = label_array[1:, :]
+    print('Image array shape: '+ str(X.shape))
+    print('Label array shape: '+ str(y.shape))
+    print(np.mean(X))
+    print(np.var(X))
+
+    # 将数据分割为训练集（80），测试集（20），验证集
+    X_train, X_valid, y_train, y_valid = train_test_split(X, y, test_size=0.2, random_state=0)
+
+    return X_train, X_valid, y_train, y_valid
+
+
+# step2 建立模型
+def build_model(keep_prob):
+    model = Sequential
+    model.add(Lambda(lambda x: (x/102.83 - 1), input_shape = INPUT_SHAPE))
+    model.add(Conv2D(24, (5, 5), activation='elu', strides=(2, 2)))
+    model.add(Conv2D(36, (5, 5), activation='elu', strides=(2, 2)))
+    model.add(Conv2D(48, (5, 5), activation='elu', strides=(2, 2)))
+
+    #model.add(Dropout(0.5))
+    model.add(Conv2D(64, (3, 3),activation='elu'))
+    #model.add(Dropout(0.3))
+    model.add(Conv2D(64, (3, 3),activation='elu'))
+    model.add(Dropout(keep_prob))  # Dropout将在训练过程中每次更新参数时随机断开一定百分比（p）的输入神经元连接
+    model.add(Flatten())
+    model.add(Dense(500, activation='elu'))
+    model.add(Dropout(0.1))
+    model.add(Dense(250, activation='elu'))
+    model.add(Dropout(0.1))
+    model.add(Dense(50, activation='elu'))
+    model.add(Dropout(0.1))
+    model.add(Dense(5))
+    model.summary()
+
+    return model
+
+# step3 训练模型
+def train_model(model, learning_rate, nb_epoch, samples_per_epoch,
+                batch_size, X_train, X_valid, y_train, y_valid):
+    checkpoint = ModelCheckpoint('model-{epoch:03d}.h5', monitor='val_loss',
+                                 verbose=0,
+                                 save_best_only=True,
+                                 mode='min')
+    # EarlyStopping patience：当earlystop被激活（如发现loss相比上一个epoch训练没有下降），
+    # 则经过patience个epoch后停止训练。
+    # mode：‘auto’，‘min’，‘max’之一，在min模式下，如果检测值停止下降则中止训练。在max模式下，当检测值不再上升则停止训练。
+    early_stop = EarlyStopping(monitor='val_loss', min_delta=.0005, patience=4,
+                               verbose=1, mode='min')
+    tensorboard = TensorBoard(log_dir='./logs', histogram_freq=0, batch_size=20, write_graph=True,
+                              write_images=True, embeddings_freq=0, embeddings_layer_names=None,
+                              embeddings_metadata=None)
+    # 编译神经网络模型，loss损失函数，optimizer优化器， metrics列表，包含评估模型在训练和测试时网络性能的指标
+    model.compile(loss='mean_squared_error', optimzier=Adam(lr=learning_rate), metrics=['accuracy'])
+    # 训练神经网络模型，batch_size梯度下降时每个batch包含的样本数，epochs训练多少轮结束，
+    # verbose是否显示日志信息，validation_data用来验证的数据集
+    model.fit_generator(batch_generator(X_train, y_train, batch_size),
+                        steps_per_epoch=samples_per_epoch/batch_size,
+                        max_queue_size=1,
+                        validation_data=batch_generator(X_valid, y_valid, batch_size),
+                        validation_steps=len(X_valid)/batch_size,
+                        callbacks=[tensorboard, checkpoint, early_stop],
+                        verbose=2)
+
+
+
+
+
+# step4
+
+def batch_generator(X, y, batch_size):
+    images = np.empty([batch_size, IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANNELS])
+    steers = np.empty(batch_size, 5)
+    while True:
+        i = 0
+        for index in np.random.permutation(X.shape[0]):
+            images[i] = X[index]
+            steers[i] = y[index]
+            i += 1
+            if i == batch_size:
+                break
+        yield (images, steers)
+
+
+# step5 评估模型
+#def evaluate(x_test, y_test):
+    #score = model.evaluate(x_test, y_test, verbose=0)
+    #print('Test loss:', score[0])
+    #print('Test accuracy:', score[1])
+
+
+
+
+
+# step6
+
+
+
+
+# step7
+
+
+
+
+def main():
+    # 打印出超参数
+
+    print('-'*30)
+    print('parameters')
+    print('-'*30)
+
+
+    keep_prob = 0.5
+    learning_rate = 0.0001
+    nb_epoch = 100
+    samples_per_epoch = 3000
+    batch_size = 40
+
+    print('keep_prob = %f', keep_prob)
+    print('learning_rate = %f', learning_rate)
+    print('nb_epoch = %d', nb_epoch)
+    print('samples_per_epoch = %d', samples_per_epoch)
+    print('batch_size = %d', batch_size)
+    print('-' * 30)
+
+    # 开始载入数据
+    data = load_data()
+    # 编译模型
+    model = build_model(keep_prob)
+    # 在数据集上训练模型，保存成model.h5
+    train_model(model, learning_rate, nb_epoch, samples_per_epoch, batch_size, *data)
+
+
+if __name__ == '__main__':
+    main()
+
+
+
+
+
+