DeepTesla

數據來自deeptesla

安裝要求

• Python3
• Numpy
• Pandas
• TensorFlow
• Keras
• opencv
• imageio

數據預處理

• 將視頻轉化為圖片

import cv2nimport imageionimport pandas as pdnimport numpy as npnepoch_ids = [i+1 for i in range(10)]nimg_data=[[] for i in range(10)]nlabel_data=[[] for i in range(10)]nnfor epoch_id in epoch_ids:n print(---------- processing video {} ----------.format(epoch_id))n csv_path = epochs/epoch{:0>2}_steering.csv.format(epoch_id)n df = pd.read_csv(csv_path)n label_data[epoch_id-1] = df[wheel].valuesn mkv_path = epochs/epoch{:0>2}_front.mkv.format(epoch_id)n vid = imageio.get_reader(mkv_path, ffmpeg)n for i in range(len(df)):n img = cv2.resize(vid.get_data(i),(224,224))n img_data[epoch_id-1].append(img)n print(end of video {},.format(epoch_id),n img_count:,len(img_data[epoch_id-1]),img_shape:,img_data[epoch_id-1][0].shape,labels:,label_data[epoch_id-1].shape)nnimg_data = np.concatenate(img_data, axis=0)nlabel_data = np.concatenate(label_data, axis=0)nprint(-------------end of all---------------,nimg_data:,img_data.shape,label_data:,label_data.shape)n

• 由於我們此次會用到遷移學習，使用的預訓練模型是VGG16，所以我們的圖片形狀為(224, 224, 3)

• 數據混洗、分割

##shufflenp = np.random.permutation(len(img_data))nimg_data,label_data = img_data[p],label_data[p]n##splitnindex = int(len(img_data)*0.8)ntrain_image,test_image = img_data[:index], img_data[index:]ntrain_label,test_label = label_data[:index], label_data[index:]nprint(train_image.shape,test_image.shape)nprint(train_label.shape,test_label.shape)n

模型構建與訓練

from keras.models import Modelnfrom keras.optimizers import SGD,RMSpropnfrom keras.callbacks import TensorBoardnfrom keras.layers import Input, Flatten, Dropout,Dense,Reshape,Convolution2D, MaxPooling2D,GlobalAveragePooling2Dnfrom keras.applications.vgg16 import VGG16nbase_model = VGG16(weights=imagenet,include_top=False)nnx = GlobalAveragePooling2D()(base_model.output)nx = Dense(1024, activation=relu)(x)nx = Dense(512, activation=relu)(x)nx = Dropout(0.7)(x)noutput = Dense(1)(x)nmodel = Model(input=base_model.input,output=output)nnfor layer in base_model.layers:n layer.trainable = Falsennopt = RMSprop(lr=1e-6)nmodel.compile(optimizer=Adadelta,n loss=mse)nnmodel.fit(train_image,train_label,validation_split=0.2,batch_size=32, callbacks=[TensorBoard(log_dir=./log)])nmodel.evaluate(test_image, test_label)nnmodel.save(./models/model.h5)nwith open(./models/model.json, w) as f:n f.write(model.to_json())n

評估與預測

model.evaluate(test_image, test_label)n

模型的loss為0.7065757193388762

• 預測可視化

pred = model.predict(train_image[0:100])nplt.plot(pred,r)nplt.plot(train_label[0:100],b)nplt.show()n

藍色為真實值，紅色為預測值，相比上次的行為克隆真是太好了

模擬

黑色方向盤為人類駕駛，綠色方向盤為模型預測，右下角的曲線圖展示了兩者方向盤角度的區別。上圖只是告訴你模擬方式，並不是具體效果。

結尾

代碼/數據地址ciozhang

推薦閱讀：