Time: January 2017

Name:Cuiqing Li

In this project, I will train data sets of photos taken from cameras of the simulation cars using neural network model, and then output the steer angle of the cars! Then, save my model data into a file, and test my model in my simulation car software to see whether my neural network have good performance

you can download your software for training data in this website:https://d17h27t6h515a5.cloudfront.net/topher/2016/(Linux)November/5831f0f7_simulator-linux/simulator-linux.zip, and make sure the IMG folder is in the same directory as self_driving_car_simulation directory.

Data Processing:

As for the image data, I need to read all the image files from log file, and then extract angle values for every pictures. After that I need to get data from IMG file, and resize my IMG file into 32*64 so that the model can be trained in my CPU computer. If not, it is a liitle slow. Originally, I use 160*320 size to train my model, but the data's total size exceeds my PC's memory when training. That's why I changed it into 32*64,but it is not bad at the end. Here is two pictures to show the extracted image data after precessing:

Here is The result of the final review of my simulation

You can see the car can drive by itself based on my neural network model!

Let's take a look at my neural network model:

cropping2d_9 (Cropping2D) (None, 22, 64, 3) 0 cropping2d_input_9[0][0]

---

batchnormalization_9 (BatchNorma (None, 22, 64, 3) 88 cropping2d_9[0][0]

---

convolution2d_50 (Convolution2D) (None, 20, 62, 8) 224 batchnormalization_9[0][0]

---

dropout_40 (Dropout) (None, 20, 62, 8) 0 convolution2d_50[0][0]

---

convolution2d_51 (Convolution2D) (None, 18, 60, 8) 584 dropout_40[0][0]

---

dropout_41 (Dropout) (None, 18, 60, 8) 0 convolution2d_51[0][0]

---

convolution2d_52 (Convolution2D) (None, 16, 58, 16) 1168 dropout_41[0][0]

---

dropout_42 (Dropout) (None, 16, 58, 16) 0 convolution2d_52[0][0]

---

convolution2d_53 (Convolution2D) (None, 14, 56, 24) 3480 dropout_42[0][0]

---

dropout_43 (Dropout) (None, 14, 56, 24) 0 convolution2d_53[0][0]

---

convolution2d_54 (Convolution2D) (None, 12, 54, 36) 7812 dropout_43[0][0]

---

maxpooling2d_2 (MaxPooling2D) (None, 6, 27, 36) 0 convolution2d_54[0][0]

---

convolution2d_55 (Convolution2D) (None, 5, 26, 45) 6525 maxpooling2d_2[0][0]

---

dropout_44 (Dropout) (None, 5, 26, 45) 0 convolution2d_55[0][0]

---

convolution2d_56 (Convolution2D) (None, 4, 25, 45) 8145 dropout_44[0][0]

---

flatten_6 (Flatten) (None, 4500) 0 convolution2d_56[0][0]

---

dense_16 (Dense) (None, 560) 2520560 flatten_6[0][0]

---

dropout_45 (Dropout) (None, 560) 0 dense_16[0][0]

---

activation_11 (Activation) (None, 560) 0 dropout_45[0][0]

---

dense_17 (Dense) (None, 10) 5610 activation_11[0][0]

---

activation_12 (Activation) (None, 10) 0 dense_17[0][0]

---

dense_18 (Dense) (None, 1) 11 activation_12[0][0]

Total params: 2,554,207 Trainable params: 2,554,163 Non-trainable params: 44

You will see that the neural network has input size 32*64, but I cropped the 10 pixels from top for every picture, and it turns out that it is very useful since the the top 10 rows' pixels are not very useful in this case

final loss and val_loss data:

Also, I drawed a picture to show how loss and validation loss change over training my neural model:

This project is from udacity!