High Score 144 , Need More Data and Better Algorithm

.gitignore

@@ -1,2 +1,4 @@
 # Don't track content of these folders
 screens/  
+1/
+temp/

helper.py

@@ -8,7 +8,11 @@
 from PIL import Image
 
 
-for i in range(0,50):
-    im = Image.open("screenshot-{}.png".format(i)) # Can be many different formats.
-    pix = im.load()
-    print sum(pix[330,291]),
+import pandas
+
+l = [1,2,3,4]
+ll = [5,6,7,8]
+
+
+df = pandas.DataFrame(data={"col1": l, "col2": ll})
+df.to_csv("file.csv", sep=',',index=False)

image_opt.py

@@ -2,17 +2,17 @@
 import numpy as np
 
 from PIL import Image
-
+import keyboard
 import cv2
 
 def mouse_callback(event, x, y, flags, params):
     global thresh_img
     print x , y , thresh_img[y,x]
 
 
-img = cv2.imread('screens/screenshot-213.png',0)
+img = cv2.imread('screens/screenshot-22.png',0)
 
-img = img[ 169 : 400 , 90: ]
+img = img[ 168 : 400 , 105: ]
 
 print img[119,315]
 
@@ -27,9 +27,11 @@ def mouse_callback(event, x, y, flags, params):
 
 cv2.setMouseCallback('image', mouse_callback)
 
+
 
-cv2.imshow("image" , thresh_img)
 
 
+cv2.imshow("image" , thresh_img)
+
 
 cv2.waitKey(0)

main.py

@@ -3,31 +3,60 @@
 
 import numpy as np
 
-from PIL import Image
-
+import keyboard
 import cv2
 
-driver = webdriver.Chrome()
+import pandas
+
+import os
+
+chrome_options = webdriver.ChromeOptions()
+chrome_options.add_argument('--no-sandbox')
+driver = webdriver.Chrome(chrome_options=chrome_options)
+
 driver.get("chrome://dino")
 elem = driver.find_element_by_id("t")
-    
+space = 0
 variable = 0
+
+batch = 1
+
+image = []
+action = []
+
+try:
+    # Create target Directory
+    os.mkdir(str(batch))
+except :
+    print "Folder Exists"
+
+
+
 while True:
-    driver.save_screenshot("screens/screenshot-{}.png".format(variable))
-    img = cv2.imread("screens/screenshot-{}.png".format(variable))
+    if( elem.get_attribute("class") == "offline" ):continue
+    if(driver.execute_script("return Runner.instance_.playing;") == False ):continue
 
-    img = img[ 169 : 400 , 90: ]
+    driver.save_screenshot("{}/screenshot-{}.png".format(batch,variable))
+    img = cv2.imread("{}/screenshot-{}.png".format(batch , variable))
 
-    ret,thresh_img = cv2.threshold(img,127,255,cv2.THRESH_BINARY)
+    img = img[ 168 : 400 , 105: ]
 
+    ret,thresh_img = cv2.threshold(img,127,255,cv2.THRESH_BINARY)
 
+    if(keyboard.is_pressed("space")):space = 1
+    if(keyboard.is_pressed("esc")):break
 
-    print thresh_img[119,291] 
+    print variable
 
-    cv2.imwrite("screens/screenshot-{}-converted.png".format(variable), thresh_img)
+    cv2.imwrite("{}/train-{}.png".format(batch,variable), thresh_img)
+
+    image.append("{}/train-{}.png".format(batch,variable))
+    action.append(space)
+    space = 0
 
-    if sum(thresh_img[119,291]) == 0 :
-        elem.send_keys(Keys.SPACE)
-        print "Sent Key"
     variable+=1
+
+df = pandas.DataFrame(data={"image": image, "action": action})
+df.to_csv("1/out.csv", sep=',',index=False)
+
 driver.close()

model.py

@@ -1,12 +1,118 @@
 
 import tensorflow as tf
+import csv
+from PIL import Image
+import numpy as np
+import PIL.ImageOps    
+import cv2
+from selenium import webdriver
+from selenium.webdriver.common.keys import Keys
 
-# First Layer 
+files = []
+action = []
 
-X = tf.Placeholder( size = (None , 768))
-Y = tf.Placeholder( size = (None ))
+with open("1/out.csv" , "r") as f :
+    data = csv.reader(f, delimiter=',')
+    for i in data:
+        if(i[0] == 1):
+            action.append(0)
+        else:
+            action.append(1)
+        files.append(i[1])
+
 
+def parse_function(filename,basewidth):
+    img = Image.open(filename).convert('L')
+    img = PIL.ImageOps.invert(img)
+    wpercent = (basewidth/float(img.size[0]))
+    hsize = int((float(img.size[1])*float(wpercent)))
+    img = img.resize((basewidth,hsize), Image.ANTIALIAS)    
+    return img
 
-# layer 1
 
-W1 = tf.Variable(type=tf.float32  )
+x_v = [np.array(parse_function(i,200)).flatten()  for i in files ] 
+y_v = action 
+
+print "Loaded all Images"
+
+print "Starting Training on NN ( Tensorflow ) " 
+
+'''
+Using a Neural Network with No Hidden Layers with 55 input layers with 200 values each. giving a total of 55 * 200 Values.
+
+The weight dimentions are 200 * 1 , input is 1 * 200
+
+Output is single neuron ( 1 - Neuron Fired -- Jump ) else ( Leave )
+
+'''
+
+print sum(x_v[0])
+
+# Parameters
+#learning_rate = .00000000001
+learning_rate =  .00000001
+training_epochs = 100
+batch_size = 100
+display_step = 1
+
+tf.set_random_seed(777)  
+x = tf.placeholder(tf.float32, [None, 11000])
+y = tf.placeholder(tf.float32, [None]) 
+
+W = tf.Variable(tf.random_normal([11000, 1]))
+b = tf.Variable(tf.random_normal([1]))
+
+pred = tf.nn.sigmoid(tf.matmul(x, W) + b ) 
+
+
+#cost = tf.reduce_mean(-1 * (tf.reduce_sum(y*tf.log(pred) + tf.reduce_sum( (y-1)*tf.log(pred -1) ) )))
+cost = tf.reduce_mean(-1 * (tf.reduce_sum(y*tf.log(pred  + 1e-30 ))  ) ) 
+
+optimizer = tf.train.GradientDescentOptimizer(learning_rate).minimize(cost)
+
+
+correct_prediction = tf.equal(tf.cast(tf.argmax(pred, 1),"float64"),tf.cast(y , "float64") )
+accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float32"))
+
+init = tf.global_variables_initializer()
+
+with tf.Session() as sess:
+    sess.run(init)
+
+    for epoch in range(training_epochs):
+
+        _, c , w = sess.run([optimizer, cost, W], feed_dict={x: x_v,y: y_v})
+        print sum(w)
+        if (epoch+1) % display_step == 0:
+            print("Epoch:", '%04d' % (epoch+1), "cost=", "{}".format(c))
+
+    print(sess.run([accuracy , pred] ,feed_dict={x: x_v,y: y_v }))
+
+
+    # Start Selenium Program
+
+    import math
+
+    driver = webdriver.Chrome()
+
+    driver.get("chrome://dino")
+    elem = driver.find_element_by_id("t")
+
+    variable = 1
+    while True:
+        if( elem.get_attribute("class") == "offline" ):continue
+        if(driver.execute_script("return Runner.instance_.playing;") == False ):continue
+
+        driver.save_screenshot("temp/screenshot-{}.png".format(variable))
+        img = cv2.imread("temp/screenshot-{}.png".format( variable))
+
+        img = img[ 168 : 400 , 105: ]
+
+        ret,thresh_img = cv2.threshold(img,127,255,cv2.THRESH_BINARY)
+        cv2.imwrite("temp/train-1.png", thresh_img)
+
+        value = parse_function("temp/train-1.png",200)
+
+        if(round(sess.run(pred , feed_dict={x : [np.array(value).flatten()]})) == 0) :
+            elem.send_keys(Keys.SPACE)
+