In [23]:

    

import tensorflow as tf
import numpy as np
import gym


    

In [24]:

    

import matplotlib.pyplot as plt
%matplotlib inline


    

In [25]:

    

sess = tf.InteractiveSession()


    

    




Exception AssertionError: AssertionError() in <bound method InteractiveSession.__del__ of <tensorflow.python.client.session.InteractiveSession object at 0x7f66fc3fe190>> ignored

In [26]:

    

e = gym.make('Breakout-v0')
ob = e.reset()


    

    




[2016-05-19 17:09:21,721] Making new env: Breakout-v0

In [27]:

    

print ob.shape, ob.dtype
plt.imshow(ob)


    

    




(210, 160, 3) uint8






    
Out[27]:





<matplotlib.image.AxesImage at 0x7f66fc382150>






    

In [34]:

    

# reshape is needed so we can use plt.imshow
rgb_to_gray = tf.reshape(tf.image.rgb_to_grayscale(ob), [ob.shape[0], ob.shape[1]])


    

In [35]:

    

gray_ob = rgb_to_gray.eval()
gray_ob.shape, gray_ob.dtype


    

    
Out[35]:





((210, 160), dtype('uint8'))

In [36]:

    

plt.gray()
plt.imshow(gray_ob)


    

    
Out[36]:





<matplotlib.image.AxesImage at 0x7f66e008fc50>






    

In [39]:

    

# let's get the current ratio
from __future__ import division
ratio = ob.shape[0] / ob.shape[1]
print ratio


    

    




1.3125

In [87]:

    

def resize_op(img, h, w):
    resized_ob = tf.image.resize_bilinear(tf.reshape(ob, [1, ob.shape[0], ob.shape[1], ob.shape[2]]), [h,w ])
    return tf.reshape(resized_ob, [h, w, 3])


    

In [88]:

    

# 84 x 84 is the size of the img used in the original DQN paper, 21168 or 7056 (grayscale) points per img
plt.imshow(resize_op(ob, 84, 84).eval())


    

    
Out[88]:





<matplotlib.image.AxesImage at 0x7f66d00740d0>






    

In [111]:

    

w = 40
h = int(w * ratio)
rgb_pixels = 3 * h * w
grayscale_pixels = h * w
print 'height =', h, 'width =', w
print 'pixels with rgb =', rgb_pixels, 'grayscale =', grayscale_pixels
plt.imshow(resize_op(ob, h, w).eval())


    

    




height = 52 width = 40
pixels with rgb = 6240 grayscale = 2080






    
Out[111]:





<matplotlib.image.AxesImage at 0x7f66aca028d0>






    

In [92]:

    

scaled_ob = tf.image.convert_image_dtype(ob, tf.float32, saturate=True).eval()
print scaled_ob.min(), scaled_ob.max()


    

    




0.0 0.784314

In [75]:

    

np.unique(ob)


    

    
Out[75]:





array([  0,  42,  48,  58,  66,  72, 108, 122, 130, 142, 158, 160, 162,
       180, 198, 200], dtype=uint8)

In [77]:

    

np.unique(gray_ob)


    

    
Out[77]:





array([  0,  84, 110, 123, 127, 129, 131, 142, 148], dtype=uint8)

In [94]:

    

np.unique(scaled_ob)


    

    
Out[94]:





array([ 0.        ,  0.16470589,  0.18823531,  0.227451  ,  0.25882354,
        0.28235295,  0.42352945,  0.4784314 ,  0.50980395,  0.55686277,
        0.61960787,  0.627451  ,  0.63529414,  0.70588237,  0.77647066,
        0.7843138 ], dtype=float32)

In [97]:

    

# assumes img.shape is (batch_size, h, w, c)
def img_preprocess(img, h, w):
    img = tf.convert_to_tensor(img)
    rgb2y = tf.image.rgb_to_grayscale(img)
    resized = tf.image.resize_bilinear(rgb2y, [h, w])
    return resized


    

In [105]:

    

obs = np.reshape(ob, [1] + list(ob.shape))
obs.shape


    

    
Out[105]:





(1, 210, 160, 3)

In [112]:

    

preprocessed_img = img_preprocess(obs, 84, 84).eval()


    

In [113]:

    

plt.gray()
plt.imshow(preprocessed_img.reshape(84, 84))


    

    
Out[113]:





<matplotlib.image.AxesImage at 0x7f66ac95ae10>






    

In [ ]: