| 2 | |
| 3 | |
| 4 | def build_feature_extractor(input_): |
| 5 | # We only want to create the weights once |
| 6 | # In all future calls we should set reuse = True |
| 7 | |
| 8 | # scale the inputs from 0..255 to 0..1 |
| 9 | input_ = tf.to_float(input_) / 255.0 |
| 10 | |
| 11 | # conv layers |
| 12 | conv1 = tf.contrib.layers.conv2d( |
| 13 | input_, |
| 14 | 16, # num output feature maps |
| 15 | 8, # kernel size |
| 16 | 4, # stride |
| 17 | activation_fn=tf.nn.relu, |
| 18 | scope="conv1") |
| 19 | conv2 = tf.contrib.layers.conv2d( |
| 20 | conv1, |
| 21 | 32, # num output feature maps |
| 22 | 4, # kernel size |
| 23 | 2, # stride |
| 24 | activation_fn=tf.nn.relu, |
| 25 | scope="conv2") |
| 26 | |
| 27 | # image -> feature vector |
| 28 | flat = tf.contrib.layers.flatten(conv2) |
| 29 | |
| 30 | # dense layer |
| 31 | fc1 = tf.contrib.layers.fully_connected( |
| 32 | inputs=flat, |
| 33 | num_outputs=256, |
| 34 | scope="fc1") |
| 35 | |
| 36 | return fc1 |
| 37 | |
| 38 | class PolicyNetwork: |
| 39 | def __init__(self, num_outputs, reg=0.01): |