Method __init__

tensorflow_agent/net.py:8–48

(self, x, num_targets=6, is_training=True)

class Net(object):
    """AlexNet architecture with modified final fully-connected layers regressed on driving control outputs (steering, throttle, etc...)"""
    def __init__(self, x, num_targets=6, is_training=True):
        self.x = x

        # phase = tf.placeholder(tf.bool, name='phase')  # Used for batch norm

        conv1 = tf.nn.relu(conv2d(x, "conv1", 96, 11, 4, 1))
        lrn1 = lrn(conv1)
        maxpool1 = max_pool_2x2(lrn1)
        conv2 = tf.nn.relu(conv2d(maxpool1, "conv2", 256, 5, 1, 2))
        lrn2 = lrn(conv2)
        maxpool2 = max_pool_2x2(lrn2)
        conv3 = tf.nn.relu(conv2d(maxpool2, "conv3", 384, 3, 1, 1))  # Not sure why this isn't 2 groups, but pretrained net was trained like this so we're going with it.

        # Avoid diverging from pretrained weights with things like batch norm for now.
        # Perhaps try a modern small net like Inception V1, ResNet 18, or Resnet 50
        # conv3 = tf.contrib.layers.batch_norm(conv3, scope='batchnorm3', is_training=phase,
        #     # fused=True,
        #     # data_format='NCHW',
        #     # renorm=True
        # )

        conv4 = tf.nn.relu(conv2d(conv3, "conv4", 384, 3, 1, 2))
        conv5 = tf.nn.relu(conv2d(conv4, "conv5", 256, 3, 1, 2))
        maxpool5 = max_pool_2x2(conv5)
        fc6 = tf.nn.relu(linear(maxpool5, "fc6", 4096))
        if is_training:
            fc6 = tf.nn.dropout(fc6, 0.5)
        else:
            fc6 = tf.nn.dropout(fc6, 1.0)

        fc7 = tf.nn.relu(linear(fc6, "fc7", 4096))
        # fc7 = tf.contrib.layers.batch_norm(fc7, scope='batchnorm7', is_training=phase)
        if is_training:
            fc7 = tf.nn.dropout(fc7, 0.95)
        else:
            fc7 = tf.nn.dropout(fc7, 1.0)

        fc8 = linear(fc7, "fc8", num_targets)
        self.p = fc8
        self.global_step = tf.get_variable("global_step", [], tf.int32, initializer=tf.zeros_initializer,
                                           trainable=False)