Number of actions about vin_tensorflow HOT 3 CLOSED

Hi @WangChen100 thank you for your feedback. The motivation here is that each Q layer corresponds to a 'virtual' action, so the number of layers is not necessarily to be the number of valid actions. Empirically, a bit more layers lead to better performance.

from vin_tensorflow.

Thank you @zuoxingdong very much for your reply. It's a really useful suggestion and trick. And I have another question why activations of convolution in many VIN projects are set None as well as yours.
In my mind, if the activation of tf.layers.conv2d is set None, the two linear convolution layers can be combined into one convolution layer. Waiting for your reply, thank you.

h = tf.layers.conv2d(inputs=X, filters=ch_h, kernel_size=[3, 3], strides=[1, 1], padding='same', activation=None, use_bias=True, kernel_initializer=tf.truncated_normal_initializer(stddev=0.01), bias_initializer=tf.zeros_initializer(), name='h0', reuse=tf.AUTO_REUSE) r = tf.layers.conv2d(inputs=h, filters=1, kernel_size=[3, 3], strides=[1, 1], padding='same', activation=None, use_bias=False, kernel_initializer=tf.truncated_normal_initializer(stddev=0.01), bias_initializer=None, name='r', reuse=None)

from vin_tensorflow.

More convolution layers increase the parameters of networks.

from vin_tensorflow.