def encoder_net(self, img, latent_dim): x = img x = tf.nn.relu(U.conv2d(x, 32, "c1", [4, 4], [2, 2], pad = "SAME")) # [32, 32, 32] x = tf.nn.relu(U.conv2d(x, 32, "c2", [4, 4], [2, 2], pad = "SAME")) # [16, 16, 32] x = tf.nn.relu(U.conv2d(x, 64, "c3", [4, 4], [2, 2], pad = "SAME")) # [8, 8, 64] x = tf.nn.relu(U.conv2d(x, 64, "c4", [4, 4], [2, 2], pad = "SAME")) # [4, 4, 64] x = U.flattenallbut0(x) # [1024] x = tf.nn.relu(U.dense(x, 256, 'l1', U.normc_initializer(1.0))) # 1024 mu = U.dense(x, latent_dim, 'l1_1', U.normc_initializer(1.0)) # 32 logvar = U.dense(x, latent_dim, 'l1_2', U.normc_initializer(1.0)) # 32 return mu, logvar
def proj_net(scope, img, latent_dim):
with tf.variable_scope(scope):
x = img
x = tf.nn.relu(U.conv2d(x, 64, "c1", [8, 8], [2, 2], pad="SAME"))
x = tf.nn.relu(U.conv2d(x, 128, "c2", [6, 6], [2, 2], pad="SAME"))
x = tf.nn.relu(U.conv2d(x, 128, "c3", [6, 6], [2, 2], pad="SAME"))
x = tf.nn.relu(U.conv2d(x, 128, "c4", [4, 4], [2, 2], pad="SAME"))
x = U.flattenallbut0(x)
x = tf.nn.relu(U.dense(x, 2048, 'l1', U.normc_initializer(1.0)))
x = U.dense(x, latent_dim, 'l2', U.normc_initializer(1.0))
return x
def _init(self, ob_space, ac_space):
assert isinstance(ob_space, gym.spaces.Box)
self.pdtype = pdtype = make_pdtype(ac_space)
sequence_length = None
sy_ob = U.get_placeholder(name="sy_ob",
dtype=tf.float32,
shape=[sequence_length] +
list(ob_space.shape))
obscaled = sy_ob / 255.0
with tf.variable_scope("pol"):
x = obscaled
x = tf.nn.relu(U.conv2d(x, 8, "l1", [8, 8], [4, 4], pad="VALID"))
x = tf.nn.relu(U.conv2d(x, 16, "l2", [4, 4], [2, 2], pad="VALID"))
x = U.flattenallbut0(x)
x = tf.nn.relu(U.dense(x, 128, 'lin', U.normc_initializer(1.0)))
logits = U.dense(x,
pdtype.param_shape()[0], "logits",
U.normc_initializer(0.01))
self.pd = pdtype.pdfromflat(logits)
with tf.variable_scope("vf"):
x = obscaled
x = tf.nn.relu(U.conv2d(x, 8, "l1", [8, 8], [4, 4], pad="VALID"))
x = tf.nn.relu(U.conv2d(x, 16, "l2", [4, 4], [2, 2], pad="VALID"))
x = U.flattenallbut0(x)
x = tf.nn.relu(U.dense(x, 128, 'lin', U.normc_initializer(1.0)))
self.vpred = U.dense(x, 1, "value", U.normc_initializer(1.0))
self.vpredz = self.vpred
self.state_in = []
self.state_out = []
stochastic = tf.placeholder(dtype=tf.bool, shape=())
sy_ac = self.pd.sample() # XXX
self._act = U.function([stochastic, sy_ob], [sy_ac, self.vpred])