def __init__(self, name, dimensions, gate_fun, loss_fun, SEED=None):
# placeholders
# n1 x n2 = image dimensions
# d0 = number of channels (rgb or greyscale)
# d1 = number of independent filters
# f1 = filter size
n1, n2, d0, f1, d1, hidden, m = dimensions
self.x = tf.placeholder(tf.float32, shape=(None, n1, n2, d0)) # input
self.y = tf.placeholder(tf.float32, shape=(None, m)) # target
# layer 1: conv
W_1, b_1, z_hat_1, r_hat_1 = layers.convolution_2d(
name, "layer_1", self.x, f1, d0, d1,
tf.random_normal_initializer(stddev=1.0 / np.sqrt(f1 * f1 * d0 + 1), seed=SEED),
gate_fun)
# layer 1.5: pool
s_hat_1 = tf.nn.max_pool(
r_hat_1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
shape_1 = s_hat_1.get_shape().as_list()
y_hat_1 = tf.reshape(s_hat_1, [-1, shape_1[1] * shape_1[2] * shape_1[3]])
# layer 2: full
W_2, b_2, z_hat_2, y_hat_2 = layers.fully_connected(
name, "layer_2", y_hat_1, (n1 * n2 * d1) // 4, hidden,
tf.random_normal_initializer(stddev=1.0 / np.sqrt((n1 * n2 * d1) // 4),
seed=SEED),
gate_fun)
# layer 3: full
W_3, b_3, z_hat_3, y_hat_3 = layers.fully_connected(
name, "layer_3", y_hat_2, hidden, m,
tf.random_normal_initializer(stddev=1.0 / np.sqrt(hidden + 1), seed=SEED),
tf.nn.softmax)
# loss
self.train_loss = tf.reduce_sum(loss_fun(z_hat_3, self.y))
self.train_vars = [W_1, b_1, W_2, b_2, W_3, b_3]
self.misclass_err = tf.reduce_sum(tf.cast(
tf.not_equal(tf.argmax(y_hat_3, 1), tf.argmax(self.y, 1)), tf.float32))
def __init__(self, name, dimensions, gate_fun, loss_fun, SEED=None):
# placeholders
dim_in, hidden, dim_out = dimensions
self.x = tf.placeholder(tf.float32, shape=(None, dim_in)) # input
self.y = tf.placeholder(tf.float32, shape=(None, dim_out)) # target
# layer 1: full
W_1, b_1, z_hat_1, y_hat_1 = layers.fully_connected(
name, "layer_1", self.x, dim_in, hidden,
tf.random_normal_initializer(stddev=1.0/np.sqrt(dim_in+1), seed=SEED),
gate_fun)
# layer 2: full
W_2, b_2, z_hat_2, y_hat_2 = layers.fully_connected(
name, "layer_2", y_hat_1, hidden, hidden,
tf.random_normal_initializer(stddev=1.0/np.sqrt(hidden+1), seed=SEED),
gate_fun)
# layer 3: full
W_3, b_3, z_hat, self.y_hat = layers.fully_connected(
name, "layer_3", y_hat_2, hidden, dim_out,
tf.random_normal_initializer(stddev=1.0 / np.sqrt(hidden + 1), seed=SEED),
tf.nn.softmax)
# loss
self.train_loss = tf.reduce_sum(loss_fun(z_hat, self.y))
self.train_vars = [W_1, b_1, W_2, b_2, W_3, b_3]
self.misclass_err = tf.reduce_sum(tf.cast(
tf.not_equal(tf.argmax(self.y_hat, 1), tf.argmax(self.y, 1)), tf.float32))
def __init__(self, name, dimensions, gate_fun, loss_fun, SEED=None):
# placeholders
n1, n2, d0, f1, d1, f2, d2, m = dimensions
self.x = tf.placeholder(tf.float32, shape=(None, n1, n2, d0)) # input
self.y = tf.placeholder(tf.float32, shape=(None, m)) # target
# layer 1: conv
W_1, b_1, z_hat_1, r_hat_1 = layers.convolution_2d(
name, "layer_1", self.x, f1, d0, d1,
tf.random_normal_initializer(stddev=1.0/np.sqrt(f1*f1*d0+1), seed=SEED),
gate_fun)
# layer 1.5: pool
s_hat_1 = tf.nn.max_pool(
r_hat_1, ksize=[1,2,2,1], strides=[1,2,2,1], padding="SAME")
# layer 2: conv
W_2, b_2, z_hat_2, r_hat_2 = layers.convolution_2d(
name, "layer_2", s_hat_1, f2, d1, d2,
tf.random_normal_initializer(stddev=1.0/np.sqrt(f2*f2*d1+1), seed=SEED),
gate_fun)
# layer 2.5: pool
s_hat_2 = tf.nn.max_pool(
r_hat_2, ksize=[1,2,2,1], strides=[1,2,2,1], padding="SAME")
shape_2 = s_hat_2.get_shape().as_list()
y_hat_2 = tf.reshape(s_hat_2, [-1, shape_2[1]*shape_2[2]*shape_2[3]])
# layer 3: full
W_3, b_3, z_hat, y_hat = layers.fully_connected(
name, "layer_3", y_hat_2, (n1*n2*d2)//16, m,
tf.random_normal_initializer(stddev=1.0/np.sqrt((n1*n2*d2)//16),
seed=SEED),
tf.nn.softmax)
# loss
self.train_loss = tf.reduce_sum(loss_fun(z_hat, self.y))
self.train_vars = [W_1, b_1, W_2, b_2, W_3, b_3]
self.misclass_err = tf.reduce_sum(tf.cast(
tf.not_equal(tf.argmax(y_hat, 1), tf.argmax(self.y, 1)), tf.float32))