# binarize data and create a y "placeholder"
with tf.name_scope('x_binarized'):
xb = tf.cast(tf.greater(x, tf.random_uniform(tf.shape(x), 0, 1)),
tf.float32)
with tf.name_scope('y_'):
y_ = tf.fill(tf.pack([tf.shape(x)[0], k]), 0.0)
# propose distribution over y
qy_logit, qy = qy_graph(xb, k)
# for each proposed y, infer z and reconstruct x
z, zm, zv, zm_prior, zv_prior, px_logit = [[None] * k for i in xrange(6)]
for i in xrange(k):
with tf.name_scope('graphs/hot_at{:d}'.format(i)):
y = tf.add(y_, Constant(np.eye(k)[i], name='hot_at_{:d}'.format(i)))
z[i], zm[i], zv[i] = qz_graph(xb, y)
zm_prior[i], zv_prior[i], px_logit[i] = px_graph(z[i], y)
# Aggressive name scoping for pretty graph visualization :P
with tf.name_scope('loss'):
with tf.name_scope('neg_entropy'):
nent = -cross_entropy_with_logits(qy_logit, qy)
losses = [None] * k
for i in xrange(k):
with tf.name_scope('loss_at{:d}'.format(i)):
losses[i] = labeled_loss(xb, px_logit[i], z[i], zm[i], zv[i],
zm_prior[i], zv_prior[i])
with tf.name_scope('final_loss'):
loss = tf.add_n([nent] + [qy[:, i] * losses[i] for i in xrange(k)])
train_step = tf.train.AdamOptimizer().minimize(loss)
# h1 = tf.layers.conv2d(inputs=x,filters=16,kernel_size=[3, 3],padding="same",activation=tf.nn.relu) # pool1 = tf.layers.max_pooling2d(inputs=h1, pool_size=[2, 2], strides=2) # h2 = tf.layers.conv2d(inputs=pool1,filters=32,kernel_size=[3, 3],padding="same",activation=tf.nn.relu) # pool2 = tf.layers.max_pooling2d(inputs=h2, pool_size=[2, 2], strides=2) # pool2_flat = tf.contrib.layers.flatten(pool2) # dense = tf.layers.dense(inputs=pool2_flat, units=256, activation=tf.nn.relu) # dense = tf.layers.dense(inputs=dense, units=14*14*28, activation=tf.nn.relu) # dense = tf.reshape(dense,[-1,14,14,28]) # conv1 = tf.layers.conv2d_transpose(dense,28,[3,3],(1,1),padding="same",activation=tf.nn.relu) # conv2 = tf.layers.conv2d_transpose(conv1,28,[3,3],(2,2),padding="same",activation=tf.nn.relu) # conv3 = tf.layers.conv2d_transpose(conv2,1,[3,3],(1,1),padding="same",activation=tf.nn.relu) # logits = tf.contrib.layers.flatten(conv3) xb = tf.cast(tf.greater(x, tf.random_uniform(tf.shape(x), 0, 1)), tf.float32) y_ = tf.fill(tf.stack([tf.shape(x)[0], 10]), 0.0) z, zm, zv, zm_prior, zv_prior, px_logit = [[None] * 10 for i in xrange(6)] # for i in range(10): # y = tf.add(y_, Constant(np.eye(10)[i])) # z[i], zm[i], zv[i] = qz_graph(xb, y) # zm_prior[i], zv_prior[i], px_logit[i] = px_graph(z[i], y) y = tf.add(y_, Constant(np.eye(10)[1])) z[1], zm[1], zv[1] = qz_graph(xb, y) y = tf.add(y_, Constant(np.eye(10)[2])) z[2], zm[2], zv[2] = qz_graph(xb, y) y = tf.add(y_, Constant(np.eye(10)[3])) z[3], zm[3], zv[3] = qz_graph(xb, y) sess = tf.Session() sess.run(tf.initialize_all_variables()) print(np.shape(sess.run(z[2])))
