1e-10, dtype=tf.float32) # simulate hard sort
)
experiment_id = 'sort-%s-M%d-n%d-l%d-t%d' % (method, M, n, l, tau * 10)
checkpoint_path = 'checkpoints/%s/' % experiment_id
handle = tf.placeholder(tf.string, ())
X_iterator = tf.data.Iterator.from_string_handle(
handle,
(tf.float32, tf.float32, tf.float32, tf.float32),
((M, n, l * 28, 28), (M,), (M, n), (M, n))
)
X, y, median_scores, true_scores = X_iterator.get_next()
true_scores = tf.expand_dims(true_scores, 2)
P_true = util.neuralsort(true_scores, 1e-10)
if method == 'vanilla':
representations = multi_mnist_cnn.deepnn(l, X, n)
concat_reps = tf.reshape(representations, [M, n * n])
fc1 = tf.layers.dense(concat_reps, n * n)
fc2 = tf.layers.dense(fc1, n * n)
P_hat_raw = tf.layers.dense(fc2, n * n)
P_hat_raw_square = tf.reshape(P_hat_raw, [M, n, n])
P_hat = tf.nn.softmax(P_hat_raw_square, dim=-1) # row-stochastic!
losses = tf.nn.softmax_cross_entropy_with_logits_v2(
labels=P_true, logits=P_hat_raw_square, dim=2)
losses = tf.reduce_mean(losses, axis=-1)
loss = tf.reduce_mean(losses)
temp = tf.cond(evaluation,
false_fn=lambda: tf.convert_to_tensor(tau, dtype=tf.float32),
true_fn=lambda: tf.convert_to_tensor(1e-10, dtype=tf.float32))
experiment_id = 'median-%s-M%d-n%d-l%d-t%d' % (method, M, n, l, tau * 10)
checkpoint_path = 'checkpoints/%s/' % experiment_id
handle = tf.placeholder(tf.string, ())
X_iterator = tf.data.Iterator.from_string_handle(
handle, (tf.float32, tf.float32, tf.float32, tf.float32),
((M, n, l * 28, 28), (M, ), (M, n), (M, n)))
X, y, median_scores, true_scores = X_iterator.get_next()
true_scores = tf.expand_dims(true_scores, 2)
P_true = util.neuralsort(true_scores, 1e-10)
n_prime = n
def get_median_probs(P):
median_strip = P[:, n // 2, :]
median_total = tf.reduce_sum(median_strip, axis=1, keepdims=True)
probs = median_strip / median_total
# print(probs)
return probs
if method == 'vanilla':
with tf.variable_scope("phi"):
representations = multi_mnist_cnn.deepnn(l, X, 10)
representations = tf.reshape(representations, [M, n * 10])