tf scoring test

scoring_test.py

import tensorflow as tf
import numpy as np

predicates_in_batch = 2
seq_len = 3
vn_labels = 4
d = 5

np.random.seed(5)
tf.set_random_seed(5)

# [predicates_in_batch, seq_len, vn_labels]
vn_scores = [
  [
    [1, 0, 0, 0],
    [0.5, 0, 0.5, 0],
    [0.2, 0.2, 0.3, 0.3]
  ],
  [
    [0.1, 0.2, 0.3, 0.4],
    [0, 0.1, 0.1, 0.8],
    [0, 1, 0, 0]
  ],
]

# [predicates_in_batch, seq_len, 1, vn_labels]
vn_scores_add_dim = tf.expand_dims(tf.constant(vn_scores, dtype=tf.float64), 2)

# [predicates_in_batch, vn_labels, d]
linear = tf.constant(np.random.rand(predicates_in_batch, vn_labels, d))

# [predicates_in_batch, seq_len, vn_labels, d]
linear_tiled = tf.tile(tf.expand_dims(linear, 1), [1, seq_len, 1, 1])

actual = tf.squeeze(tf.matmul(vn_scores_add_dim, linear_tiled), 2)

np.set_printoptions(threshold=np.inf)
with tf.Session() as sess:
  sess.run(tf.tables_initializer())
  linear_np, linear_tiled_np, vn_scores_add_dim_np, actual_np = sess.run([linear, linear_tiled, vn_scores_add_dim, actual])
  print("actual", actual_np)

  expected = np.empty((predicates_in_batch, seq_len, d))
  for p_idx, predicate in enumerate(vn_scores):
    # [vn_labels, d]
    p_lin = linear_np[p_idx]
    for t_idx, tok in enumerate(predicate):
      # [vn_labels]
      composed = np.dot(tok, p_lin)
      expected[p_idx, t_idx] = composed

  print("expected", expected)

  np.testing.assert_almost_equal(actual_np, expected)