def get_hit_rate_and_ndcg(self,
predicted_scores_by_user,
items_by_user,
top_k=rconst.TOP_K,
match_mlperf=False):
rconst.TOP_K = top_k
rconst.NUM_EVAL_NEGATIVES = predicted_scores_by_user.shape[1] - 1
g = tf.Graph()
with g.as_default():
logits = tf.convert_to_tensor(
predicted_scores_by_user.reshape((-1, 1)), tf.float32)
softmax_logits = tf.concat(
[tf.zeros(logits.shape, dtype=logits.dtype), logits], axis=1)
duplicate_mask = tf.convert_to_tensor(
stat_utils.mask_duplicates(items_by_user, axis=1), tf.float32)
metric_ops = neumf_model.compute_eval_loss_and_metrics(
logits=logits,
softmax_logits=softmax_logits,
duplicate_mask=duplicate_mask,
num_training_neg=NUM_TRAIN_NEG,
match_mlperf=match_mlperf).eval_metric_ops
hr = metric_ops[rconst.HR_KEY]
ndcg = metric_ops[rconst.NDCG_KEY]
init = [
tf.global_variables_initializer(),
tf.local_variables_initializer()
]
with self.test_session(graph=g) as sess:
sess.run(init)
return sess.run([hr[1], ndcg[1]])
def get_hit_rate_and_ndcg(self, predicted_scores_by_user, items_by_user,
top_k=rconst.TOP_K, match_mlperf=False):
rconst.TOP_K = top_k
rconst.NUM_EVAL_NEGATIVES = predicted_scores_by_user.shape[1] - 1
g = tf.Graph()
with g.as_default():
logits = tf.convert_to_tensor(
predicted_scores_by_user.reshape((-1, 1)), tf.float32)
softmax_logits = tf.concat([tf.zeros(logits.shape, dtype=logits.dtype),
logits], axis=1)
duplicate_mask = tf.convert_to_tensor(
stat_utils.mask_duplicates(items_by_user, axis=1), tf.float32)
metric_ops = neumf_model.compute_eval_loss_and_metrics(
logits=logits, softmax_logits=softmax_logits,
duplicate_mask=duplicate_mask, num_training_neg=NUM_TRAIN_NEG,
match_mlperf=match_mlperf).eval_metric_ops
hr = metric_ops[rconst.HR_KEY]
ndcg = metric_ops[rconst.NDCG_KEY]
init = [tf.global_variables_initializer(),
tf.local_variables_initializer()]
with self.test_session(graph=g) as sess:
sess.run(init)
return sess.run([hr[1], ndcg[1]])
def get_hit_rate_and_ndcg(self, predicted_scores_by_user, items_by_user,
top_k=rconst.TOP_K, match_mlperf=False):
rconst.TOP_K = top_k
rconst.NUM_EVAL_NEGATIVES = predicted_scores_by_user.shape[1] - 1
batch_size = items_by_user.shape[0]
users = np.repeat(np.arange(batch_size)[:, np.newaxis],
rconst.NUM_EVAL_NEGATIVES + 1, axis=1)
users, items, duplicate_mask = \
data_pipeline.BaseDataConstructor._assemble_eval_batch(
users, items_by_user[:, -1:], items_by_user[:, :-1], batch_size)
g = tf.Graph()
with g.as_default():
logits = tf.convert_to_tensor(
predicted_scores_by_user.reshape((-1, 1)), tf.float32)
softmax_logits = tf.concat([tf.zeros(logits.shape, dtype=logits.dtype),
logits], axis=1)
duplicate_mask = tf.convert_to_tensor(duplicate_mask, tf.float32)
metric_ops = neumf_model.compute_eval_loss_and_metrics(
logits=logits, softmax_logits=softmax_logits,
duplicate_mask=duplicate_mask, num_training_neg=NUM_TRAIN_NEG,
match_mlperf=match_mlperf).eval_metric_ops
hr = metric_ops[rconst.HR_KEY]
ndcg = metric_ops[rconst.NDCG_KEY]
init = [tf.global_variables_initializer(),
tf.local_variables_initializer()]
with self.test_session(graph=g) as sess:
sess.run(init)
return sess.run([hr[1], ndcg[1]])
def get_hit_rate_and_ndcg(self, predicted_scores_by_user, items_by_user,
top_k=rconst.TOP_K, match_mlperf=False):
rconst.TOP_K = top_k
rconst.NUM_EVAL_NEGATIVES = predicted_scores_by_user.shape[1] - 1
batch_size = items_by_user.shape[0]
users = np.repeat(np.arange(batch_size)[:, np.newaxis],
rconst.NUM_EVAL_NEGATIVES + 1, axis=1)
users, items, duplicate_mask = \
data_pipeline.BaseDataConstructor._assemble_eval_batch(
users, items_by_user[:, -1:], items_by_user[:, :-1], batch_size)
g = tf.Graph()
with g.as_default():
logits = tf.convert_to_tensor(
predicted_scores_by_user.reshape((-1, 1)), tf.float32)
softmax_logits = tf.concat([tf.zeros(logits.shape, dtype=logits.dtype),
logits], axis=1)
duplicate_mask = tf.convert_to_tensor(duplicate_mask, tf.float32)
metric_ops = neumf_model.compute_eval_loss_and_metrics(
logits=logits, softmax_logits=softmax_logits,
duplicate_mask=duplicate_mask, num_training_neg=NUM_TRAIN_NEG,
match_mlperf=match_mlperf).eval_metric_ops
hr = metric_ops[rconst.HR_KEY]
ndcg = metric_ops[rconst.NDCG_KEY]
init = [tf.global_variables_initializer(),
tf.local_variables_initializer()]
with self.test_session(graph=g) as sess:
sess.run(init)
return sess.run([hr[1], ndcg[1]])