data_test.py 12.4 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Test NCF data pipeline."""

from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

21
from collections import defaultdict
22
import hashlib
23
24
import os

25
import mock
26
import numpy as np
27
import scipy.stats
28
29
30
31
32
import tensorflow as tf

from official.datasets import movielens
from official.recommendation import constants as rconst
from official.recommendation import data_preprocessing
33
from official.recommendation import popen_helper
34
35
36
37
38
39
40


DATASET = "ml-test"
NUM_USERS = 1000
NUM_ITEMS = 2000
NUM_PTS = 50000
BATCH_SIZE = 2048
41
EVAL_BATCH_SIZE = 4000
42
43
44
NUM_NEG = 4


45
46
47
48
49
END_TO_END_TRAIN_MD5 = "b218738e915e825d03939c5e305a2698"
END_TO_END_EVAL_MD5 = "d753d0f3186831466d6e218163a9501e"
FRESH_RANDOMNESS_MD5 = "63d0dff73c0e5f1048fbdc8c65021e22"


50
51
52
def mock_download(*args, **kwargs):
  return

53
54
55
56
# The forkpool used by data producers interacts badly with the threading
# used by TestCase. Without this patch tests will hang, and no amount
# of diligent closing and joining within the producer will prevent it.
@mock.patch.object(popen_helper, "get_forkpool", popen_helper.get_fauxpool)
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
class BaseTest(tf.test.TestCase):
  def setUp(self):
    self.temp_data_dir = self.get_temp_dir()
    ratings_folder = os.path.join(self.temp_data_dir, DATASET)
    tf.gfile.MakeDirs(ratings_folder)
    np.random.seed(0)
    raw_user_ids = np.arange(NUM_USERS * 3)
    np.random.shuffle(raw_user_ids)
    raw_user_ids = raw_user_ids[:NUM_USERS]

    raw_item_ids = np.arange(NUM_ITEMS * 3)
    np.random.shuffle(raw_item_ids)
    raw_item_ids = raw_item_ids[:NUM_ITEMS]

    users = np.random.choice(raw_user_ids, NUM_PTS)
    items = np.random.choice(raw_item_ids, NUM_PTS)
    scores = np.random.randint(low=0, high=5, size=NUM_PTS)
    times = np.random.randint(low=1000000000, high=1200000000, size=NUM_PTS)

76
    self.rating_file = os.path.join(ratings_folder, movielens.RATINGS_FILE)
77
78
    self.seen_pairs = set()
    self.holdout = {}
79
    with tf.gfile.Open(self.rating_file, "w") as f:
80
81
82
83
84
85
86
87
88
89
90
91
92
      f.write("user_id,item_id,rating,timestamp\n")
      for usr, itm, scr, ts in zip(users, items, scores, times):
        pair = (usr, itm)
        if pair in self.seen_pairs:
          continue
        self.seen_pairs.add(pair)
        if usr not in self.holdout or (ts, itm) > self.holdout[usr]:
          self.holdout[usr] = (ts, itm)

        f.write("{},{},{},{}\n".format(usr, itm, scr, ts))

    movielens.download = mock_download
    movielens.NUM_RATINGS[DATASET] = NUM_PTS
93
94
    data_preprocessing.DATASET_TO_NUM_USERS_AND_ITEMS[DATASET] = (NUM_USERS,
                                                                  NUM_ITEMS)
95

96
97
98
99
100
101
102
103
104
105
106
107
108
  def make_params(self, train_epochs=1):
    return {
        "train_epochs": train_epochs,
        "batches_per_step": 1,
        "use_seed": False,
        "batch_size": BATCH_SIZE,
        "eval_batch_size": EVAL_BATCH_SIZE,
        "num_neg": NUM_NEG,
        "match_mlperf": True,
        "use_tpu": False,
        "use_xla_for_gpu": False,
    }

109
110
  def test_preprocessing(self):
    # For the most part the necessary checks are performed within
111
112
    # _filter_index_sort()

113
114
115
116
117
118
    cache_path = os.path.join(self.temp_data_dir, "test_cache.pickle")
    data, valid_cache = data_preprocessing._filter_index_sort(
        self.rating_file, cache_path=cache_path)

    assert len(data[rconst.USER_MAP]) == NUM_USERS
    assert len(data[rconst.ITEM_MAP]) == NUM_ITEMS
119
120
121
122
123
124
125
126
127
128
129
130
131
132

  def drain_dataset(self, dataset, g):
    # type: (tf.data.Dataset, tf.Graph) -> list
    with self.test_session(graph=g) as sess:
      with g.as_default():
        batch = dataset.make_one_shot_iterator().get_next()
      output = []
      while True:
        try:
          output.append(sess.run(batch))
        except tf.errors.OutOfRangeError:
          break
    return output

133
  def _test_end_to_end(self, constructor_type):
134
135
    params = self.make_params(train_epochs=1)
    _, _, producer = data_preprocessing.instantiate_pipeline(
136
137
        dataset=DATASET, data_dir=self.temp_data_dir, params=params,
        constructor_type=constructor_type, deterministic=True)
138
139
140
141

    producer.start()
    producer.join()
    assert producer._fatal_exception is None
142

143
144
145
146
147
148
    user_inv_map = {v: k for k, v in producer.user_map.items()}
    item_inv_map = {v: k for k, v in producer.item_map.items()}

    # ==========================================================================
    # == Training Data =========================================================
    # ==========================================================================
149
150
    g = tf.Graph()
    with g.as_default():
151
152
153
      input_fn = producer.make_input_fn(is_training=True)
      dataset = input_fn(params)

154
155
    first_epoch = self.drain_dataset(dataset=dataset, g=g)

156
    counts = defaultdict(int)
157
158
159
160
    train_examples = {
        True: set(),
        False: set(),
    }
161

162
    md5 = hashlib.md5()
163
    for features, labels in first_epoch:
164
      data_list = [
165
          features[movielens.USER_COLUMN], features[movielens.ITEM_COLUMN],
166
          features[rconst.VALID_POINT_MASK], labels]
Taylor Robie's avatar
Taylor Robie committed
167
168
169
      for i in data_list:
        md5.update(i.tobytes())

170
      for u, i, v, l in zip(*data_list):
171
172
        if not v:
          continue  # ignore padding
173

174
175
176
177
178
179
180
        u_raw = user_inv_map[u]
        i_raw = item_inv_map[i]
        if ((u_raw, i_raw) in self.seen_pairs) != l:
          # The evaluation item is not considered during false negative
          # generation, so it will occasionally appear as a negative example
          # during training.
          assert not l
181
          self.assertEqual(i_raw, self.holdout[u_raw][1])
182
        train_examples[l].add((u_raw, i_raw))
183
184
        counts[(u_raw, i_raw)] += 1

185
    self.assertRegexpMatches(md5.hexdigest(), END_TO_END_TRAIN_MD5)
186

187
    num_positives_seen = len(train_examples[True])
188
    self.assertEqual(producer._train_pos_users.shape[0], num_positives_seen)
189
190
191

    # This check is more heuristic because negatives are sampled with
    # replacement. It only checks that negative generation is reasonably random.
192
193
194
195
196
197
198
199
200
201
202
203
204
205
    self.assertGreater(
        len(train_examples[False]) / NUM_NEG / num_positives_seen, 0.9)

    # This checks that the samples produced are independent by checking the
    # number of duplicate entries. If workers are not properly independent there
    # will be lots of repeated pairs.
    self.assertLess(np.mean(list(counts.values())), 1.1)

    # ==========================================================================
    # == Eval Data =============================================================
    # ==========================================================================
    with g.as_default():
      input_fn = producer.make_input_fn(is_training=False)
      dataset = input_fn(params)
206

207
    eval_data = self.drain_dataset(dataset=dataset, g=g)
208

209
    current_user = None
210
    md5 = hashlib.md5()
211
    for features in eval_data:
212
213
214
      data_list = [
          features[movielens.USER_COLUMN], features[movielens.ITEM_COLUMN],
          features[rconst.DUPLICATE_MASK]]
Taylor Robie's avatar
Taylor Robie committed
215
216
217
      for i in data_list:
        md5.update(i.tobytes())

218
      for idx, (u, i, d) in enumerate(zip(*data_list)):
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
        u_raw = user_inv_map[u]
        i_raw = item_inv_map[i]
        if current_user is None:
          current_user = u

        # Ensure that users appear in blocks, as the evaluation logic expects
        # this structure.
        self.assertEqual(u, current_user)

        # The structure of evaluation data is 999 negative examples followed
        # by the holdout positive.
        if not (idx + 1) % (rconst.NUM_EVAL_NEGATIVES + 1):
          # Check that the last element in each chunk is the holdout item.
          self.assertEqual(i_raw, self.holdout[u_raw][1])
          current_user = None

        elif i_raw == self.holdout[u_raw][1]:
          # Because the holdout item is not given to the negative generation
          # process, it can appear as a negative. In that case, it should be
          # masked out as a duplicate. (Since the true positive is placed at
          # the end and would therefore lose the tie.)
          assert d

        else:
          # Otherwise check that the other 999 points for a user are selected
          # from the negatives.
          assert (u_raw, i_raw) not in self.seen_pairs

247
248
249
    self.assertRegexpMatches(md5.hexdigest(), END_TO_END_EVAL_MD5)

  def _test_fresh_randomness(self, constructor_type):
250
251
252
    train_epochs = 5
    params = self.make_params(train_epochs=train_epochs)
    _, _, producer = data_preprocessing.instantiate_pipeline(
253
254
        dataset=DATASET, data_dir=self.temp_data_dir, params=params,
        constructor_type=constructor_type, deterministic=True)
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269

    producer.start()

    results = []
    g = tf.Graph()
    with g.as_default():
      for _ in range(train_epochs):
        input_fn = producer.make_input_fn(is_training=True)
        dataset = input_fn(params)
        results.extend(self.drain_dataset(dataset=dataset, g=g))

    producer.join()
    assert producer._fatal_exception is None

    positive_counts, negative_counts = defaultdict(int), defaultdict(int)
270
    md5 = hashlib.md5()
271
    for features, labels in results:
272
      data_list = [
273
          features[movielens.USER_COLUMN], features[movielens.ITEM_COLUMN],
274
          features[rconst.VALID_POINT_MASK], labels]
Taylor Robie's avatar
Taylor Robie committed
275
276
277
      for i in data_list:
        md5.update(i.tobytes())

278
      for u, i, v, l in zip(*data_list):
279
280
281
282
283
284
285
286
        if not v:
          continue  # ignore padding

        if l:
          positive_counts[(u, i)] += 1
        else:
          negative_counts[(u, i)] += 1

287
288
    self.assertRegexpMatches(md5.hexdigest(), FRESH_RANDOMNESS_MD5)

289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
    # The positive examples should appear exactly once each epoch
    self.assertAllEqual(list(positive_counts.values()),
                        [train_epochs for _ in positive_counts])

    # The threshold for the negatives is heuristic, but in general repeats are
    # expected, but should not appear too frequently.

    pair_cardinality = NUM_USERS * NUM_ITEMS
    neg_pair_cardinality = pair_cardinality - len(self.seen_pairs)

    # Approximation for the expectation number of times that a particular
    # negative will appear in a given epoch. Implicit in this calculation is the
    # treatment of all negative pairs as equally likely. Normally is not
    # necessarily reasonable; however the generation in self.setUp() will
    # approximate this behavior sufficiently for heuristic testing.
    e_sample = len(self.seen_pairs) * NUM_NEG / neg_pair_cardinality

    # The frequency of occurance of a given negative pair should follow an
    # approximately binomial distribution in the limit that the cardinality of
    # the negative pair set >> number of samples per epoch.
    approx_pdf = scipy.stats.binom.pmf(k=np.arange(train_epochs+1),
                                       n=train_epochs, p=e_sample)

    # Tally the actual observed counts.
    count_distribution = [0 for _ in range(train_epochs + 1)]
    for i in negative_counts.values():
      i = min([i, train_epochs])  # round down tail for simplicity.
      count_distribution[i] += 1
    count_distribution[0] = neg_pair_cardinality - sum(count_distribution[1:])

    # Check that the frequency of negative pairs is approximately binomial.
    for i in range(train_epochs + 1):
      if approx_pdf[i] < 0.05:
        continue  # Variance will be high at the tails.

      observed_fraction = count_distribution[i] / neg_pair_cardinality
      deviation = (2 * abs(observed_fraction - approx_pdf[i]) /
                   (observed_fraction + approx_pdf[i]))

      self.assertLess(deviation, 0.2)
329

Taylor Robie's avatar
Taylor Robie committed
330
331
  def test_end_to_end_materialized(self):
    self._test_end_to_end("materialized")
332

Taylor Robie's avatar
Taylor Robie committed
333
334
335
336
337
338
339
340
  def test_end_to_end_bisection(self):
    self._test_end_to_end("bisection")

  def test_fresh_randomness_materialized(self):
    self._test_fresh_randomness("materialized")

  def test_fresh_randomness_bisection(self):
    self._test_fresh_randomness("bisection")
341

342
343
344

if __name__ == "__main__":
  tf.logging.set_verbosity(tf.logging.INFO)
Taylor Robie's avatar
Taylor Robie committed
345
  tf.test.main()