ncf_common.py 13 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
# 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.
# ==============================================================================
Shining Sun's avatar
Shining Sun committed
15
"""Common functionalities used by both Keras and Estimator implementations.
16
"""
17

18
19
20
21
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function

22
import json
23
24
import os

25
# pylint: disable=g-bad-import-order
26
import numpy as np
27
from absl import flags
28
from absl import logging
29
import tensorflow as tf
30
# pylint: enable=g-bad-import-order
31

32
from official.datasets import movielens
33
from official.recommendation import constants as rconst
34
from official.recommendation import data_pipeline
35
from official.recommendation import data_preprocessing
36
from official.utils.flags import core as flags_core
37
from official.utils.misc import distribution_utils
38
39


Reed's avatar
Reed committed
40
41
42
FLAGS = flags.FLAGS


Shining Sun's avatar
Shining Sun committed
43
44
45
46
def get_inputs(params):
  """Returns some parameters used by the model."""
  if FLAGS.download_if_missing and not FLAGS.use_synthetic_data:
    movielens.download(FLAGS.dataset, FLAGS.data_dir)
47

Shining Sun's avatar
Shining Sun committed
48
49
  if FLAGS.seed is not None:
    np.random.seed(FLAGS.seed)
50

Shining Sun's avatar
Shining Sun committed
51
52
53
54
55
56
  if FLAGS.use_synthetic_data:
    producer = data_pipeline.DummyConstructor()
    num_users, num_items = data_preprocessing.DATASET_TO_NUM_USERS_AND_ITEMS[
        FLAGS.dataset]
    num_train_steps = rconst.SYNTHETIC_BATCHES_PER_EPOCH
    num_eval_steps = rconst.SYNTHETIC_BATCHES_PER_EPOCH
57
  else:
Shining Sun's avatar
Shining Sun committed
58
59
60
61
    num_users, num_items, producer = data_preprocessing.instantiate_pipeline(
        dataset=FLAGS.dataset, data_dir=FLAGS.data_dir, params=params,
        constructor_type=FLAGS.constructor_type,
        deterministic=FLAGS.seed is not None)
62

Shining Sun's avatar
Shining Sun committed
63
64
65
66
67
68
    num_train_steps = (producer.train_batches_per_epoch //
                       params["batches_per_step"])
    num_eval_steps = (producer.eval_batches_per_epoch //
                      params["batches_per_step"])
    assert not producer.train_batches_per_epoch % params["batches_per_step"]
    assert not producer.eval_batches_per_epoch % params["batches_per_step"]
69

Shining Sun's avatar
Shining Sun committed
70
  return num_users, num_items, num_train_steps, num_eval_steps, producer
71
72
73


def parse_flags(flags_obj):
Taylor Robie's avatar
Taylor Robie committed
74
  """Convenience function to turn flags into params."""
75
76
77
  num_gpus = flags_core.get_num_gpus(flags_obj)
  num_devices = FLAGS.num_tpu_shards if FLAGS.tpu else num_gpus or 1

Taylor Robie's avatar
Taylor Robie committed
78
  batch_size = (flags_obj.batch_size + num_devices - 1) // num_devices
79
80

  eval_divisor = (rconst.NUM_EVAL_NEGATIVES + 1) * num_devices
Taylor Robie's avatar
Taylor Robie committed
81
82
83
  eval_batch_size = flags_obj.eval_batch_size or flags_obj.batch_size
  eval_batch_size = ((eval_batch_size + eval_divisor - 1) //
                     eval_divisor * eval_divisor // num_devices)
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106

  return {
      "train_epochs": flags_obj.train_epochs,
      "batches_per_step": num_devices,
      "use_seed": flags_obj.seed is not None,
      "batch_size": batch_size,
      "eval_batch_size": eval_batch_size,
      "learning_rate": flags_obj.learning_rate,
      "mf_dim": flags_obj.num_factors,
      "model_layers": [int(layer) for layer in flags_obj.layers],
      "mf_regularization": flags_obj.mf_regularization,
      "mlp_reg_layers": [float(reg) for reg in flags_obj.mlp_regularization],
      "num_neg": flags_obj.num_neg,
      "num_gpus": num_gpus,
      "use_tpu": flags_obj.tpu is not None,
      "tpu": flags_obj.tpu,
      "tpu_zone": flags_obj.tpu_zone,
      "tpu_gcp_project": flags_obj.tpu_gcp_project,
      "beta1": flags_obj.beta1,
      "beta2": flags_obj.beta2,
      "epsilon": flags_obj.epsilon,
      "match_mlperf": flags_obj.ml_perf,
      "use_xla_for_gpu": flags_obj.use_xla_for_gpu,
107
108
      "clone_model_in_keras_dist_strat":
          flags_obj.clone_model_in_keras_dist_strat,
109
      "epochs_between_evals": FLAGS.epochs_between_evals,
Shining Sun's avatar
Shining Sun committed
110
      "turn_off_distribution_strategy": FLAGS.turn_off_distribution_strategy,
111
      "keras_use_ctl": flags_obj.keras_use_ctl,
112
      "hr_threshold": flags_obj.hr_threshold,
113
  }
114
115


Shining Sun's avatar
Shining Sun committed
116
117
118
119
def get_distribution_strategy(params):
  """Returns the distribution strategy to use."""
  if params["turn_off_distribution_strategy"]:
    return None
120

Shining Sun's avatar
Shining Sun committed
121
122
123
124
125
  if params["use_tpu"]:
    # Some of the networking libraries are quite chatty.
    for name in ["googleapiclient.discovery", "googleapiclient.discovery_cache",
                 "oauth2client.transport"]:
      logging.getLogger(name).setLevel(logging.ERROR)
126

127
    tpu_cluster_resolver = tf.distribute.cluster_resolver.TPUClusterResolver(
Shining Sun's avatar
Shining Sun committed
128
129
130
131
132
        tpu=params["tpu"],
        zone=params["tpu_zone"],
        project=params["tpu_gcp_project"],
        coordinator_name="coordinator"
    )
133

134
    logging.info("Issuing reset command to TPU to ensure a clean state.")
Shining Sun's avatar
Shining Sun committed
135
    tf.Session.reset(tpu_cluster_resolver.get_master())
136

Shining Sun's avatar
Shining Sun committed
137
138
139
140
141
142
143
144
145
146
    # Estimator looks at the master it connects to for MonitoredTrainingSession
    # by reading the `TF_CONFIG` environment variable, and the coordinator
    # is used by StreamingFilesDataset.
    tf_config_env = {
        "session_master": tpu_cluster_resolver.get_master(),
        "eval_session_master": tpu_cluster_resolver.get_master(),
        "coordinator": tpu_cluster_resolver.cluster_spec()
                       .as_dict()["coordinator"]
    }
    os.environ['TF_CONFIG'] = json.dumps(tf_config_env)
147

148
    distribution = tf.distribute.experimental.TPUStrategy(
Shining Sun's avatar
Shining Sun committed
149
        tpu_cluster_resolver, steps_per_run=100)
150

Shining Sun's avatar
Shining Sun committed
151
152
153
  else:
    distribution = distribution_utils.get_distribution_strategy(
        num_gpus=params["num_gpus"])
154

Shining Sun's avatar
Shining Sun committed
155
  return distribution
156

157
158
159
160
161
162
163
164
165

def define_ncf_flags():
  """Add flags for running ncf_main."""
  # Add common flags
  flags_core.define_base(export_dir=False)
  flags_core.define_performance(
      num_parallel_calls=False,
      inter_op=False,
      intra_op=False,
166
      synthetic_data=True,
167
      max_train_steps=False,
168
169
      dtype=False,
      all_reduce_alg=False
170
  )
171
  flags_core.define_device(tpu=True)
172
173
174
175
176
177
178
179
180
  flags_core.define_benchmark()

  flags.adopt_module_key_flags(flags_core)

  flags_core.set_defaults(
      model_dir="/tmp/ncf/",
      data_dir="/tmp/movielens-data/",
      train_epochs=2,
      batch_size=256,
181
182
183
      hooks="ProfilerHook",
      tpu=None
  )
184
185
186
187
188
189
190
191

  # Add ncf-specific flags
  flags.DEFINE_enum(
      name="dataset", default="ml-1m",
      enum_values=["ml-1m", "ml-20m"], case_sensitive=False,
      help=flags_core.help_wrap(
          "Dataset to be trained and evaluated."))

192
193
194
195
  flags.DEFINE_boolean(
      name="download_if_missing", default=True, help=flags_core.help_wrap(
          "Download data to data_dir if it is not already present."))

196
  flags.DEFINE_integer(
197
198
199
200
201
202
      name="eval_batch_size", default=None, help=flags_core.help_wrap(
          "The batch size used for evaluation. This should generally be larger"
          "than the training batch size as the lack of back propagation during"
          "evaluation can allow for larger batch sizes to fit in memory. If not"
          "specified, the training batch size (--batch_size) will be used."))

203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
  flags.DEFINE_integer(
      name="num_factors", default=8,
      help=flags_core.help_wrap("The Embedding size of MF model."))

  # Set the default as a list of strings to be consistent with input arguments
  flags.DEFINE_list(
      name="layers", default=["64", "32", "16", "8"],
      help=flags_core.help_wrap(
          "The sizes of hidden layers for MLP. Example "
          "to specify different sizes of MLP layers: --layers=32,16,8,4"))

  flags.DEFINE_float(
      name="mf_regularization", default=0.,
      help=flags_core.help_wrap(
          "The regularization factor for MF embeddings. The factor is used by "
          "regularizer which allows to apply penalties on layer parameters or "
          "layer activity during optimization."))

  flags.DEFINE_list(
      name="mlp_regularization", default=["0.", "0.", "0.", "0."],
      help=flags_core.help_wrap(
          "The regularization factor for each MLP layer. See mf_regularization "
          "help for more info about regularization factor."))

  flags.DEFINE_integer(
      name="num_neg", default=4,
      help=flags_core.help_wrap(
          "The Number of negative instances to pair with a positive instance."))

  flags.DEFINE_float(
      name="learning_rate", default=0.001,
      help=flags_core.help_wrap("The learning rate."))

236
237
238
239
240
241
242
243
244
245
246
247
248
  flags.DEFINE_float(
      name="beta1", default=0.9,
      help=flags_core.help_wrap("beta1 hyperparameter for the Adam optimizer."))

  flags.DEFINE_float(
      name="beta2", default=0.999,
      help=flags_core.help_wrap("beta2 hyperparameter for the Adam optimizer."))

  flags.DEFINE_float(
      name="epsilon", default=1e-8,
      help=flags_core.help_wrap("epsilon hyperparameter for the Adam "
                                "optimizer."))

249
  flags.DEFINE_float(
250
      name="hr_threshold", default=1.0,
251
252
253
254
255
256
      help=flags_core.help_wrap(
          "If passed, training will stop when the evaluation metric HR is "
          "greater than or equal to hr_threshold. For dataset ml-1m, the "
          "desired hr_threshold is 0.68 which is the result from the paper; "
          "For dataset ml-20m, the threshold can be set as 0.95 which is "
          "achieved by MLPerf implementation."))
257

258
259
260
261
262
263
264
265
  flags.DEFINE_enum(
      name="constructor_type", default="bisection",
      enum_values=["bisection", "materialized"], case_sensitive=False,
      help=flags_core.help_wrap(
          "Strategy to use for generating false negatives. materialized has a"
          "precompute that scales badly, but a faster per-epoch construction"
          "time and can be faster on very large systems."))

266
  flags.DEFINE_bool(
267
      name="ml_perf", default=False,
268
269
270
271
272
273
274
275
276
277
278
279
280
      help=flags_core.help_wrap(
          "If set, changes the behavior of the model slightly to match the "
          "MLPerf reference implementations here: \n"
          "https://github.com/mlperf/reference/tree/master/recommendation/"
          "pytorch\n"
          "The two changes are:\n"
          "1. When computing the HR and NDCG during evaluation, remove "
          "duplicate user-item pairs before the computation. This results in "
          "better HRs and NDCGs.\n"
          "2. Use a different soring algorithm when sorting the input data, "
          "which performs better due to the fact the sorting algorithms are "
          "not stable."))

Reed's avatar
Reed committed
281
282
283
284
285
286
287
288
289
290
291
292
  flags.DEFINE_bool(
      name="output_ml_perf_compliance_logging", default=False,
      help=flags_core.help_wrap(
          "If set, output the MLPerf compliance logging. This is only useful "
          "if one is running the model for MLPerf. See "
          "https://github.com/mlperf/policies/blob/master/training_rules.adoc"
          "#submission-compliance-logs for details. This uses sudo and so may "
          "ask for your password, as root access is needed to clear the system "
          "caches, which is required for MLPerf compliance."
      )
  )

293
294
295
296
  flags.DEFINE_integer(
      name="seed", default=None, help=flags_core.help_wrap(
          "This value will be used to seed both NumPy and TensorFlow."))

Shining Sun's avatar
Shining Sun committed
297
298
299
300
301
302
  flags.DEFINE_boolean(
      name="turn_off_distribution_strategy",
      default=False,
      help=flags_core.help_wrap(
          "If set, do not use any distribution strategy."))

303
304
305
  @flags.validator("eval_batch_size", "eval_batch_size must be at least {}"
                   .format(rconst.NUM_EVAL_NEGATIVES + 1))
  def eval_size_check(eval_batch_size):
Taylor Robie's avatar
Taylor Robie committed
306
307
    return (eval_batch_size is None or
            int(eval_batch_size) > rconst.NUM_EVAL_NEGATIVES)
308

Reed's avatar
Reed committed
309
310
311
312
313
  flags.DEFINE_bool(
      name="use_xla_for_gpu", default=False, help=flags_core.help_wrap(
          "If True, use XLA for the model function. Only works when using a "
          "GPU. On TPUs, XLA is always used"))

314
315
316
317
  xla_message = "--use_xla_for_gpu is incompatible with --tpu"
  @flags.multi_flags_validator(["use_xla_for_gpu", "tpu"], message=xla_message)
  def xla_validator(flag_dict):
    return not flag_dict["use_xla_for_gpu"] or not flag_dict["tpu"]
Reed's avatar
Reed committed
318

319
320
321
322
323
324
325
  flags.DEFINE_bool(
      name="clone_model_in_keras_dist_strat",
      default=True,
      help=flags_core.help_wrap(
          'If False, then the experimental code path is used that doesn\'t '
          "clone models for distribution."))

326
327
328
329
330
331
  flags.DEFINE_bool(
      name="early_stopping",
      default=False,
      help=flags_core.help_wrap(
          'If True, we stop the training when it reaches hr_threshold'))

332
333
334
335
336
  flags.DEFINE_bool(
      name="keras_use_ctl",
      default=False,
      help=flags_core.help_wrap(
          'If True, we use a custom training loop for keras.'))
337

Shining Sun's avatar
Shining Sun committed
338
339
340
341
342
343
344
def convert_to_softmax_logits(logits):
  '''Convert the logits returned by the base model to softmax logits.

  Softmax with the first column of zeros is equivalent to sigmoid.
  '''
  softmax_logits = tf.concat([logits * 0, logits], axis=1)
  return softmax_logits