Sync to latest.

CODEOWNERS

@@ -17,6 +17,7 @@
 /research/inception/ @shlens @vincentvanhoucke
 /research/learned_optimizer/ @olganw @nirum
 /research/learning_to_remember_rare_events/ @lukaszkaiser @ofirnachum
+/research/learning_unsupervised_learning/ @lukemetz @nirum
 /research/lexnet_nc/ @vered1986 @waterson
 /research/lfads/ @jazcollins @susillo
 /research/lm_1b/ @oriolvinyals @panyx0718

official/utils/testing/scripts/presubmit.sh

@@ -53,12 +53,14 @@ py_test() {
 
 py2_test() {
   local PY_BINARY=$(which python2)
-  return $(py_test "${PY_BINARY}")
+  py_test "$PY_BINARY"
+  return $?
 }
 
 py3_test() {
   local PY_BINARY=$(which python3)
-  return $(py_test "${PY_BINARY}")
+  py_test "$PY_BINARY"
+  return $?
 }
 
 test_result=0

research/README.md

@@ -36,6 +36,8 @@ installation](https://www.tensorflow.org/install).
 -   [inception](inception): deep convolutional networks for computer vision.
 -   [learning_to_remember_rare_events](learning_to_remember_rare_events): a
     large-scale life-long memory module for use in deep learning.
+-   [learning_unsupervised_learning](learning_unsupervised_learning): a
+    meta-learned unsupervised learning update rule.
 -   [lexnet_nc](lexnet_nc): a distributed model for noun compound relationship
     classification.
 -   [lfads](lfads): sequential variational autoencoder for analyzing

research/compression/image_encoder/decoder.py

@@ -71,7 +71,7 @@ def main(_):
     return
 
   contents = ''
-  with tf.gfile.FastGFile(FLAGS.input_codes, 'r') as code_file:
+  with tf.gfile.FastGFile(FLAGS.input_codes, 'rb') as code_file:
     contents = code_file.read()
     loaded_codes = np.load(io.BytesIO(contents))
     assert ['codes', 'shape'] not in loaded_codes.files

research/compression/image_encoder/encoder.py

@@ -59,7 +59,7 @@ def main(_):
     print('\n--iteration must be between 0 and 15 inclusive.\n')
     return
 
-  with tf.gfile.FastGFile(FLAGS.input_image) as input_image:
+  with tf.gfile.FastGFile(FLAGS.input_image, 'rb') as input_image:
     input_image_str = input_image.read()
 
   with tf.Graph().as_default() as graph:

research/compression/image_encoder/msssim.py

@@ -199,9 +199,9 @@ def main(_):
     return
 
   with tf.gfile.FastGFile(FLAGS.original_image) as image_file:
-    img1_str = image_file.read()
+    img1_str = image_file.read('rb')
   with tf.gfile.FastGFile(FLAGS.compared_image) as image_file:
-    img2_str = image_file.read()
+    img2_str = image_file.read('rb')
 
   input_img = tf.placeholder(tf.string)
   decoded_image = tf.expand_dims(tf.image.decode_png(input_img, channels=3), 0)

research/learning_unsupervised_learning/.gitignore

+*.pyc

research/learning_unsupervised_learning/README.md

+# Learning Unsupervised Learning Rules
+This repository contains code and weights for the learned update rule
+presented in "Learning Unsupervised Learning Rules." At this time, this
+code can not meta-train the update rule.
+
+
+### Structure
+`run_eval.py` contains the main training loop. This constructs an op
+that runs one iteration of the learned update rule and assigns the
+results to variables. Additionally, it loads the weights from our
+pre-trained model.
+
+The base model and the update rule architecture definition can be found in
+`architectures/more_local_weight_update.py`. For a complete description
+of the model, see our [paper](https://arxiv.org/abs/1804.00222).
+
+### Dependencies
+[absl]([https://github.com/abseil/abseil-py), [tensorflow](https://tensorflow.org), [sonnet](https://github.com/deepmind/sonnet)
+
+### Usage
+
+First, download the [pre-trained optimizer model weights](https://storage.googleapis.com/learning_unsupervised_learning/200_tf_graph.zip) and extract it.
+
+```bash
+# move to the folder above this folder
+cd path_to/research/learning_unsupervised_learning/../
+
+# launch the eval script
+python -m learning_unsupervised_learning.run_eval \
+--train_log_dir="/tmp/learning_unsupervised_learning" \
+--checkpoint_dir="/path/to/downloaded/model/tf_graph_data.ckpt"
+```
+
+### Contact
+Luke Metz, Niru Maheswaranathan, Github: @lukemetz, @nirum. Email: {lmetz, nirum}@google.com
+
+

research/learning_unsupervised_learning/architectures/__init__.py

+# Copyright 2018 Google, Inc. 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.
+# ==============================================================================
+
+
+import more_local_weight_update

research/learning_unsupervised_learning/architectures/common.py

+# Copyright 2018 Google, Inc. 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.
+# ==============================================================================
+
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import sonnet as snt
+import tensorflow as tf
+import numpy as np
+import collections
+from learning_unsupervised_learning import utils
+
+from tensorflow.python.util import nest
+
+from learning_unsupervised_learning import variable_replace
+
+
+class LinearBatchNorm(snt.AbstractModule):
+  """Module that does a Linear layer then a BatchNorm followed by an activation fn"""
+  def __init__(self, size, activation_fn=tf.nn.relu, name="LinearBatchNorm"):
+    self.size = size
+    self.activation_fn = activation_fn
+    super(LinearBatchNorm, self).__init__(name=name)
+
+  def _build(self, x):
+    x = tf.to_float(x)
+    initializers={"w": tf.truncated_normal_initializer(stddev=0.01)}
+    lin = snt.Linear(self.size, use_bias=False, initializers=initializers)
+    z = lin(x)
+
+    scale = tf.constant(1., dtype=tf.float32)
+    offset = tf.get_variable(
+        "b",
+        shape=[1, z.shape.as_list()[1]],
+        initializer=tf.truncated_normal_initializer(stddev=0.1),
+        dtype=tf.float32
+    )
+
+    mean, var = tf.nn.moments(z, [0], keep_dims=True)
+    z = ((z - mean) * tf.rsqrt(var + 1e-6)) * scale + offset
+
+    x_p = self.activation_fn(z)
+
+    return z, x_p
+
+  # This needs to work by string name sadly due to how the variable replace
+  # works and would also work even if the custom getter approuch was used.
+  # This is verbose, but it should atleast be clear as to what is going on.
+  # TODO(lmetz) a better way to do this (the next 3 functions:
+  #    _raw_name, w(), b() )
+  def _raw_name(self, var_name):
+    """Return just the name of the variable, not the scopes."""
+    return var_name.split("/")[-1].split(":")[0]
+
+
+  @property
+  def w(self):
+    var_list = snt.get_variables_in_module(self)
+    w = [x for x in var_list if self._raw_name(x.name) == "w"]
+    assert len(w) == 1
+    return w[0]
+
+  @property
+  def b(self):
+    var_list = snt.get_variables_in_module(self)
+    b = [x for x in var_list if self._raw_name(x.name) == "b"]
+    assert len(b) == 1
+    return b[0]
+
+
+
+class Linear(snt.AbstractModule):
+  def __init__(self, size, use_bias=True, init_const_mag=True):
+    self.size = size
+    self.use_bias = use_bias
+    self.init_const_mag = init_const_mag
+    super(Linear, self).__init__(name="commonLinear")
+
+  def _build(self, x):
+    if self.init_const_mag:
+      initializers={"w": tf.truncated_normal_initializer(stddev=0.01)}
+    else:
+      initializers={}
+    lin = snt.Linear(self.size, use_bias=self.use_bias, initializers=initializers)
+    z = lin(x)
+    return z
+
+  # This needs to work by string name sadly due to how the variable replace
+  # works and would also work even if the custom getter approuch was used.
+  # This is verbose, but it should atleast be clear as to what is going on.
+  # TODO(lmetz) a better way to do this (the next 3 functions:
+  #    _raw_name, w(), b() )
+  def _raw_name(self, var_name):
+    """Return just the name of the variable, not the scopes."""
+    return var_name.split("/")[-1].split(":")[0]
+
+  @property
+  def w(self):
+    var_list = snt.get_variables_in_module(self)
+    if self.use_bias:
+      assert len(var_list) == 2, "Found not 2 but %d" % len(var_list)
+    else:
+      assert len(var_list) == 1, "Found not 1 but %d" % len(var_list)
+    w = [x for x in var_list if self._raw_name(x.name) == "w"]
+    assert len(w) == 1
+    return w[0]
+
+  @property
+  def b(self):
+    var_list = snt.get_variables_in_module(self)
+    assert len(var_list) == 2, "Found not 2 but %d" % len(var_list)
+    b = [x for x in var_list if self._raw_name(x.name) == "b"]
+    assert len(b) == 1
+    return b[0]
+
+
+def transformer_at_state(base_model, new_variables):
+  """Get the base_model that has been transformed to use the variables
+  in final_state.
+  Args:
+    base_model: snt.Module
+      Goes from batch to features
+    new_variables: list
+      New list of variables to use
+  Returns:
+    func: callable of same api as base_model.
+  """
+  assert not variable_replace.in_variable_replace_scope()
+
+  def _feature_transformer(input_data):
+    """Feature transformer at the end of training."""
+    initial_variables = base_model.get_variables()
+    replacement = collections.OrderedDict(
+        utils.eqzip(initial_variables, new_variables))
+    with variable_replace.variable_replace(replacement):
+      features = base_model(input_data)
+    return features
+
+  return _feature_transformer

research/learning_unsupervised_learning/datasets/__init__.py

+# Copyright 2018 Google, Inc. 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.
+# ==============================================================================
+
+import mnist

research/learning_unsupervised_learning/datasets/common.py

+# Copyright 2018 Google, Inc. 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.
+# ==============================================================================
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+
+import tensorflow as tf
+import numpy as np
+
+ImageLabelOnehot = collections.namedtuple('ImageLabelOnehot',
+                                          ['image', 'label', 'label_onehot'])
+ImageLabelOnehotRegression = collections.namedtuple(
+    "ImageLabelOnehotRegression",
+    ["image", "label", "label_onehot", "regression_target"])

research/learning_unsupervised_learning/datasets/mnist.py

+# Copyright 2018 Google, Inc. 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.
+# ==============================================================================
+
+
+import sonnet as snt
+import tensorflow as tf
+from tensorflow.python.keras.datasets import mnist
+from learning_unsupervised_learning.datasets import common
+
+class Mnist(snt.AbstractModule):
+  def __init__(self, device, batch_size=128, name="Mnist"):
+    self.device = device
+    self.batch_size = batch_size
+
+    self._make_dataset()
+    self.iterator = None
+
+    super(Mnist, self).__init__(name=name)
+
+  def _make_dataset(self):
+    (x_train, y_train), (x_test, y_test) = mnist.load_data()
+
+    x_train = x_train.reshape(60000, 784)
+    x_test = x_test.reshape(10000, 784)
+
+    dataset = tf.data.Dataset.from_tensor_slices((x_train, y_train))
+    dataset = dataset.repeat()
+    dataset = dataset.shuffle(self.batch_size * 3)
+    dataset = dataset.batch(self.batch_size)
+    def _map_fn(image, label):
+      image = tf.to_float(image) / 255.
+      label.set_shape([self.batch_size])
+      label = tf.cast(label, dtype=tf.int32)
+      label_onehot = tf.one_hot(label, 10)
+      image = tf.reshape(image, [self.batch_size, 28, 28, 1])
+      return common.ImageLabelOnehot(
+          image=image, label=label, label_onehot=label_onehot)
+
+    self.dataset = dataset.map(_map_fn)
+
+  def _build(self):
+    if self.iterator is None:
+      self.iterator = self.dataset.make_one_shot_iterator()
+    batch = self.iterator.get_next()
+    [b.set_shape([self.batch_size] + b.shape.as_list()[1:]) for b in batch]
+    return batch
+
+
+class TinyMnist(Mnist):
+  def __init__(self, *args, **kwargs):
+    kwargs.setdefault("name", "TinyMnist")
+    super(TinyMnist, self).__init__(*args, **kwargs)
+
+  def _make_dataset(self):
+    super(TinyMnist, self)._make_dataset()
+
+    def _map_fn(batch):
+      new_img = tf.image.resize_images(batch.image, [14, 14])
+      return common.ImageLabelOnehot(
+          image=new_img, label=batch.label, label_onehot=batch.label_onehot)
+
+    self.dataset = self.dataset.map(_map_fn)

research/learning_unsupervised_learning/evaluation.py

+# Copyright 2018 Google, Inc. 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.
+# ==============================================================================
+
+
+"""Evaluation job.
+
+This sits on the side and performs evaluation on a saved model.
+This is a separate process for ease of use and stability of numbers.
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorflow as tf
+
+from learning_unsupervised_learning import utils
+
+
+def construct_evaluation_graph(theta_process_fn=None,
+                               w_learner_fn=None,
+                               dataset_fn=None,
+                               meta_objectives=None,
+                              ):
+  """Construct the evaluation graph.
+  """
+  if meta_objectives is None:
+    meta_objectives = []
+
+  tf.train.create_global_step()
+
+  local_device = ""
+  remote_device = ""
+
+  meta_opt = theta_process_fn(
+      remote_device=remote_device, local_device=local_device)
+
+  base_model = w_learner_fn(
+      remote_device=remote_device, local_device=local_device)
+
+  train_dataset = dataset_fn(device=local_device)
+
+  # construct variables
+  x, outputs = base_model(train_dataset())
+  initial_state = base_model.initial_state(meta_opt, max_steps=10)
+  next_state = base_model.compute_next_state(outputs, meta_opt, initial_state)
+  with utils.state_barrier_context(next_state):
+    train_one_step_op = meta_opt.assign_state(base_model, next_state)
+
+  meta_objs = []
+  for meta_obj_fn in meta_objectives:
+    meta_obj = meta_obj_fn(local_device="", remote_device="")
+    meta_objs.append(meta_obj)
+    J = meta_obj(train_dataset, lambda x: base_model(x)[0])
+    tf.summary.scalar(str(meta_obj.__class__.__name__)+"_J", tf.reduce_mean(J))
+
+  # TODO(lmetz) this is kinda error prone.
+  # We should share the construction of the global variables across train and
+  # make sure both sets of savable variables are the same
+  checkpoint_vars = meta_opt.remote_variables() + [tf.train.get_global_step()]
+  for meta_obj in meta_objs:
+    checkpoint_vars.extend(meta_obj.remote_variables())
+
+  return checkpoint_vars, train_one_step_op, (base_model, train_dataset)

research/learning_unsupervised_learning/meta_objective/__init__.py

+# Copyright 2018 Google, Inc. 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.
+# ==============================================================================
+
+
+import sklearn
+import linear_regression

research/learning_unsupervised_learning/meta_objective/linear_regression.py

+# Copyright 2018 Google, Inc. 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.
+# ==============================================================================
+
+
+
+"""Closed form linear regression.
+
+Can be differentiated through.
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import numpy as np
+import sonnet as snt
+import tensorflow as tf
+
+from learning_unsupervised_learning import utils
+from learning_unsupervised_learning import variable_replace
+
+
+def solve_ridge(x, y, ridge_factor):
+  with tf.name_scope("solve_ridge"):
+    # Added a column of ones to the end of the feature matrix for bias
+    A = tf.concat([x, tf.ones((x.shape.as_list()[0], 1))], axis=1)
+
+    # Analytic solution for the ridge regression loss
+    inv_target = tf.matmul(A, A, transpose_a=True)
+    np_diag_penalty = ridge_factor * np.ones(
+        A.shape.as_list()[1], dtype="float32")
+    # Remove penalty on bias component of weights
+    np_diag_penalty[-1] = 0.
+    diag_penalty = tf.constant(np_diag_penalty)
+    inv_target += tf.diag(diag_penalty)
+
+    inv = tf.matrix_inverse(inv_target)
+    w = tf.matmul(inv, tf.matmul(A, y, transpose_a=True))
+    return w
+
+
+class LinearRegressionMetaObjective(snt.AbstractModule):
+  """A meta objective based on training Ridge Regression with analytic solution.
+
+  This is used to evaluate the performance of a given feature set trained in
+  some other manner.
+  """
+
+  def __init__(self,
+               local_device=None,
+               remote_device=None,
+               zero_one_labels=True,
+               normalize_y_hat=True,
+               normalize_act=False,
+               averages=1,
+               ridge_factor=0.1,
+               center_y=True,
+               hinge_loss=False,
+               samples_per_class=10,
+               test_train_scalar=1.0,
+              ):
+    self._local_device = local_device
+    self._remote_device = remote_device
+    self.zero_one_labels = zero_one_labels
+    self.normalize_y_hat = normalize_y_hat
+    self.normalize_act = normalize_act
+    self.ridge_factor = ridge_factor
+    self.averages = averages
+    self.samples_per_class = samples_per_class
+    self.center_y=center_y
+    self.test_train_scalar=test_train_scalar
+    self.hinge_loss = hinge_loss
+
+    self.dataset_map = {}
+
+    super(LinearRegressionMetaObjective,
+          self).__init__(name="LinearRegressionMetaObjective")
+
+  def _build(self, dataset, feature_transformer):
+    if self.samples_per_class is not None:
+      if dataset not in self.dataset_map:
+        # datasets are outside of frames from while loops
+        with tf.control_dependencies(None):
+          self.dataset_map[dataset] = utils.sample_n_per_class(
+              dataset, self.samples_per_class)
+
+      dataset = self.dataset_map[dataset]
+
+    stats = collections.defaultdict(list)
+    losses = []
+    # TODO(lmetz) move this to ingraph control flow?
+    for _ in xrange(self.averages):
+      loss, stat = self._build_once(dataset, feature_transformer)
+      losses.append(loss)
+      for k, v in stat.items():
+        stats[k].append(v)
+    stats = {k: tf.add_n(v) / float(len(v)) for k, v in stats.items()}
+
+    summary_updates = []
+    for k, v in stats.items():
+      tf.summary.scalar(k, v)
+
+    with tf.control_dependencies(summary_updates):
+      return tf.add_n(losses) / float(len(losses))
+
+  def _build_once(self, dataset, feature_transformer):
+    with tf.device(self._local_device):
+      batch = dataset()
+      num_classes = batch.label_onehot.shape.as_list()[1]
+
+      regression_mod = snt.Linear(num_classes)
+
+      if self.normalize_act:
+
+        def normalize_transformer(x):
+          unnorm_x = feature_transformer(x)
+          return tf.nn.l2_normalize(unnorm_x, 0)
+
+        feature_transformer_wrap = normalize_transformer
+      else:
+        feature_transformer_wrap = feature_transformer
+
+      # construct the variables of the right shape in the sonnet module by
+      # calling a forward pass through the regressor.
+      with utils.assert_no_new_variables():
+        dummy_features = feature_transformer_wrap(batch)
+      regression_mod(dummy_features)
+      reg_w = regression_mod.w
+      reg_b = regression_mod.b
+
+      batch_test = dataset()
+      all_batch = utils.structure_map_multi(lambda x: tf.concat(x, 0), [batch, batch_test])
+      #all_batch = tf.concat([batch, batch_test], 0)
+      # Grab a new batch of data from the dataset.
+      features = feature_transformer_wrap(all_batch)
+      features, features_test = utils.structure_map_split(lambda x: tf.split(x, 2, axis=0), features)
+
+      def center_y(y):
+        y -= tf.reduce_mean(y)
+        y *= tf.rsqrt(tf.reduce_mean(tf.reduce_sum(y**2, axis=[1], keep_dims=True)))
+        return y
+      def get_y_vec(batch):
+        y_pieces = []
+        if hasattr(batch, "label_onehot"):
+          if self.zero_one_labels:
+            y_pieces += [batch.label_onehot]
+          else:
+            y_pieces += [2. * batch.label_onehot - 1.]
+        if hasattr(batch, "regression_target"):
+          y_pieces += [batch.regression_target]
+        y = tf.concat(y_pieces, 1)
+        if self.center_y:
+          y = center_y(y)
+        return y
+
+      y_train = get_y_vec(batch)
+
+      w = solve_ridge(features, y_train, self.ridge_factor)
+
+      # Generate features from another batch to evaluate loss on the validation
+      # set. This provide a less overfit signal to the learned optimizer.
+      y_test = get_y_vec(batch_test)
+
+      def compute_logit(features):
+        # We have updated the classifier mod in previous steps, we need to
+        # substitute out those variables to get new values.
+        replacement = collections.OrderedDict([(reg_w, w[:-1]), (reg_b, w[-1])])
+        with variable_replace.variable_replace(replacement):
+          logits = regression_mod(features)
+
+        return logits
+
+      batch_size = y_train.shape.as_list()[0]
+
+      logit_train = compute_logit(features)
+      logit_test_unnorm = compute_logit(features_test)
+      if self.normalize_y_hat:
+        logit_test = logit_test_unnorm / tf.sqrt(
+            tf.reduce_sum(logit_test_unnorm**2, axis=[1], keep_dims=True))
+      else:
+        logit_test = logit_test_unnorm
+
+      stats = {}
+
+      if self.hinge_loss:
+        # slightly closer to the true classification loss
+        # any distance smaller than 1 is guaranteed to map to the correct class
+        mse_test = tf.reduce_sum(tf.nn.relu(tf.reduce_sum(tf.square(logit_test - y_test), axis=1)-1.)) / batch_size
+      else:
+        mse_test = tf.reduce_sum(tf.square(logit_test - y_test)) / batch_size
+
+      stats["mse_test"] = mse_test
+
+      mse_train = tf.reduce_sum(tf.square(logit_train - y_train)) / batch_size
+      stats["mse_train"] = mse_train
+
+      is_correct_test = tf.equal(tf.argmax(logit_test, 1), tf.argmax(y_test, 1))
+      accuracy_test = tf.reduce_mean(tf.cast(is_correct_test, tf.float32))
+      stats["accuracy_test"] = accuracy_test
+
+      def test_confusion_fn():
+        test_confusion = tf.confusion_matrix(tf.argmax(y_test, 1), tf.argmax(logit_test, 1))
+        test_confusion = tf.to_float(test_confusion) / tf.constant((logit_test.shape.as_list()[0] / float(logit_test.shape.as_list()[1])), dtype=tf.float32)
+        test_confusion = tf.expand_dims(tf.expand_dims(test_confusion, 0), 3)
+        return test_confusion
+      tf.summary.image("test_confusion", test_confusion_fn())
+
+      def train_confusion_fn():
+        train_confusion = tf.confusion_matrix(tf.argmax(y_train, 1), tf.argmax(logit_train, 1))
+        train_confusion = tf.to_float(train_confusion) / tf.constant((logit_train.shape.as_list()[0] / float(logit_train.shape.as_list()[1])), dtype=tf.float32)
+        train_confusion = tf.expand_dims(tf.expand_dims(train_confusion, 0), 3)
+        return train_confusion
+      tf.summary.image("train_confusion", train_confusion_fn())
+
+      is_correct = tf.equal(tf.argmax(logit_train, 1), tf.argmax(y_train, 1))
+      accuracy_train = tf.reduce_mean(tf.cast(is_correct, tf.float32))
+      stats["accuracy_train"] = accuracy_train
+
+      reg = self.ridge_factor * tf.reduce_sum(tf.square(w[:-1])) / batch_size
+      stats["ridge_component"] = reg
+
+      stats["total_loss"] = mse_test + reg
+
+      loss_to_train_at = (reg+ mse_test) * self.test_train_scalar + (mse_train + reg)*(1 - self.test_train_scalar)
+
+      loss_to_train_at = tf.identity(loss_to_train_at)
+
+      # Minimizing the test loss should not require regurization because the
+      # metaobjective is solved for the training loss
+      return loss_to_train_at, stats
+
+  def local_variables(self):
+    """List of variables that need to be updated for each evaluation.
+
+    These variables should not be stored on a parameter server and
+    should be reset every computation of a meta_objective loss.
+
+    Returns:
+      vars: list of tf.Variable
+    """
+    return list(
+        snt.get_variables_in_module(self, tf.GraphKeys.TRAINABLE_VARIABLES))
+
+  def remote_variables(self):
+    return []

research/learning_unsupervised_learning/meta_objective/sklearn.py

+# Copyright 2018 Google, Inc. 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.
+# ==============================================================================
+
+
+"""
+
+Can NOT be differentiated through.
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import numpy as np
+import sonnet as snt
+import tensorflow as tf
+from tensorflow.python.framework import function
+
+from learning_unsupervised_learning import utils
+
+from learning_unsupervised_learning.meta_objective import utils as meta_obj_utils
+
+from sklearn import svm
+from sklearn import linear_model
+
+
+def build_fit(device, model_fn, num_classes, probs=True):
+
+  def _py_fit_predict(trX, trY, teX):
+    assert len(np.unique(trY)) == num_classes
+    model = model_fn()
+    model.fit(trX, trY)
+    trP = model.predict(trX)
+    teP = model.predict(teX)
+    if probs:
+      teP_probs = model.predict_log_proba(teX)
+      return trP.astype(np.int64), teP.astype(np.int64), teP_probs.astype(
+          np.float32)
+    else:
+      teP = model.predict(teX)
+      return trP.astype(np.int64), teP.astype(np.int64)
+
+  def return_fn(trX, trY, teX):
+    with tf.device(device):
+      with tf.device("/cpu:0"):
+        if probs:
+          return tf.py_func(
+              _py_fit_predict,
+              [tf.identity(trX),
+               tf.identity(trY),
+               tf.identity(teX)], [tf.int64, tf.int64, tf.float32])
+        else:
+          return tf.py_func(
+              _py_fit_predict,
+              [tf.identity(trX),
+               tf.identity(trY),
+               tf.identity(teX)], [tf.int64, tf.int64])
+
+  return return_fn
+
+
+class SKLearn(meta_obj_utils.MultiTrialMetaObjective):
+
+  def __init__(
+      self,
+      local_device=None,
+      remote_device=None,
+      averages=1,
+      samples_per_class=10,
+      probs=False,
+      stddev=0.01,
+      n_samples=10,
+      name="SKLearn",
+  ):
+    self._local_device = local_device
+    self._remote_device = remote_device
+    self.name = name
+    self.probs = probs
+    self.n_samples = n_samples
+    self.stddev = stddev
+
+    super(SKLearn, self).__init__(
+        name=name, samples_per_class=samples_per_class, averages=averages)
+
+  def _get_model(self):
+    raise NotImplemented()
+
+  def _build_once(self, dataset, feature_transformer):
+    with tf.device(self._local_device):
+      tr_batch = dataset()
+      te_batch = dataset()
+      num_classes = tr_batch.label_onehot.shape.as_list()[1]
+      all_batch = utils.structure_map_multi(lambda x: tf.concat(x, 0),
+                                            [tr_batch, te_batch])
+      features = feature_transformer(all_batch)
+      trX, teX = utils.structure_map_split(lambda x: tf.split(x, 2, axis=0),
+                                           features)
+      trY = tf.to_int64(tr_batch.label)
+      trY_onehot = tf.to_int32(tr_batch.label_onehot)
+      teY = tf.to_int64(te_batch.label)
+      teY_shape = teY.shape.as_list()
+
+      def blackbox((trX, trY, teX, teY)):
+        trY = tf.to_int32(tf.rint(trY))
+        teY = tf.to_int32(tf.rint(teY))
+        tf_fn = build_fit(
+            self._local_device,
+            self._get_model,
+            num_classes=num_classes,
+            probs=self.probs)
+        if self.probs:
+          trP, teP, teP_probs = tf_fn(trX, trY, teX)
+        else:
+          trP, teP = tf_fn(trX, trY, teX)
+
+        teY.set_shape(teY_shape)
+        if self.probs:
+          onehot = tf.one_hot(teY, num_classes)
+          crossent = -tf.reduce_sum(onehot * teP_probs, [1])
+          return tf.reduce_mean(crossent)
+        else:
+          # use error rate as the loss if no surrogate is avalible.
+          return 1 - tf.reduce_mean(
+              tf.to_float(tf.equal(teY, tf.to_int32(teP))))
+
+      test_loss = blackbox((trX, tf.to_float(trY), teX, tf.to_float(teY)))
+
+      stats = {}
+
+      tf_fn = build_fit(
+          self._local_device,
+          self._get_model,
+          num_classes=num_classes,
+          probs=self.probs)
+      if self.probs:
+        trP, teP, teP_probs = tf_fn(trX, trY, teX)
+      else:
+        trP, teP = tf_fn(trX, trY, teX)
+      stats["%s/accuracy_train" % self.name] = tf.reduce_mean(
+          tf.to_float(tf.equal(tf.to_int32(trY), tf.to_int32(trP))))
+      stats["%s/accuracy_test" % self.name] = tf.reduce_mean(
+          tf.to_float(tf.equal(tf.to_int32(teY), tf.to_int32(teP))))
+      stats["%s/test_loss" % self.name] = test_loss
+      return test_loss, stats
+
+
+class LogisticRegression(SKLearn):
+
+  def __init__(self, C=1.0, name="LogisticRegression", probs=True, **kwargs):
+    self.C = C
+    super(LogisticRegression, self).__init__(name=name, probs=probs, **kwargs)
+
+  def _get_model(self):
+    return linear_model.LogisticRegression(C=self.C)

research/learning_unsupervised_learning/meta_objective/utils.py

+# Copyright 2018 Google, Inc. 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.
+# ==============================================================================
+
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import collections
+import numpy as np
+import sonnet as snt
+import tensorflow as tf
+
+from learning_unsupervised_learning import optimizers
+from learning_unsupervised_learning import utils
+from learning_unsupervised_learning import summary_utils
+from learning_unsupervised_learning import variable_replace
+
+class MultiTrialMetaObjective(snt.AbstractModule):
+  def __init__(self, samples_per_class, averages, **kwargs):
+    self.samples_per_class = samples_per_class
+    self.averages = averages
+    self.dataset_map = {}
+
+    super(MultiTrialMetaObjective,
+          self).__init__(**kwargs)
+
+  def _build(self, dataset, feature_transformer):
+    if self.samples_per_class is not None:
+      if dataset not in self.dataset_map:
+        # datasets are outside of frames from while loops
+        with tf.control_dependencies(None):
+          self.dataset_map[dataset] = utils.sample_n_per_class(
+              dataset, self.samples_per_class)
+
+      dataset = self.dataset_map[dataset]
+
+    stats = collections.defaultdict(list)
+    losses = []
+    # TODO(lmetz) move this to ingraph control flow?
+    for _ in xrange(self.averages):
+      loss, stat = self._build_once(dataset, feature_transformer)
+      losses.append(loss)
+      for k, v in stat.items():
+        stats[k].append(v)
+    stats = {k: tf.add_n(v) / float(len(v)) for k, v in stats.items()}
+
+    for k, v in stats.items():
+      tf.summary.scalar(k, v)
+
+    return tf.add_n(losses) / float(len(losses))
+
+  def local_variables(self):
+    """List of variables that need to be updated for each evaluation.
+
+    These variables should not be stored on a parameter server and
+    should be reset every computation of a meta_objective loss.
+
+    Returns:
+      vars: list of tf.Variable
+    """
+    return list(
+        snt.get_variables_in_module(self, tf.GraphKeys.TRAINABLE_VARIABLES))
+
+  def remote_variables(self):
+    return []