Merge pull request #49 from otiliastr:master

PiperOrigin-RevId: 301199341

Author: tensorflow-copybara

research/gam/README.md

@@ -34,23 +34,48 @@ folder on a strict "as is" basis, without warranties or conditions of any kind.
 Also, these implementations may not be compatible with certain TensorFlow
 versions (such as 2.0 or above) or Python versions.
 
+More details can be found in our
+[paper](https://papers.nips.cc/paper/9076-graph-agreement-models-for-semi-supervised-learning.pdf),
+[supplementary material](https://papers.nips.cc/paper/9076-graph-agreement-models-for-semi-supervised-learning-supplemental.zip),
+[slides](https://drive.google.com/open?id=1tWEMoyrbLnzfSfTfYFi9eWgZWaPKF3Uu) or
+[poster](https://drive.google.com/file/d/1BZNR4B-xM41hdLLqx4mLsQ4KKJOhjgqV/view).
+
 ## How to run
 
 To run GAM on a graph-based dataset (e.g., Cora, Citeseer, Pubmed), from this
-folder run: `bash python3.7 -m gam.experiments.run_train_gam_graph
+folder run: `$ python3.7 -m gam.experiments.run_train_gam_graph
 --data_path=<path_to_data>`
 
-To run GAM on datasets without a graph (e.g., CIFAR10), from this folder run:
-`bash python3.7 -m gam.experiments.run_train_gam`
+To run GAM on datasets without a graph (e.g., CIFAR10), from this folder run: `$
+python3.7 -m gam.experiments.run_train_gam`
+
+We recommend running on a GPU. With CUDA, this can be done by prepending
+`CUDA_VISIBLE_DEVICES=<your-gpu-number>` in front of the run command.
 
 For running on different datasets and configuration, please check the command
-line flags in each of the run scripts.
+line flags in each of the run scripts. The configurations used in our paper can
+be found in the file `run_configs.txt`.
+
+## Visualizing the results.
+
+To visualize the results in Tensorboard, use the following command, adjusting
+the dataset name accordingly: `$ tensorboard --logdir=outputs/summaries/cora`
+
+An example of such visualization for Cora with GCN + GAM model on the Pubmed
+dataset is the following:
+![Tensorboard plot](gam_gcn_pubmed.png?raw=true "GCN + GAM on Pubmed")
+
+Similarly, we can run with multiple different parameter configurations and plot
+the results together for comparison. An example showing the accuracy per
+co-train iteration of a GCN + GAM model on the Cora dataset for 3 runs with 3
+different random seeds is the following:
+![Tensorboard plot](gam_gcn_cora_multiple_seeds.png?raw=true "GCN + GAM on Cora")
 
 ## References
 
-[[1] O. Stretcu, K. Viswanathan, D. Movshovitz-Attias, E.A. Platanios, A.
-Tomkins, S. Ravi. "Graph Agreement Models for Semi-Supervised Learning." NeurIPS
-2019](https://nips.cc/Conferences/2019/Schedule?showEvent=13925)
+[[1] O. Stretcu, K. Viswanathan, D. Movshovitz-Attias, E.A. Platanios, S. Ravi,
+A. Tomkins. "Graph Agreement Models for Semi-Supervised Learning." NeurIPS
+2019](https://papers.nips.cc/paper/9076-graph-agreement-models-for-semi-supervised-learning)
 
 [[2] T. Bui, S. Ravi and V. Ramavajjala. "Neural Graph Learning: Training Neural
 Networks Using Graphs." WSDM 2018](https://research.google/pubs/pub46568.pdf)

research/gam/gam/experiments/helper.py

@@ -62,7 +62,7 @@ def get_model_cls(model_name,
         num_classes=data.num_classes,
         lrelu_leakiness=0.1,
         horizontal_flip=dataset_name in ('cifar10',),
-        random_translation=True,
+        random_translation=False,
         gaussian_noise=dataset_name not in ('svhn', 'svhn_cropped'),
         width=2,
         num_residual_units=4,
@@ -125,7 +125,7 @@ def get_model_agr(model_name,
         num_classes=1,
         lrelu_leakiness=0.1,
         horizontal_flip=dataset_name in ('cifar10',),
-        random_translation=True,
+        random_translation=False,
         gaussian_noise=dataset_name not in ('svhn', 'svhn_cropped'),
         width=2,
         num_residual_units=4,

research/gam/gam/models/models_base.py

@@ -142,7 +142,7 @@ class attribute. The valid options are:
     else:
       raise NotImplementedError()
 
-  def _project(self, inputs, reuse=tf.AUTO_REUSE):
+  def _project(self, inputs, reuse=tf.compat.v1.AUTO_REUSE):
     """Projects the provided inputs using a multilayer perceptron.
 
     Arguments:

research/gam/gam/models/wide_resnet.py

@@ -22,7 +22,6 @@
 
 import numpy as np
 import tensorflow as tf
-import tensorflow_addons as tfa
 
 
 def fast_flip(images, is_training):
@@ -38,27 +37,6 @@ def func(inp):
   return tf.cond(is_training, lambda: func(images), lambda: images)
 
 
-def jitter(input_data, is_training):
-  """Applies random noise to input data when training."""
-
-  def func(inp):
-    """Wrap functionality in a subfunction."""
-    bsz = tf.shape(inp)[0]
-    inp = tf.pad(inp, [[0, 0], [2, 2], [2, 2], [0, 0]], mode="REFLECT")
-    base = tf.constant([1, 0, 0, 0, 1, 0, 0, 0], shape=[1, 8], dtype=tf.float32)
-    base = tf.tile(base, [bsz, 1])
-    mask = tf.constant([0, 0, 1, 0, 0, 1, 0, 0], shape=[1, 8], dtype=tf.float32)
-    mask = tf.tile(mask, [bsz, 1])
-    jit = tf.random_uniform([bsz, 8], minval=-2, maxval=3, dtype=tf.int32)
-    jit = tf.cast(jit, tf.float32)
-    xforms = base + jit * mask
-    processed_data = tfa.image.transform(images=inp, transforms=xforms)
-    cropped_data = processed_data[:, 2:-2, 2:-2, :]
-    return cropped_data
-
-  return tf.cond(is_training, lambda: func(input_data), lambda: input_data)
-
-
 class WideResnet(Model):
   """Resnet implementation from `Realistic Evaluation of Deep SSL Algorithms`.
 
@@ -289,7 +267,7 @@ def _residual(x,
     if self.horizontal_flip:
       x = fast_flip(x, is_training=is_train)
     if self.random_translation:
-      x = jitter(x, is_training=is_train)
+      raise NotImplementedError("Random translations are not implemented yet.")
     if self.gaussian_noise:
       x = tf.cond(is_train, lambda: x + tf.random_normal(tf.shape(x)) * 0.15,
                   lambda: x)

research/gam/gam/trainer/trainer_classification_gcn.py

@@ -21,8 +21,8 @@
 import logging
 import os
 
-from .adversarial_dense import entropy_y_x
-from .adversarial_dense import get_loss_vat
+from .adversarial_sparse import entropy_y_x
+from .adversarial_sparse import get_loss_vat
 import numpy as np
 import tensorflow as tf
 from .trainer_base import batch_iterator