1. Splits train_image_classifier into library and binary rule, to simplify reuse. (#4552)

2. Flag that allows to prevent imagenet.py  from downloading label_to_names from github and/or dumping into training directory (which might be read-only)
3. Adds some comments about how decay steps are computed, since it computed differently when there are clones vs sync replicas.
4. Updates mobilenet.md to describe the training process using train_image_classifer
5. Add citation for TF-Slim model library.

PiperOrigin-RevId: 191955231

PiperOrigin-RevId: 193254125

PiperOrigin-RevId: 193371562

PiperOrigin-RevId: 194085628

PiperOrigin-RevId: 194857067

PiperOrigin-RevId: 196125653

PiperOrigin-RevId: 196589070

PiperOrigin-RevId: 199522873

PiperOrigin-RevId: 200351305

research/slim/BUILD

@@ -694,10 +694,9 @@ py_test(
     ],
 )
 
-py_binary(
-    name = "train_image_classifier",
+py_library(
+    name = "train_image_classifier_lib",
     srcs = ["train_image_classifier.py"],
-    paropts = ["--compress"],
     deps = [
         ":dataset_factory",
         ":model_deploy",
@@ -707,6 +706,15 @@ py_binary(
     ],
 )
 
+py_binary(
+    name = "train_image_classifier",
+    srcs = ["train_image_classifier.py"],
+    paropts = ["--compress"],
+    deps = [
+        ":train_image_classifier_lib",
+    ],
+)
+
 py_binary(
     name = "eval_image_classifier",
     srcs = ["eval_image_classifier.py"],

research/slim/README.md

@@ -25,6 +25,11 @@ Maintainers of TF-slim:
   github: [nathansilberman](https://github.com/nathansilberman)
 * Sergio Guadarrama, github: [sguada](https://github.com/sguada)
 
+## Citation
+"TensorFlow-Slim image classification model library"
+N. Silberman and S. Guadarrama, 2016.
+https://github.com/tensorflow/models/tree/master/research/slim
+
 ## Table of contents
 
 <a href="#Install">Installation and setup</a><br>

research/slim/datasets/imagenet.py

@@ -58,6 +58,10 @@ _ITEMS_TO_DESCRIPTIONS = {
 
 _NUM_CLASSES = 1001
 
+# If set to false, will not try to set label_to_names in dataset
+# by reading them from labels.txt or github.
+LOAD_READABLE_NAMES = True
+
 
 def create_readable_names_for_imagenet_labels():
   """Create a dict mapping label id to human readable string.
@@ -177,6 +181,7 @@ def get_split(split_name, dataset_dir, file_pattern=None, reader=None):
       keys_to_features, items_to_handlers)
 
   labels_to_names = None
+  if LOAD_READABLE_NAMES:
     if dataset_utils.has_labels(dataset_dir):
       labels_to_names = dataset_utils.read_label_file(dataset_dir)
     else:

research/slim/nets/mobilenet/README.md

@@ -47,7 +47,27 @@ are no published size numbers. We estimate it to be comparable to MobileNetV2 nu
 | [mobilenet_v2_0.35_128](https://storage.googleapis.com/mobilenet_v2/checkpoints/mobilenet_v2_0.35_128.tgz) | 20 | 1.66 | 50.8 | 75.0 | 6.9
 | [mobilenet_v2_0.35_96](https://storage.googleapis.com/mobilenet_v2/checkpoints/mobilenet_v2_0.35_96.tgz) | 11 | 1.66 | 45.5 | 70.4 | 4.5
 
+# Training
+The numbers above can be reproduced using slim's `train_image_classifier`.
+Below is the set of parameters that achieves 72.0% for full size MobileNetV2, after about 700K when trained on 8 GPU.
+If trained on a single GPU the full convergence is after 5.5M steps. Also note that learning rate and
+num_epochs_per_decay both need to be adjusted depending on how many GPUs are being
+used due to slim's internal averaging.
+
+```bash
+--model_name="mobilenet_v2"
+--learning_rate=0.045 * NUM_GPUS   #slim internally averages clones so we compensate
+--preprocessing_name="inception_v2"
+--label_smoothing=0.1
+--moving_average_decay=0.9999
+--batch_size= 96
+--num_clones = NUM_GPUS # you can use any number here between 1 and 8 depending on your hardware setup.
+--learning_rate_decay_factor=0.98
+--num_epochs_per_decay = 2.5 / NUM_GPUS # train_image_classifier does per clone epochs
+```
+
 # Example
 
+
 See this [ipython notebook](mobilenet_example.ipynb) or open and run the network directly in [Colaboratory](https://colab.research.google.com/github/tensorflow/models/blob/master/research/slim/nets/mobilenet/mobilenet_example.ipynb).
 

research/slim/nets/nasnet/nasnet_utils.py

@@ -133,8 +133,10 @@ def drop_path(net, keep_prob, is_training=True):
     noise_shape = [batch_size, 1, 1, 1]
     random_tensor = keep_prob
     random_tensor += tf.random_uniform(noise_shape, dtype=tf.float32)
-    binary_tensor = tf.floor(random_tensor)
-    net = tf.div(net, keep_prob) * binary_tensor
+    binary_tensor = tf.cast(tf.floor(random_tensor), net.dtype)
+    keep_prob_inv = tf.cast(1.0 / keep_prob, net.dtype)
+    net = net * keep_prob_inv * binary_tensor
+
   return net
 
 

research/slim/train_image_classifier.py

@@ -35,7 +35,10 @@ tf.app.flags.DEFINE_string(
     'Directory where checkpoints and event logs are written to.')
 
 tf.app.flags.DEFINE_integer('num_clones', 1,
-                            'Number of model clones to deploy.')
+                            'Number of model clones to deploy. Note For '
+                            'historical reasons loss from all clones averaged '
+                            'out and learning rate decay happen per clone '
+                            'epochs')
 
 tf.app.flags.DEFINE_boolean('clone_on_cpu', False,
                             'Use CPUs to deploy clones.')
@@ -145,7 +148,10 @@ tf.app.flags.DEFINE_float(
 
 tf.app.flags.DEFINE_float(
     'num_epochs_per_decay', 2.0,
-    'Number of epochs after which learning rate decays.')
+    'Number of epochs after which learning rate decays. Note: this flag counts '
+    'epochs per clone but aggregates per sync replicas. So 1.0 means that '
+    'each clone will go over full epoch individually, but replicas will go '
+    'once across all replicas.')
 
 tf.app.flags.DEFINE_bool(
     'sync_replicas', False,
@@ -233,8 +239,12 @@ def _configure_learning_rate(num_samples_per_epoch, global_step):
   Raises:
     ValueError: if
   """
-  decay_steps = int(num_samples_per_epoch / FLAGS.batch_size *
-                    FLAGS.num_epochs_per_decay)
+  # Note: when num_clones is > 1, this will actually have each clone to go
+  # over each epoch FLAGS.num_epochs_per_decay times. This is different
+  # behavior from sync replicas and is expected to produce different results.
+  decay_steps = int(num_samples_per_epoch * FLAGS.num_epochs_per_decay /
+                    FLAGS.batch_size)
+
   if FLAGS.sync_replicas:
     decay_steps /= FLAGS.replicas_to_aggregate
 
@@ -256,7 +266,7 @@ def _configure_learning_rate(num_samples_per_epoch, global_step):
                                      cycle=False,
                                      name='polynomial_decay_learning_rate')
   else:
-    raise ValueError('learning_rate_decay_type [%s] was not recognized',
+    raise ValueError('learning_rate_decay_type [%s] was not recognized' %
                      FLAGS.learning_rate_decay_type)
 
 
@@ -308,7 +318,7 @@ def _configure_optimizer(learning_rate):
   elif FLAGS.optimizer == 'sgd':
     optimizer = tf.train.GradientDescentOptimizer(learning_rate)
   else:
-    raise ValueError('Optimizer [%s] was not recognized', FLAGS.optimizer)
+    raise ValueError('Optimizer [%s] was not recognized' % FLAGS.optimizer)
   return optimizer