Merge branch 'panoptic-segmentation' into panoptic-deeplab-modeling

official/vision/image_classification/optimizer_factory_test.py → official/legacy/image_classification/optimizer_factory_test.py

@@ -21,8 +21,8 @@ from __future__ import print_function
 from absl.testing import parameterized
 
 import tensorflow as tf
-from official.vision.image_classification import optimizer_factory
-from official.vision.image_classification.configs import base_configs
+from official.legacy.image_classification import optimizer_factory
+from official.legacy.image_classification.configs import base_configs
 
 
 class OptimizerFactoryTest(tf.test.TestCase, parameterized.TestCase):

official/vision/image_classification/preprocessing.py → official/legacy/image_classification/preprocessing.py

@@ -17,11 +17,9 @@
 from __future__ import absolute_import
 from __future__ import division
 from __future__ import print_function
-
-import tensorflow as tf
 from typing import List, Optional, Text, Tuple
-
-from official.vision.image_classification import augment
+import tensorflow as tf
+from official.legacy.image_classification import augment
 
 
 # Calculated from the ImageNet training set
@@ -272,8 +270,11 @@ def resize_image(image_bytes: tf.Tensor,
     A tensor containing the resized image.
 
   """
+  print(height, width)
   return tf.compat.v1.image.resize(
-      image_bytes, [height, width], method=tf.image.ResizeMethod.BILINEAR,
+      image_bytes,
+      tf.convert_to_tensor([height, width]),
+      method=tf.image.ResizeMethod.BILINEAR,
       align_corners=False)
 
 

official/vision/image_classification/resnet/common.py → official/legacy/image_classification/resnet/common.py

@@ -245,7 +245,7 @@ def define_keras_flags(model=False,
   flags.DEFINE_boolean(
       name='enable_tensorboard',
       default=False,
-      help='Whether to enable Tensorboard callback.')
+      help='Whether to enable TensorBoard callback.')
   flags.DEFINE_string(
       name='profile_steps',
       default=None,

official/vision/image_classification/resnet/resnet_config.py → official/legacy/image_classification/resnet/resnet_config.py

@@ -19,9 +19,8 @@ from __future__ import division
 from __future__ import print_function
 
 import dataclasses
-
+from official.legacy.image_classification.configs import base_configs
 from official.modeling.hyperparams import base_config
-from official.vision.image_classification.configs import base_configs
 
 
 @dataclasses.dataclass
@@ -30,12 +29,14 @@ class ResNetModelConfig(base_configs.ModelConfig):
   name: str = 'ResNet'
   num_classes: int = 1000
   model_params: base_config.Config = dataclasses.field(
+      # pylint: disable=g-long-lambda
       default_factory=lambda: {
           'num_classes': 1000,
           'batch_size': None,
           'use_l2_regularizer': True,
           'rescale_inputs': False,
       })
+  # pylint: enable=g-long-lambda
   loss: base_configs.LossConfig = base_configs.LossConfig(
       name='sparse_categorical_crossentropy')
   optimizer: base_configs.OptimizerConfig = base_configs.OptimizerConfig(

official/vision/image_classification/resnet/resnet_ctl_imagenet_main.py → official/legacy/image_classification/resnet/resnet_ctl_imagenet_main.py

@@ -24,13 +24,13 @@ from absl import logging
 import orbit
 import tensorflow as tf
 from official.common import distribute_utils
+from official.legacy.image_classification.resnet import common
+from official.legacy.image_classification.resnet import imagenet_preprocessing
+from official.legacy.image_classification.resnet import resnet_runnable
 from official.modeling import performance
 from official.utils.flags import core as flags_core
 from official.utils.misc import keras_utils
 from official.utils.misc import model_helpers
-from official.vision.image_classification.resnet import common
-from official.vision.image_classification.resnet import imagenet_preprocessing
-from official.vision.image_classification.resnet import resnet_runnable
 
 flags.DEFINE_boolean(name='use_tf_function', default=True,
                      help='Wrap the train and test step inside a '

official/vision/image_classification/resnet/resnet_model.py → official/legacy/image_classification/resnet/resnet_model.py

@@ -28,7 +28,7 @@ from __future__ import division
 from __future__ import print_function
 
 import tensorflow as tf
-from official.vision.image_classification.resnet import imagenet_preprocessing
+from official.legacy.image_classification.resnet import imagenet_preprocessing
 
 layers = tf.keras.layers
 

official/vision/image_classification/resnet/resnet_runnable.py → official/legacy/image_classification/resnet/resnet_runnable.py

@@ -16,12 +16,12 @@
 
 import orbit
 import tensorflow as tf
+from official.legacy.image_classification.resnet import common
+from official.legacy.image_classification.resnet import imagenet_preprocessing
+from official.legacy.image_classification.resnet import resnet_model
 from official.modeling import grad_utils
 from official.modeling import performance
 from official.utils.flags import core as flags_core
-from official.vision.image_classification.resnet import common
-from official.vision.image_classification.resnet import imagenet_preprocessing
-from official.vision.image_classification.resnet import resnet_model
 
 
 class ResnetRunnable(orbit.StandardTrainer, orbit.StandardEvaluator):
@@ -71,15 +71,9 @@ class ResnetRunnable(orbit.StandardTrainer, orbit.StandardEvaluator):
     self.optimizer = common.get_optimizer(lr_schedule)
     # Make sure iterations variable is created inside scope.
     self.global_step = self.optimizer.iterations
-
-    use_graph_rewrite = flags_obj.fp16_implementation == 'graph_rewrite'
-    if use_graph_rewrite and not flags_obj.use_tf_function:
-      raise ValueError('--fp16_implementation=graph_rewrite requires '
-                       '--use_tf_function to be true')
     self.optimizer = performance.configure_optimizer(
         self.optimizer,
         use_float16=self.dtype == tf.float16,
-        use_graph_rewrite=use_graph_rewrite,
         loss_scale=flags_core.get_loss_scale(flags_obj, default_for_fp16=128))
 
     self.train_loss = tf.keras.metrics.Mean('train_loss', dtype=tf.float32)
@@ -105,7 +99,8 @@ class ResnetRunnable(orbit.StandardTrainer, orbit.StandardEvaluator):
         datasets_num_private_threads=self.flags_obj
         .datasets_num_private_threads,
         dtype=self.dtype,
-        drop_remainder=True)
+        drop_remainder=True,
+        training_dataset_cache=self.flags_obj.training_dataset_cache)
     orbit.StandardTrainer.__init__(
         self,
         train_dataset,

official/vision/image_classification/resnet/tfhub_export.py → official/legacy/image_classification/resnet/tfhub_export.py

@@ -26,8 +26,8 @@ from absl import flags
 
 import tensorflow as tf
 
-from official.vision.image_classification.resnet import imagenet_preprocessing
-from official.vision.image_classification.resnet import resnet_model
+from official.legacy.image_classification.resnet import imagenet_preprocessing
+from official.legacy.image_classification.resnet import resnet_model
 
 FLAGS = flags.FLAGS
 

official/nlp/transformer/README.md → official/legacy/transformer/README.md

@@ -32,7 +32,7 @@ model.
 # https://github.com/tensorflow/models/tree/master/official#requirements
 export PYTHONPATH="$PYTHONPATH:/path/to/models"
 
-cd /path/to/models/official/nlp/transformer
+cd /path/to/models/official/legacy/transformer
 
 # Export variables
 PARAM_SET=big

official/nlp/transformer/attention_layer.py → official/legacy/transformer/attention_layer.py

@@ -16,7 +16,6 @@
 import math
 
 import tensorflow as tf
-from official.nlp.modeling import layers
 
 
 class Attention(tf.keras.layers.Layer):
@@ -51,28 +50,31 @@ class Attention(tf.keras.layers.Layer):
 
     attention_initializer = _glorot_initializer(input_shape.as_list()[-1],
                                                 self.hidden_size)
-    self.query_dense_layer = layers.DenseEinsum(
-        output_shape=(self.num_heads, size_per_head),
+    self.query_dense_layer = tf.keras.layers.experimental.EinsumDense(
+        "BTE,ENH->BTNH",
+        output_shape=(None, self.num_heads, size_per_head),
         kernel_initializer=attention_initializer,
-        use_bias=False,
+        bias_axes=None,
         name="query")
-    self.key_dense_layer = layers.DenseEinsum(
-        output_shape=(self.num_heads, size_per_head),
+    self.key_dense_layer = tf.keras.layers.experimental.EinsumDense(
+        "BTE,ENH->BTNH",
+        output_shape=(None, self.num_heads, size_per_head),
         kernel_initializer=attention_initializer,
-        use_bias=False,
+        bias_axes=None,
         name="key")
-    self.value_dense_layer = layers.DenseEinsum(
-        output_shape=(self.num_heads, size_per_head),
+    self.value_dense_layer = tf.keras.layers.experimental.EinsumDense(
+        "BTE,ENH->BTNH",
+        output_shape=(None, self.num_heads, size_per_head),
         kernel_initializer=attention_initializer,
-        use_bias=False,
+        bias_axes=None,
         name="value")
 
     output_initializer = _glorot_initializer(self.hidden_size, self.hidden_size)
-    self.output_dense_layer = layers.DenseEinsum(
-        output_shape=self.hidden_size,
-        num_summed_dimensions=2,
+    self.output_dense_layer = tf.keras.layers.experimental.EinsumDense(
+        "BTNH,NHE->BTE",
+        output_shape=(None, self.hidden_size),
         kernel_initializer=output_initializer,
-        use_bias=False,
+        bias_axes=None,
         name="output_transform")
     super(Attention, self).build(input_shape)
 

official/nlp/transformer/compute_bleu.py → official/legacy/transformer/compute_bleu.py

@@ -24,12 +24,13 @@ import unicodedata
 
 from absl import app
 from absl import flags
+from absl import logging
 import six
 from six.moves import range
 import tensorflow as tf
 
-from official.nlp.transformer.utils import metrics
-from official.nlp.transformer.utils import tokenizer
+from official.legacy.transformer.utils import metrics
+from official.legacy.transformer.utils import tokenizer
 from official.utils.flags import core as flags_core
 
 
@@ -109,11 +110,11 @@ def bleu_on_list(ref_lines, hyp_lines, case_sensitive=False):
 def main(unused_argv):
   if FLAGS.bleu_variant in ("both", "uncased"):
     score = bleu_wrapper(FLAGS.reference, FLAGS.translation, False)
-    tf.logging.info("Case-insensitive results: %f" % score)
+    logging.info("Case-insensitive results: %f", score)
 
   if FLAGS.bleu_variant in ("both", "cased"):
     score = bleu_wrapper(FLAGS.reference, FLAGS.translation, True)
-    tf.logging.info("Case-sensitive results: %f" % score)
+    logging.info("Case-sensitive results: %f", score)
 
 
 def define_compute_bleu_flags():
@@ -142,7 +143,6 @@ def define_compute_bleu_flags():
 
 
 if __name__ == "__main__":
-  tf.logging.set_verbosity(tf.logging.INFO)
   define_compute_bleu_flags()
   FLAGS = flags.FLAGS
   app.run(main)

official/nlp/transformer/compute_bleu_test.py → official/legacy/transformer/compute_bleu_test.py

@@ -18,7 +18,7 @@ import tempfile
 
 import tensorflow as tf
 
-from official.nlp.transformer import compute_bleu
+from official.legacy.transformer import compute_bleu
 
 
 class ComputeBleuTest(tf.test.TestCase):

official/nlp/transformer/data_download.py → official/legacy/transformer/data_download.py

@@ -29,7 +29,7 @@ from six.moves import urllib
 from six.moves import zip
 import tensorflow.compat.v1 as tf
 
-from official.nlp.transformer.utils import tokenizer
+from official.legacy.transformer.utils import tokenizer
 from official.utils.flags import core as flags_core
 # pylint: enable=g-bad-import-order