Sync.

Author: mingxingtan

efficientdet/Det-AdvProp.md

@@ -1,6 +1,5 @@
 # Det-AdvProp
-[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.sandbox.google.com/github/google/automl/blob/master/efficientdet
-/det_advprop_tutorial.ipynb)
+[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.sandbox.google.com/github/google/automl/blob/master/efficientdet/det_advprop_tutorial.ipynb)
 
 [1] Xiangning Chen, Cihang Xie, Mingxing Tan, Li Zhang, Cho-Jui Hsieh, Boqing
 Gong. CVPR 2021. Arxiv link: https://arxiv.org/abs/2103.13886
@@ -31,7 +30,7 @@ EfficientDet-D5 + Det-AdvProp + AA ([ckpt](https://storage.googleapis.com/cloud-
 
 <sup>Unlike the vanilla EfficientDet that scales the image with mean and std,
 here we scale the input to the range of [-1, 1] to make it easier for performing
-adversarial attack. Please see [this Colab](??) for reproducing the
+adversarial attack. Please see [this Colab](https://github.com/google/automl/blob/master/efficientdet/det_advprop_tutorial.ipynb) for reproducing the
 results.</sup>
 
 ## 2. Robust Against Common Corruptions

efficientdet/det_advprop_tutorial.ipynb

@@ -668,6 +668,7 @@ def __init__(self,
       is_training_bn: True if we train the BatchNorm.
       act_type: String of the activation used.
       strategy: string to specify training strategy for TPU/GPU/CPU.
+      name: string of name.
       **kwargs: other parameters.
     """
     super().__init__(name=name, **kwargs)

efficientdet/keras/efficientdet_keras.py

@@ -80,6 +80,7 @@ def __init__(self, tflite_model_path, only_network=False):
   def run(self, image):
     """Runs inference with Lite model."""
     interpreter = self.interpreter
+    signature_fn = interpreter.get_signature_runner()
     input_details = self.input_details
     output_details = self.output_details
 
@@ -88,12 +89,12 @@ def run(self, image):
       scale, zero_point = input_detail['quantization']
       image = image / scale + zero_point
       image = np.array(image, dtype=input_detail['dtype'])
-    interpreter.set_tensor(input_detail['index'], image)
-    interpreter.invoke()
+
+    output = signature_fn(images=image)
 
     def get_output(idx):
       output_detail = output_details[idx]
-      output_tensor = interpreter.get_tensor(output_detail['index'])
+      output_tensor = output[f'output_{idx}']
       if output_detail['quantization'] != (DEFAULT_SCALE, DEFAULT_ZERO_POINT):
         # Dequantize the output
         scale, zero_point = output_detail['quantization']
@@ -105,14 +106,14 @@ def get_output(idx):
     if not self.only_network:
       # TFLite model with post-processing.
       # Four Outputs:
-      #   detection_boxes: a float32 tensor of shape [1, num_boxes, 4] with box
-      #     locations
-      #   detection_classes: a float32 tensor of shape [1, num_boxes]
-      #     with class indices
-      #   detection_scores: a float32 tensor of shape [1, num_boxes]
-      #     with class scores
       #   num_boxes: a float32 tensor of size 1 containing the number of
       #     detected boxes
+      #   detection_scores: a float32 tensor of shape [1, num_boxes]
+      #     with class scores
+      #   detection_classes: a float32 tensor of shape [1, num_boxes]
+      #     with class indices
+      #   detection_boxes: a float32 tensor of shape [1, num_boxes, 4] with box
+      #     locations
       return [get_output(i) for i in range(output_size)]
     else:
       # TFLite model only contains network without post-processing.

efficientdet/keras/eval_tflite.py

@@ -90,7 +90,8 @@ def main(_):
       validation_data=test_dataset,
       callbacks=[])
 
-  model.save_weights('./testdata/segmentation')
+  model.save_weights(
+      './testdata/segmentation')
 
   print(create_mask(model(tf.ones((1, 512, 512, 3)), False)))
 

efficientdet/keras/segmentation.py

@@ -15,7 +15,7 @@
 """Common keras utils."""
 import collections
 
-from typing import Text
+from typing import Optional, Text
 from absl import logging
 import tensorflow as tf
 import utils
@@ -32,7 +32,7 @@ def build_batch_norm(is_training_bn: bool,
                      data_format: Text = 'channels_last',
                      momentum: float = 0.99,
                      epsilon: float = 1e-3,
-                     strategy: Text = None,
+                     strategy: Optional[Text] = None,
                      name: Text = 'tpu_batch_normalization'):
   """Build a batch normalization layer.
 

efficientdet/keras/util_keras.py

@@ -429,7 +429,7 @@ def convert_tr(self, graph_def, fetches):
     """Convert to TensorRT."""
     from tensorflow.python.compiler.tensorrt import trt  # pylint: disable=g-direct-tensorflow-import,g-import-not-at-top
     converter = trt.TrtGraphConverter(
-        nodes_blacklist=[t.split(':')[0] for t in fetches],
+        nodes_denylist=[t.split(':')[0] for t in fetches],
         input_graph_def=graph_def,
         precision_mode=self.tensorrt)
     infer_graph = converter.convert()

efficientdet/model_inspect.py

@@ -328,6 +328,11 @@ def dataset_parser(self, value):
       features['orig_image'] = tf.image.decode_image(image_bytes)
     return features, labels
 
+  def fetch_dataset(self, filename):
+    buffer_size = (self.file_buffer_size_m or 8) * 1024 * 1024  # 8 MiB
+    dataset = tf.data.TFRecordDataset(filename, buffer_size=buffer_size)
+    return dataset
+
   def make_source_dataset(self, index, num_hosts):
     """See base class."""
     if self.data_dir == 'null' or not self.data_dir:
@@ -354,14 +359,9 @@ def make_source_dataset(self, index, num_hosts):
     if self.is_training and not self.cache:
       dataset = dataset.repeat()
 
-    def fetch_dataset(filename):
-      buffer_size = (self.file_buffer_size_m or 8) * 1024 * 1024  # 8 MiB
-      dataset = tf.data.TFRecordDataset(filename, buffer_size=buffer_size)
-      return dataset
-
     # Read the data from disk in parallel
     dataset = dataset.interleave(
-        fetch_dataset,
+        self.fetch_dataset,
         num_parallel_calls=tf.data.experimental.AUTOTUNE,
         deterministic=self.debug)
 

efficientnetv2/datasets.py

@@ -511,6 +511,7 @@ def __init__(self,
     Args:
       model_name: A string of model name.
       model_config: A dict of model configurations or a string of hparams.
+      include_top: If True, include the top layer for classification.
       name: A string of layer name.
 
     Raises:
@@ -581,17 +582,17 @@ def summary(self, input_shape=(224, 224, 3), **kargs):
     return model.summary()
 
   def get_model_with_inputs(self, inputs, **kargs):
-    model = tf.keras.Model(inputs=[inputs], outputs=self.call(inputs, training=True))
-    return  model
+    model = tf.keras.Model(
+        inputs=[inputs], outputs=self.call(inputs, training=True))
+    return model
 
   def call(self, inputs, training, with_endpoints=False):
     """Implementation of call().
 
     Args:
       inputs: input tensors.
       training: boolean, whether the model is constructed for training.
-      features_only: build the base feature network only.
-      single_out: If true, only return the single output.
+      with_endpoints: If true, return a list of endpoints.
 
     Returns:
       output tensors.
@@ -657,7 +658,7 @@ def get_model(model_name,
               pretrained=True,
               training=True,
               with_endpoints=False,
-              **kargs):
+              **kwargs):
   """Get a EfficientNet V1 or V2 model instance.
 
   This is a simply utility for finetuning or inference.
@@ -669,23 +670,29 @@ def get_model(model_name,
     pretrained: if true, download the checkpoint. If string, load the ckpt.
     training: If true, all model variables are trainable.
     with_endpoints: whether to return all intermedia endpoints.
+    **kwargs: additional parameters for keras model, such as name=xx.
 
   Returns:
     A single tensor if with_endpoints if False; otherwise, a list of tensor.
   """
-  net = EffNetV2Model(model_name, model_config, include_top)
+  net = EffNetV2Model(model_name, model_config, include_top, **kwargs)
   net(tf.keras.Input(shape=(None, None, 3)),
       training=training,
       with_endpoints=with_endpoints)
-  if pretrained is True:
+  if pretrained is True:  # pylint: disable=g-bool-id-comparison
+    # pylint: disable=line-too-long
     # download checkpoint and set pretrained path. Supported models include:
-    #    efficientnetv2-s, efficientnetv2-m, efficientnetv2-l,
-    #    efficientnetv2-b0, efficientnetv2-b1, efficientnetv2-b2, efficientnetv2-b3, 
-    #    efficientnet-b0, efficientnet-b1, efficientnet-b2, efficientnet-b3,
-    #    efficientnet-b4, efficientnet-b5, efficientnet-b6, efficientnet-b7, efficientnet-l2
-    # More V2 ckpts: https://github.com/google/automl/tree/master/efficientnetv2
-    # More V1 ckpts: https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet
-    url = f'https://storage.googleapis.com/cloud-tpu-checkpoints/efficientnet/v2/{model_name}.tgz'
+    #   efficientnetv2-s, efficientnetv2-m, efficientnetv2-l,
+    #   efficientnetv2-b0, efficientnetv2-b1, efficientnetv2-b2, efficientnetv2-b3,
+    #   efficientnet-b0, efficientnet-b1, efficientnet-b2, efficientnet-b3,
+    #   efficientnet-b4, efficientnet-b5, efficientnet-b6, efficientnet-b7,
+    #   efficientnet-l2
+    # v2: https://github.com/google/automl/tree/master/efficientnetv2
+    # v1: https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet
+    # pylint: enable=line-too-long
+
+    url = ('https://storage.googleapis.com/cloud-tpu-checkpoints/'
+           f'efficientnet/v2/{model_name}.tgz')
     pretrained_ckpt = tf.keras.utils.get_file(model_name, url, untar=True)
   else:
     pretrained_ckpt = pretrained
@@ -695,4 +702,4 @@ def get_model(model_name,
       pretrained_ckpt = tf.train.latest_checkpoint(pretrained_ckpt)
     net.load_weights(pretrained_ckpt)
 
-  return net
+  return net

efficientnetv2/effnetv2_model.py

@@ -13,7 +13,6 @@
 # limitations under the License.
 # ==============================================================================
 """A simple example on how to use keras model for inference."""
-import copy
 import time
 from absl import app
 from absl import flags
@@ -22,12 +21,10 @@
 import tensorflow.compat.v1 as tf1
 import tensorflow_datasets as tfds
 
-import datasets
-import effnetv2_configs
 import effnetv2_model
-import hparams
 import preprocessing
 import utils
+
 FLAGS = flags.FLAGS
 
 
@@ -47,15 +44,6 @@ def define_flags():
   flags.DEFINE_bool('mixed_precision', False, 'If True, use mixed precision.')
 
 
-def get_config(model_name, dataset_cfg, hparam_str=''):
-  """Create a keras model for EffNetV2."""
-  config = copy.deepcopy(effnetv2_configs.get_model_config(model_name))
-  config.update(datasets.get_dataset_config(dataset_cfg))
-  config.override(hparam_str, allow_new_keys=True)
-  config.model.num_classes = config.data.num_classes
-  return config
-
-
 def build_tf2_model():
   """Build the tf2 model."""
   tf.config.run_functions_eagerly(FLAGS.debug)
@@ -65,10 +53,14 @@ def build_tf2_model():
     tf.keras.mixed_precision.set_global_policy(policy)
 
   model = effnetv2_model.get_model(
-    FLAGS.model_name, FLAGS.hparam_str, include_top=True, pretrained=FLAGS.model_dir or True)
+      FLAGS.model_name,
+      FLAGS.hparam_str,
+      include_top=True,
+      pretrained=FLAGS.model_dir or True)
   model.summary()
   return model
 
+
 def tf2_eval_dataset():
   """Run TF2 benchmark and inference."""
   model = build_tf2_model()
@@ -131,8 +123,7 @@ def tf1_benchmark():
   with tf1.Session() as sess:
     model = effnetv2_model.EffNetV2Model(FLAGS.model_name, FLAGS.hparam_str)
     batch_size = FLAGS.batch_size
-    run_options = tf1.RunOptions(
-        trace_level=tf1.RunOptions.FULL_TRACE)
+    run_options = tf1.RunOptions(trace_level=tf1.RunOptions.FULL_TRACE)
     run_metadata = tf1.RunMetadata()
     isize = FLAGS.image_size or model.cfg.eval.isize
     inputs = tf.ones((batch_size, isize, isize, 3), tf.float16)