Merge pull request NVIDIA-AI-IOT#29 from NVIDIA-AI-IOT/improved_model…

…_support

Improved model support

tf_trt_models/detection.py

@@ -1,168 +1,174 @@
-from object_detection.protos.pipeline_pb2 import TrainEvalPipelineConfig
-from object_detection.builders import model_builder
+from object_detection.protos import pipeline_pb2
+from object_detection.protos import image_resizer_pb2
+from object_detection import exporter
 
 import os
-import tarfile
 import subprocess
 
+from collections import namedtuple
 from google.protobuf import text_format
 
 import tensorflow as tf
 
-from .graph_utils import convert_relu6, remove_op
-
-input_name = 'input'
-output_map = {
- 'detection_scores': 'scores',
- 'detection_boxes': 'boxes',
- 'detection_classes': 'classes',
- 'detection_masks': 'masks'
+from .graph_utils import force_nms_cpu as f_force_nms_cpu
+from .graph_utils import replace_relu6 as f_replace_relu6
+from .graph_utils import remove_assert as f_remove_assert
+
+DetectionModel = namedtuple('DetectionModel', ['name', 'url', 'extract_dir'])
+
+INPUT_NAME='image_tensor'
+BOXES_NAME='detection_boxes'
+CLASSES_NAME='detection_classes'
+SCORES_NAME='detection_scores'
+MASKS_NAME='detection_masks'
+NUM_DETECTIONS_NAME='num_detections'
+FROZEN_GRAPH_NAME='frozen_inference_graph.pb'
+PIPELINE_CONFIG_NAME='pipeline.config'
+CHECKPOINT_PREFIX='model.ckpt'
+
+
+MODELS = {
+ 'ssd_mobilenet_v1_coco': DetectionModel(
+ 'ssd_mobilenet_v1_coco',
+ 'http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_2018_01_28.tar.gz',
+ 'ssd_mobilenet_v1_coco_2018_01_28',
+ ),
+ 'ssd_mobilenet_v2_coco': DetectionModel(
+ 'ssd_mobilenet_v2_coco',
+ 'http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v2_coco_2018_03_29.tar.gz',
+ 'ssd_mobilenet_v2_coco_2018_03_29',
+ ),
+ 'ssd_inception_v2_coco': DetectionModel(
+ 'ssd_inception_v2_coco',
+ 'http://download.tensorflow.org/models/object_detection/ssd_inception_v2_coco_2018_01_28.tar.gz',
+ 'ssd_inception_v2_coco_2018_01_28',
+ ),
+ 'ssd_resnet_50_fpn_coco': DetectionModel(
+ 'ssd_resnet_50_fpn_coco',
+ 'http://download.tensorflow.org/models/object_detection/ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync_2018_07_03.tar.gz',
+ 'ssd_resnet50_v1_fpn_shared_box_predictor_640x640_coco14_sync_2018_07_03',
+ ),
+ 'faster_rcnn_resnet50_coco': DetectionModel(
+ 'faster_rcnn_resnet50_coco',
+ 'http://download.tensorflow.org/models/object_detection/faster_rcnn_resnet50_coco_2018_01_28.tar.gz',
+ 'faster_rcnn_resnet50_coco_2018_01_28',
+ ),
+ 'faster_rcnn_nas': DetectionModel(
+ 'faster_rcnn_nas',
+ 'http://download.tensorflow.org/models/object_detection/faster_rcnn_nas_coco_2018_01_28.tar.gz',
+ 'faster_rcnn_nas_coco_2018_01_28',
+ ),
+ 'mask_rcnn_resnet50_atrous_coco': DetectionModel(
+ 'mask_rcnn_resnet50_atrous_coco',
+ 'http://download.tensorflow.org/models/object_detection/mask_rcnn_resnet50_atrous_coco_2018_01_28.tar.gz',
+ 'mask_rcnn_resnet50_atrous_coco_2018_01_28',
+ )
 }
 
-nets = {
- 'ssd_mobilenet_v1_coco': {
- 'config_url': 'https://raw.githubusercontent.com/tensorflow/models/master/research/object_detection/samples/configs/ssd_mobilenet_v1_coco.config',
- 'checkpoint_url': 'http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_2017_11_17.tar.gz',
- },
- 'ssd_mobilenet_v2_coco': {
- 'config_url': 'https://raw.githubusercontent.com/tensorflow/models/master/research/object_detection/samples/configs/ssd_mobilenet_v2_coco.config',
- 'checkpoint_url': 'http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v2_coco_2018_03_29.tar.gz',
- },
- 'ssd_inception_v2_coco': {
- 'config_url': 'https://raw.githubusercontent.com/tensorflow/models/master/research/object_detection/samples/configs/ssd_inception_v2_coco.config',
- 'checkpoint_url': 'http://download.tensorflow.org/models/object_detection/ssd_inception_v2_coco_2017_11_17.tar.gz',
- },
-}
 
-def download_detection_model(model, output_dir='.'):
- """Download a default detection model configuration and checkpoint.
-
- This function downloads a default detection model configuration and
- checkpoint. This is only available for a subset of models in the
- TensorFlow object detection model zoo that are known to work on Jetson.
-
- The following models are available
-
- ssd_mobilenet_v1_coco
- ssd_mobilenet_v2_coco
- ssd_inception_v2_coco
-
- :param model: the model name from the above list
- :type model: string
- :param output_dir: the directory where files are downloaded to
- :type output_dir: string
- :return config_path: path to the object detection pipeline config file
- :rtype string
- :return checkpoint_path: path to the checkpoint files prefix containing trained model params
- :rtype string
- """
- global nets
- config_path = ''
- checkpoint_path = ''
-
- if not os.path.exists(output_dir):
- os.makedirs(output_dir)
-
- modeldir_path = os.path.join(output_dir, model)
- if not os.path.exists(modeldir_path):
- os.makedirs(modeldir_path)
-
- config_path = os.path.join(output_dir, model + '.config')
- if not os.path.isfile(config_path):
- subprocess.call(['wget', '--no-check-certificate', nets[model]['config_url'], '-O', config_path]) 
-
- modeltar_path = os.path.join(output_dir, os.path.basename(nets[model]['checkpoint_url']))
- if not os.path.isfile(modeltar_path):
- subprocess.call(['wget', '--no-check-certificate', nets[model]['checkpoint_url'], '-O', modeltar_path])
-
- tar_file = tarfile.open(modeltar_path)
- for file in tar_file.getmembers():
- file_name = os.path.basename(file.name)
- if 'model.ckpt' in file_name:
- file.name = file_name
- tar_file.extract(file, modeldir_path)
-
- checkpoint_path = os.path.join(modeldir_path, 'model.ckpt')
+def get_input_names(model):
+ return [INPUT_NAME]
 
- return config_path, checkpoint_path
 
-def build_detection_graph(config, checkpoint):
- """Build an object detection model from the TensorFlow model zoo.
+def get_output_names(model):
+ output_names = [BOXES_NAME, CLASSES_NAME, SCORES_NAME, NUM_DETECTIONS_NAME]
+ if model == 'mask_rcnn_resnet50_atrous_coco':
+ output_names.append(MASKS_NAME)
+ return output_names
 
- This function creates an object detection model, sourced from the
- TensorFlow object detection API.
 
- It is necessary to use this function to generate a frozen graph that is
- compatible with TensorFlow/TensorRT integration. In addition to generating
- a graph that is compatible with TensorFlow's TensorRT package, this
- function performs other graph modifications, such as forced device
- placement, that improve performance on Jetson. These graph modifications
- are tested with a subset of the object detection API and may or may not
- work well with models not listed.
-
- The workflow when using this method is:
-
- 1. Train model using TensorFlow object detection API
- 2. Build graph configured for Jetson using this function
- 3. Optimize the graph output by this method with the TensorRT package in
- TensorFlow
- 4. Execute in regular TensorFlow, or using the high level TFModel class
+def download_detection_model(model, output_dir='.'):
+ """Downloads a pre-trained object detection model"""
+ global MODELS
 
- :param config: path to the object detection pipeline config file
- :type config: string
- :param checkpoint: path to the checkpoint files prefix containing trained model params
- :type checkpoint: string
- :returns: the configured frozen graph representing object detection model
- :rtype: a tensorflow GraphDef
- """
- global input_name, output_map
+ model_name = model
 
- if isinstance(config, str):
- with open(config, 'r') as f:
- config_str = f.read()
- config = TrainEvalPipelineConfig()
- text_format.Merge(config_str, config)
+ model = MODELS[model_name]
+ subprocess.call(['mkdir', '-p', output_dir])
+ tar_file = os.path.join(output_dir, os.path.basename(model.url))
 
+ config_path = os.path.join(output_dir, model.extract_dir, PIPELINE_CONFIG_NAME)
+ checkpoint_path = os.path.join(output_dir, model.extract_dir, CHECKPOINT_PREFIX)
 
- tf_config = tf.ConfigProto()
- tf_config.gpu_options.allow_growth = True
+ if not os.path.exists(os.path.join(output_dir, model.extract_dir)):
+ subprocess.call(['wget', model.url, '-O', tar_file])
+ subprocess.call(['tar', '-xzf', tar_file, '-C', output_dir])
 
- with tf.Graph().as_default() as tf_graph:
- with tf.Session(config=tf_config) as tf_sess:
+  # hack fix to handle mobilenet_v2 config bug
+ subprocess.call(['sed', '-i', '/batch_norm_trainable/d', config_path])
 
-  model = model_builder.build(model_config=config.model, is_training=False)
+ return config_path, checkpoint_path
 
- tf_input = tf.placeholder(tf.float32, [1, None, None, 3], name=input_name)
- tf_preprocessed, tf_true_image_shapes = model.preprocess(tf_input)
- tf_predictions = model.predict(preprocessed_inputs=tf_preprocessed,
- true_image_shapes=tf_true_image_shapes)
- tf_postprocessed = model.postprocess(
- prediction_dict=tf_predictions,
- true_image_shapes=tf_true_image_shapes
- )
 
- tf_saver = tf.train.Saver()
- tf_saver.restore(save_path=checkpoint, sess=tf_sess)
+def build_detection_graph(config, checkpoint,
+ batch_size=1,
+ score_threshold=None,
+ force_nms_cpu=True,
+ replace_relu6=True,
+ remove_assert=True,
+ input_shape=None,
+ output_dir='.generated_model'):
+ """Builds a frozen graph for a pre-trained object detection model"""
+
+ config_path = config
+ checkpoint_path = checkpoint
+
+ # parse config from file
+ config = pipeline_pb2.TrainEvalPipelineConfig()
+ with open(config_path, 'r') as f:
+ text_format.Merge(f.read(), config, allow_unknown_extension=True)
+
+ # override some config parameters
+ if config.model.HasField('ssd'):
+ config.model.ssd.feature_extractor.override_base_feature_extractor_hyperparams = True
+ if score_threshold is not None:
+ config.model.ssd.post_processing.batch_non_max_suppression.score_threshold = score_threshold 
+ if input_shape is not None:
+ config.model.ssd.image_resizer.fixed_shape_resizer.height = input_shape[0]
+ config.model.ssd.image_resizer.fixed_shape_resizer.width = input_shape[1]
+ elif config.model.HasField('faster_rcnn'):
+ if score_threshold is not None:
+ config.model.faster_rcnn.second_stage_post_processing.score_threshold = score_threshold
+ if input_shape is not None:
+ config.model.faster_rcnn.image_resizer.fixed_shape_resizer.height = input_shape[0]
+ config.model.faster_rcnn.image_resizer.fixed_shape_resizer.width = input_shape[1]
+
+ if os.path.isdir(output_dir):
+ subprocess.call(['rm', '-rf', output_dir])
 
- outputs = {}
- for key, op in tf_postprocessed.items():
- if key in output_map.keys():
- outputs[output_map[key]] = \
- tf.identity(op, name=output_map[key])
+ tf_config = tf.ConfigProto()
+ tf_config.gpu_options.allow_growth = True
 
- frozen_graph = tf.graph_util.convert_variables_to_constants(
- tf_sess,
- tf_sess.graph_def,
- output_node_names=list(outputs.keys())
+ # export inference graph to file (initial)
+ with tf.Session(config=tf_config) as tf_sess:
+ with tf.Graph().as_default() as tf_graph:
+ exporter.export_inference_graph(
+ 'image_tensor', 
+ config, 
+ checkpoint_path, 
+ output_dir, 
+ input_shape=[batch_size, None, None, 3]
  )
 
- frozen_graph = convert_relu6(frozen_graph)
+ # read frozen graph from file
+ frozen_graph = tf.GraphDef()
+ with open(os.path.join(output_dir, FROZEN_GRAPH_NAME), 'rb') as f:
+ frozen_graph.ParseFromString(f.read())
+
+ # apply graph modifications
+ if force_nms_cpu:
+ frozen_graph = f_force_nms_cpu(frozen_graph)
+ if replace_relu6:
+ frozen_graph = f_replace_relu6(frozen_graph)
+ if remove_assert:
+ frozen_graph = f_remove_assert(frozen_graph)
 
- remove_op(frozen_graph, 'Assert')
+ # get input names
+ # TODO: handle mask_rcnn 
+ input_names = [INPUT_NAME]
+ output_names = [BOXES_NAME, CLASSES_NAME, SCORES_NAME, NUM_DETECTIONS_NAME]
 
- # force CPU device placement for NMS ops
- for node in frozen_graph.node:
- if 'NonMaxSuppression' in node.name:
- node.device = '/device:CPU:0'
+ # remove temporary directory
+ subprocess.call(['rm', '-rf', output_dir])
 
- return frozen_graph, [input_name], list(outputs.keys())
+ return frozen_graph, input_names, output_names

tf_trt_models/graph_utils.py

@@ -73,3 +73,19 @@ def remove_op(graph_def, op_name):
  matches = [node for node in graph_def.node if node.op == op_name]
  for match in matches:
  remove_node(graph_def, match)
+
+
+def force_nms_cpu(frozen_graph):
+ for node in frozen_graph.node:
+ if 'NonMaxSuppression' in node.name:
+ node.device = '/device:CPU:0'
+ return frozen_graph
+
+
+def replace_relu6(frozen_graph):
+ return convert_relu6(frozen_graph)
+
+
+def remove_assert(frozen_graph):
+ remove_op(frozen_graph, 'Assert')
+ return frozen_graph