Tensorflow对象检测API多尺度推理变量重用

Question

I was trying to build multi-scale object detection inference code based on Tensorflow Object Detection API. However I don't quite know how to get around with variable reuse issue when predicting boxes on difference scale of a same image in one session. Here's what I did and where I don't understand:

In https://github.com/tensorflow/models/blob/master/research/object_detection/evaluator.py , I duplicate the line 'prediction_dict = model.predict(preprocessed_image, true_image_shapes)' one more time as you can see below.

def _extract_predictions_and_losses(model,
                                    create_input_dict_fn,
                                    ignore_groundtruth=False):
  """Constructs tensorflow detection graph and returns output tensors.
  Args:
    model: model to perform predictions with.
    create_input_dict_fn: function to create input tensor dictionaries.
    ignore_groundtruth: whether groundtruth should be ignored.
  Returns:
    prediction_groundtruth_dict: A dictionary with postprocessed tensors (keyed
      by standard_fields.DetectionResultsFields) and optional groundtruth
      tensors (keyed by standard_fields.InputDataFields).
    losses_dict: A dictionary containing detection losses. This is empty when
      ignore_groundtruth is true.
  """
  input_dict = create_input_dict_fn()
  prefetch_queue = prefetcher.prefetch(input_dict, capacity=500)
  input_dict = prefetch_queue.dequeue()
  original_image = tf.expand_dims(input_dict[fields.InputDataFields.image], 0)
  preprocessed_image, true_image_shapes = model.preprocess(
      tf.to_float(original_image))


  prediction_dict1 = model.predict(preprocessed_image, true_image_shapes)
  /****Some code to resize preprocessed_image****/
  prediction_dict2 = model.predict(preprocessed_image, true_image_shapes)


  detections = model.postprocess(prediction_dict, true_image_shapes)

  groundtruth = None
  losses_dict = {}
  if not ignore_groundtruth:

This gives me the following error:

Traceback (most recent call last):
  File "object_detection/eval_fddb.py", line 167, in <module>
    tf.app.run()
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/python/platform/app.py", line 48, in run
    _sys.exit(main(_sys.argv[:1] + flags_passthrough))
  File "object_detection/eval_fddb.py", line 163, in main
    FLAGS.checkpoint_dir, FLAGS.eval_dir)
  File "/local/mnt/workspace/chris/projects/models/object_detection/evaluator_fddb.py", line 261, in evaluate
    create_input_dict_fn=create_input_dict_fn)
  File "/local/mnt/workspace/chris/projects/models/object_detection/evaluator_fddb.py", line 187, in _extract_prediction_tensors
    prediction_dict = model.predict(preprocessed_image)
  File "/local/mnt/workspace/chris/projects/models/object_detection/meta_architectures/umd_meta_arch.py", line 362, in predict
    image_shape) = self._extract_rpn_feature_maps(preprocessed_inputs)
  File "/local/mnt/workspace/chris/projects/models/object_detection/meta_architectures/umd_meta_arch.py", line 278, in _extract_rpn_feature_maps
    preprocessed_inputs, scope=self.first_stage_feature_extractor_scope)
  File "/local/mnt/workspace/chris/projects/models/object_detection/meta_architectures/faster_rcnn_meta_arch.py", line 154, in extract_proposal_features_w_internal_layers
    return self._extract_proposal_features_w_internal_layers(preprocessed_inputs, scope)
  File "/local/mnt/workspace/chris/projects/models/object_detection/models/faster_rcnn_resnet_v1_feature_extractor.py", line 173, in _extract_proposal_features_w_internal_layers
    scope=var_scope)
  File "/local/mnt/workspace/chris/projects/models/slim/nets/resnet_v1.py", line 300, in resnet_v1_101
    reuse=reuse, scope=scope)
  File "/local/mnt/workspace/chris/projects/models/slim/nets/resnet_v1.py", line 214, in resnet_v1
    net = resnet_utils.conv2d_same(net, 64, 7, stride=2, scope='conv1')
  File "/local/mnt/workspace/chris/projects/models/slim/nets/resnet_utils.py", line 122, in conv2d_same
    rate=rate, padding='VALID', scope=scope)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/framework/python/ops/arg_scope.py", line 181, in func_with_args
    return func(*args, **current_args)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/layers/python/layers/layers.py", line 1027, in convolution
    outputs = layer.apply(inputs)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/python/layers/base.py", line 503, in apply
    return self.__call__(inputs, *args, **kwargs)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/python/layers/base.py", line 443, in __call__
    self.build(input_shapes[0])
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/python/layers/convolutional.py", line 137, in build
    dtype=self.dtype)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/python/layers/base.py", line 383, in add_variable
    trainable=trainable and self.trainable)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.py", line 1065, in get_variable
    use_resource=use_resource, custom_getter=custom_getter)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.py", line 962, in get_variable
    use_resource=use_resource, custom_getter=custom_getter)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.py", line 360, in get_variable
    validate_shape=validate_shape, use_resource=use_resource)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/layers/python/layers/layers.py", line 1561, in layer_variable_getter
    return _model_variable_getter(getter, *args, **kwargs)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/layers/python/layers/layers.py", line 1553, in _model_variable_getter
    custom_getter=getter, use_resource=use_resource)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/framework/python/ops/arg_scope.py", line 181, in func_with_args
    return func(*args, **current_args)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/framework/python/ops/variables.py", line 261, in model_variable
    use_resource=use_resource)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/framework/python/ops/arg_scope.py", line 181, in func_with_args
    return func(*args, **current_args)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/framework/python/ops/variables.py", line 216, in variable
    use_resource=use_resource)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.py", line 352, in _true_getter
    use_resource=use_resource)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/variable_scope.py", line 664, in _get_single_variable
    name, "".join(traceback.format_list(tb))))
ValueError: Variable FirstStageFeatureExtractor/resnet_v1_101/conv1/weights already exists, disallowed. Did you mean to set reuse=True in VarScope? Originally defined at:

  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/framework/python/ops/variables.py", line 216, in variable
    use_resource=use_resource)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/framework/python/ops/arg_scope.py", line 181, in func_with_args
    return func(*args, **current_args)
  File "/local/mnt/workspace/chris/anaconda2/lib/python2.7/site-packages/tensorflow/contrib/framework/python/ops/variables.py", line 261, in model_variable
    use_resource=use_resource)

My understanding is that when each time I call model.predict(), it creates a model with all variables. But in the second time, I can't create all the variables because all of them have been existing with the same name. I tried to add an variable scope with 'reuse=True' for the second model.predict(), but it wouldn't load the variables at all because mismatch names.

My questions is:

How to get around with this by running the same model on two different scale images in one sess.run()?

Thank you very much.

Answer 1

Have you tried using tf.variable_scope() before you build the model?

with tf.variable_scope('first_prediction'):
    prediction_dict1 = model.predict(preprocessed_image, true_image_shapes)

with tf.variable_scope('second_prediction'):
    prediction_dict2 = model.predict(preprocessed_image, true_image_shapes)

This way both model should have different prefix to their original variable name, hence preventing the reuse problem.