我正在使用Tensorflow API的Faster R-CNN Resnet 101。
我为class_head和box_head预测方法添加了多个滤除层，以实现更快的MC滤除采样。例如，如果添加50个滤除层，我希望接收300 * 50个检测。然后，我将计算输出的方差。
下面是我的更改后来自box_head.py的示例。
当我开始评估时，我收到一个错误。我认为该模型无法理解我正在发送一批50个输出，但是由于我使用的是同一张图片，因此将其视为一个输出。问题在于在NMS之前的第二阶段使用锚点对盒子进行解码。
我想知道是否有办法解决这个问题。我真的需要用这样的技巧来加快推理速度。
# New method
def multi_dropout(self, features, keep_prob=0.5, is_training=True, n_samples=10):
output = []
for dropout_index in range(n_samples):
output.append(slim.dropout(
features,
keep_prob=keep_prob,
is_training=is_training))
stacked = tf.stack(output)
return stacked
def predict(self, features, num_predictions_per_location=1):
"""Predicts boxes.
Args:
features: A float tensor of shape [batch_size, height, width,
channels] containing features for a batch of images.
num_predictions_per_location: Int containing number of predictions per
location.
Returns:
box_encodings: A float tensor of shape
[batch_size, 1, num_classes, code_size] representing the location of the
objects.
Raises:
ValueError: If num_predictions_per_location is not 1.
"""
if num_predictions_per_location != 1:
raise ValueError('Only num_predictions_per_location=1 is supported')
spatial_averaged_roi_pooled_features = tf.reduce_mean(
features, [1, 2], keep_dims=True, name='AvgPool')
flattened_roi_pooled_features = slim.flatten(
spatial_averaged_roi_pooled_features)
if self._use_dropout:
# New controls
if self._is_training:
flattened_roi_pooled_features = slim.dropout(
flattened_roi_pooled_features,
keep_prob=self._dropout_keep_prob,
is_training=self._is_training)
else:
flattened_roi_pooled_features = self.multi_dropout(features=flattened_roi_pooled_features, keep_prob=self._dropout_keep_prob, is_training=self._is_training, n_samples=50)
number_of_boxes = 1
if not self._share_box_across_classes:
number_of_boxes = self._num_classes
with slim.arg_scope(self._fc_hyperparams_fn()):
box_encodings = slim.fully_connected(
flattened_roi_pooled_features,
number_of_boxes * self._box_code_size,
activation_fn=None,
scope='BoxEncodingPredictor')
box_encodings = tf.reshape(box_encodings,
[-1, 1, number_of_boxes, self._box_code_size])
return box_encodings
错误：
Traceback (most recent call last):
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 1576, in _create_c_op
c_op = c_api.TF_FinishOperation(op_desc)
tensorflow.python.framework.errors_impl.InvalidArgumentError: Dimensions must be equal, but are 45000 and 900 for 'SecondStagePostprocessor/Decode/mul' (op: 'Mul') with input shapes: [45000], [900].
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "./legacy/eval.py", line 141, in <module>
tf.app.run()
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/python/platform/app.py", line 125, in run
_sys.exit(main(argv))
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/python/util/deprecation.py", line 272, in new_func
return func(*args, **kwargs)
File "./legacy/eval.py", line 138, in main
graph_hook_fn=graph_rewriter_fn)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/models/research/object_detection/legacy/evaluator.py", line 187, in evaluate
ignore_groundtruth=eval_config.ignore_groundtruth)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/models/research/object_detection/legacy/evaluator.py", line 79, in _extract_predictions_and_losses
detections = model.postprocess(prediction_dict, true_image_shapes)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/models/research/object_detection/meta_architectures/faster_rcnn_meta_arch.py", line 1175, in postprocess
mask_predictions=mask_predictions)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/models/research/object_detection/meta_architectures/faster_rcnn_meta_arch.py", line 1595, in _postprocess_box_classifier
refined_box_encodings_batch, proposal_boxes)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/models/research/object_detection/meta_architectures/faster_rcnn_meta_arch.py", line 1665, in _batch_decode_boxes
tiled_anchors_boxlist)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/models/research/object_detection/core/box_coder.py", line 86, in decode
return self._decode(rel_codes, anchors)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/models/research/object_detection/box_coders/faster_rcnn_box_coder.py", line 111, in _decode
w = tf.exp(tw) * wa
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/python/ops/math_ops.py", line 850, in binary_op_wrapper
return func(x, y, name=name)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/python/ops/math_ops.py", line 1094, in _mul_dispatch
return gen_math_ops.mul(x, y, name=name)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/python/ops/gen_math_ops.py", line 4936, in mul
"Mul", x=x, y=y, name=name)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/python/framework/op_def_library.py", line 787, in _apply_op_helper
op_def=op_def)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/python/util/deprecation.py", line 454, in new_func
return func(*args, **kwargs)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 3155, in create_op
op_def=op_def)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 1731, in __init__
control_input_ops)
File "/home/a21/.virtualenvs/tf/lib/python3.5/site-packages/tensorflow/python/framework/ops.py", line 1579, in _create_c_op
raise ValueError(str(e))
ValueError: Dimensions must be equal, but are 45000 and 900 for 'SecondStagePostprocessor/Decode/mul' (op: 'Mul') with input shapes: [45000], [900].
我也尝试将输出一一发送到盒式解码器，但再次收到错误
ValueError: Shape must be rank 3 but is rank 4 for 'BatchMultiClassNonMaxSuppression/map/while/Slice' (op: 'Slice') with input shapes: [1,1,1,4], [3], [3].