def build_prediction(self, examples, **kwargs):
"""Builds tf graph for prediction.
Args:
examples: dict of input tensors keyed by name.
prediction_task: the specific prediction task.
Returns:
predictions: dict of prediction results keyed by name.
"""
options = self._model_proto
is_training = self._is_training
if is_training or len(options.eval_min_dimension) == 0:
predictions = self._build_prediction(examples)
predictions.update(self._postprocess(examples, predictions))
return predictions
inputs = examples[InputDataFields.image]
assert inputs.get_shape()[0].value == 1
proposal_scores_list = [[] for _ in range(1 + options.oicr_iterations)]
# Get predictions from different resolutions.
reuse = False
for min_dimension in options.eval_min_dimension:
inputs_resized = tf.expand_dims(
imgproc.resize_image_to_min_dimension(inputs[0],
min_dimension)[0],
axis=0)
examples[InputDataFields.image] = inputs_resized
with tf.variable_scope(tf.get_variable_scope(), reuse=reuse):
predictions = self._build_prediction(examples)
for i in range(1 + options.oicr_iterations):
proposals_scores = predictions[
Cap2DetPredictions.oicr_proposal_scores +
'_at_{}'.format(i)]
proposal_scores_list[i].append(proposals_scores)
reuse = True
# Aggregate predictions from different resolutions.
predictions_aggregated = predictions
for i in range(1 + options.oicr_iterations):
proposal_scores = tf.stack(proposal_scores_list[i], axis=-1)
proposal_scores = tf.reduce_mean(proposal_scores, axis=-1)
predictions_aggregated[Cap2DetPredictions.oicr_proposal_scores +
'_at_{}'.format(i)] = proposal_scores
predictions_aggregated.update(
self._postprocess(inputs, predictions_aggregated))
return predictions_aggregated
def test_resize_image_to_min_dimension(self):
tf.reset_default_graph()
image = tf.placeholder(tf.float32, shape=[None, None, 3])
resized_image = imgproc.resize_image_to_min_dimension(
image, min_dimension=900)
with self.test_session() as sess:
img, img_shape = sess.run(resized_image,
feed_dict={
image: np.zeros([300, 400, 3]),
})
self.assertEqual(img.shape, (900, 1200, 3))
self.assertAllEqual(img_shape, [900, 1200, 3])
img, img_shape = sess.run(resized_image,
feed_dict={
image: np.zeros([400, 300, 3]),
})
self.assertEqual(img.shape, (1200, 900, 3))
self.assertAllEqual(img_shape, [1200, 900, 3])
def _keep_aspect_ratio_resize_fn(image):
return imgproc.resize_image_to_min_dimension(
image, min_dimension=options.min_dimension)