def accuracy_function(self, logits, labels, data_type):
"""Returns the ops to measure the mean precision of the model."""
try:
import ssd_dataloader # pylint: disable=g-import-not-at-top
from object_detection.box_coders import faster_rcnn_box_coder # pylint: disable=g-import-not-at-top
from object_detection.core import box_coder # pylint: disable=g-import-not-at-top
from object_detection.core import box_list # pylint: disable=g-import-not-at-top
except ImportError:
raise ImportError('To use the COCO dataset, you must clone the '
'repo https://github.com/tensorflow/models and add '
'tensorflow/models and tensorflow/models/research to '
'the PYTHONPATH, and compile the protobufs by '
'following https://github.com/tensorflow/models/blob/'
'master/research/object_detection/g3doc/installation.md'
'#protobuf-compilation ; To evaluate using COCO'
'metric, download and install Python COCO API from'
'https://github.com/cocodataset/cocoapi')
# Unpack model output back to locations and confidence scores of predictions
# pred_locs: relative locations (coordiates) of objects in all SSD boxes
# shape: [batch_size, NUM_SSD_BOXES, 4]
# pred_labels: confidence scores of objects being of all categories
# shape: [batch_size, NUM_SSD_BOXES, label_num]
pred_locs, pred_labels = tf.split(logits, [4, self.label_num], 2)
ssd_box_coder = faster_rcnn_box_coder.FasterRcnnBoxCoder(
scale_factors=ssd_constants.BOX_CODER_SCALES)
anchors = box_list.BoxList(
tf.convert_to_tensor(ssd_dataloader.DefaultBoxes()('ltrb')))
pred_boxes = box_coder.batch_decode(
encoded_boxes=pred_locs, box_coder=ssd_box_coder, anchors=anchors)
pred_scores = tf.nn.softmax(pred_labels, axis=2)
boxes_classes, id_shape = tf.split(
labels, [ssd_constants.MAX_NUM_EVAL_BOXES, 1], 1)
# TODO(haoyuzhang): maybe use these values for visualization.
gt_boxes, gt_classes = tf.split(boxes_classes, [4, 1], 2) # pylint: disable=unused-variable
id_shape = tf.squeeze(id_shape, 1)
source_id, raw_shape, _ = tf.split(id_shape, [1, 3, 1], 1)
source_id = tf.squeeze(source_id, 1)
return {
(constants.UNREDUCED_ACCURACY_OP_PREFIX +
ssd_constants.PRED_BOXES): pred_boxes,
(constants.UNREDUCED_ACCURACY_OP_PREFIX +
ssd_constants.PRED_SCORES): pred_scores,
# TODO(haoyuzhang): maybe use these values for visualization.
# constants.UNREDUCED_ACCURACY_OP_PREFIX+'gt_boxes': gt_boxes,
# constants.UNREDUCED_ACCURACY_OP_PREFIX+'gt_classes': gt_classes,
(constants.UNREDUCED_ACCURACY_OP_PREFIX +
ssd_constants.SOURCE_ID): source_id,
(constants.UNREDUCED_ACCURACY_OP_PREFIX +
ssd_constants.RAW_SHAPE): raw_shape
}
def loss_function(self, build_network_result, labels):
logits = build_network_result.logits
# Unpack model output back to locations and confidence scores of predictions
# Shape of pred_loc: [batch_size, 4, NUM_SSD_BOXES]
# Shape of pred_label: [batch_size, label_num, NUM_SSD_BOXES]
pred_loc, pred_label = tf.split(logits, [4, self.label_num], 1)
# Unpack ground truth labels to number of boxes, locations, and classes
# initial shape: [batch_size, NUM_SSD_BOXES, 5]
# Shape of labels: [batch_size, NUM_SSD_BOXES, 5]
# Shape of num_gt: [batch_size, 1, 5] -- 5 identical copies
labels, num_gt = tf.split(labels, [ssd_constants.NUM_SSD_BOXES, 1], 1)
# Shape of num_gt: [batch_size]
num_gt = tf.squeeze(tf.cast(num_gt[:, :, 0], tf.int32))
# Shape of gt_loc: [batch_size, NUM_SSD_BOXES, 4]
# Shape of gt_label: [batch_size, NUM_SSD_BOXES, 1]
gt_loc, gt_label = tf.split(labels, [4, 1], 2)
gt_label = tf.cast(gt_label, tf.int32)
cross_entropy = tf.losses.sparse_softmax_cross_entropy(
gt_label,
tf.transpose(pred_label, [0, 2, 1]),
reduction=tf.losses.Reduction.NONE)
default_boxes = tf.tile(
tf.convert_to_tensor(
ssd_dataloader.DefaultBoxes()('xywh'))[tf.newaxis, :, :],
[gt_loc.get_shape()[0], 1, 1])
# To performance people: MLPerf uses this transposed convention.
# I (taylorrobie) have matched it to make it easier to compare to the
# reference. If this hurts performance, feel free to adjust accordingly.
gt_label = tf.squeeze(gt_label)
# pred_loc, pred_label, gt_loc, default_boxes = [
# tf.transpose(i, (0, 2, 1)) for i in
# [pred_loc, pred_label, gt_loc, default_boxes]
# ]
# Shape of gt_loc: [batch_size, 4, NUM_SSD_BOXES]
gt_loc = tf.transpose(gt_loc, [0, 2, 1])
# Shape of default_boxes: [batch_size, 4, NUM_SSD_BOXES]
default_boxes = tf.transpose(default_boxes, [0, 2, 1])
mask = tf.greater(gt_label, 0)
float_mask = tf.cast(mask, tf.float32)
gt_location_vectors = tf.concat(
[
(ssd_constants.SCALE_XY *
(gt_loc[:, :2, :] - default_boxes[:, :2, :]) /
default_boxes[:, 2:, :]),
# The gt_loc height and width have already had the log taken.
# See FasterRcnnBoxCoder for more details.
(ssd_constants.SCALE_HW *
(gt_loc[:, 2:, :] - tf.log(default_boxes[:, 2:, :])))
],
axis=1)
smooth_l1 = tf.reduce_sum(tf.losses.huber_loss(
gt_location_vectors, pred_loc, reduction=tf.losses.Reduction.NONE),
axis=1)
smooth_l1 = tf.multiply(smooth_l1, float_mask)
box_loss = tf.reduce_sum(smooth_l1, axis=1)
# Hard example mining
neg_masked_cross_entropy = cross_entropy * (1 - float_mask)
relative_position = tf.contrib.framework.argsort(
tf.contrib.framework.argsort(neg_masked_cross_entropy,
direction='DESCENDING'))
num_neg_boxes = num_gt * ssd_constants.NEGS_PER_POSITIVE
top_k_neg_mask = tf.cast(
tf.less(
relative_position,
tf.tile(num_neg_boxes[:, tf.newaxis],
(1, ssd_constants.NUM_SSD_BOXES))), tf.float32)
class_loss = tf.reduce_sum(tf.multiply(cross_entropy,
float_mask + top_k_neg_mask),
axis=1)
class_loss = tf.reduce_mean(class_loss / tf.cast(num_gt, tf.float32))
box_loss = tf.reduce_mean(box_loss / tf.cast(num_gt, tf.float32))
return class_loss + box_loss
