def metric_fn(predictions, labels, params=None):
images = params['inputs']
images = denormalize(images)
pred_prob = predictions['prob']
flat_pred_prob = tf.reshape(pred_prob, [-1])
pred_class = tf.to_float(tf.greater_equal(pred_prob, score_threshold))
# add summary for prediction results.
pred_images = pred_class * 255
vis_preds = tf.concat([
tf.cast(images, pred_class.dtype),
tf.tile(pred_images, multiples=[1, 1, 1, 3])
],
axis=2)
vis_preds = tf.cast(vis_preds, tf.uint8)
tf.summary.image('prediction', vis_preds, max_outputs=10)
labels = tf.reshape(labels, [-1])
pred_class = tf.reshape(pred_class, [-1])
miou = tf.metrics.mean_iou(labels, pred_class, num_classes=2)
p_iou = positive_iou(labels,
tf.reshape(pred_class, [-1]),
num_classes=2)
pr_curve('eval/prc',
tf.cast(labels, tf.bool),
flat_pred_prob,
num_thresholds=201)
return {'eval/miou': miou, 'eval/piou': p_iou}
def loss_fn(predictions, labels, params=None):
pred_logit = predictions['logit']
labels = tf.reshape(labels, [-1])
flat_logit = tf.reshape(pred_logit, [-1, num_classes])
if num_classes > 1:
onehot_labels = tf.one_hot(labels, num_classes)
else:
onehot_labels = tf.reshape(labels, [-1, 1])
bpn_weights = balance_positive_negative_weight(labels,
positive_weight=1.,
negative_weight=1.)
# bpn_weights = 1e-3 * bpn_weights
ce_loss = tf.losses.sigmoid_cross_entropy(onehot_labels,
flat_logit,
weights=bpn_weights[:, None])
pred_scores = tf.sigmoid(tf.reshape(flat_logit, [-1]))
# debug
# pred_scores = tf.Print(pred_scores, [pred_scores], message='pred_score = ')
# dice_loss_v = dice_loss(pred_scores, labels)
tf.summary.scalar('loss/cross_entropy_loss', ce_loss)
# tf.summary.scalar('loss/dice_loss', dice_loss_v)
# add metric for training
# mean iou
pred_class = tf.to_float(tf.greater_equal(pred_scores,
score_threshold))
miou = tf.metrics.mean_iou(labels,
tf.reshape(pred_class, [-1]),
num_classes=2)
miou_v = compute_mean_iou(None, miou[1])
tf.identity(miou_v, 'train_miou')
tf.summary.scalar('train/miou', miou_v)
# postive iou
p_iou = positive_iou(labels,
tf.reshape(pred_class, [-1]),
num_classes=2)
p_iou_v = compute_positive_iou(None, p_iou[1])
tf.identity(p_iou_v, 'train_piou')
tf.summary.scalar('train/piou', p_iou_v)
# negative iou
n_iou = negative_iou(labels,
tf.reshape(pred_class, [-1]),
num_classes=2)
n_iou_v = compute_negative_iou(None, n_iou[1])
tf.identity(n_iou_v, 'train_niou')
tf.summary.scalar('train/niou', n_iou_v)
# add pr curve
pr_curve('train/prc',
tf.cast(labels, tf.bool),
pred_scores,
num_thresholds=201)
return ce_loss