def _dualloss(logits, mid_logits, labels, class_hist, num_labels, is_training=True):
#loss1 = losses.cross_entropy_loss(logits, labels, weights, num_labels)
#loss2 = losses.cross_entropy_loss(mid_logits, labels, weights, num_labels)
#max_weight = 10
max_weight = 1
loss1 = losses.weighted_cross_entropy_loss(logits, labels, class_hist,
max_weight=max_weight)
loss2 = losses.weighted_cross_entropy_loss(mid_logits, labels, class_hist,
max_weight=max_weight)
#loss1 = losses.weighted_cross_entropy_loss_dense(logits, labels, weights, num_labels,
# max_weight=max_weight)
#loss2 = losses.weighted_cross_entropy_loss_dense(mid_logits, labels, weights, num_labels,
# max_weight=max_weight)
#wgt = 0.4
#xent_loss = loss1 + wgt * loss2
wgt = 0.3 # best
#wgt = 0.2
#wgt = 0.4
xent_loss = (1-wgt)*loss1 + wgt*loss2
all_losses = [xent_loss]
# get losses + regularization
total_loss = losses.total_loss_sum(all_losses)
if is_training:
loss_averages_op = losses.add_loss_summaries(total_loss)
with tf.control_dependencies([loss_averages_op]):
total_loss = tf.identity(total_loss)
return total_loss
def _multiloss(logits, aux_logits, labels, num_labels, class_hist, is_training):
max_weight = FLAGS.max_weight
xent_loss = 0
#main_wgt = 0.6
if len(aux_logits) > 0:
main_wgt = 0.7
aux_wgt = (1 - main_wgt) / len(aux_logits)
else:
main_wgt = 1.0
aux_wgt = 0
xent_loss = main_wgt * losses.weighted_cross_entropy_loss(
logits, labels, class_hist, max_weight=max_weight)
for i, l in enumerate(aux_logits):
print('loss' + str(i), ' --> ' , l)
xent_loss += aux_wgt * losses.weighted_cross_entropy_loss(
l, labels, class_hist, max_weight=max_weight)
all_losses = [xent_loss]
# get losses + regularization
total_loss = losses.total_loss_sum(all_losses)
if is_training:
loss_averages_op = losses.add_loss_summaries(total_loss)
with tf.control_dependencies([loss_averages_op]):
total_loss = tf.identity(total_loss)
return total_loss
def inference(image, labels=None, constant_shape=True, is_training=False):
global known_shape
known_shape = constant_shape
x = normalize_input(image)
logits, aux_logits = _build(x, is_training=is_training)
if labels:
main_wgt = 0.7
xent_loss = main_wgt * losses.weighted_cross_entropy_loss(logits, labels)
xent_loss = (1-main_wgt) * losses.weighted_cross_entropy_loss(aux_logits, labels)
return logits, aux_logits, xent_loss
return logits, aux_logits
def get_loss(logits, labels, weights, is_training):
#xent_loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits, labels))
xent_loss = losses.weighted_cross_entropy_loss(logits, labels, weights)
total_loss = total_loss_sum([xent_loss])
if is_training:
loss_averages_op = losses.add_loss_summaries(total_loss)
with tf.control_dependencies([loss_averages_op]):
total_loss = tf.identity(total_loss)
return total_loss
def train_step(x, y_true_labels):
print('Tracing training step...')
# Forward pass
with tf.GradientTape() as tape:
y_pred_logits = network(
x) # (batch_size, img_h, img_w, num_classes)
loss = weighted_cross_entropy_loss(y_true_labels, y_pred_logits,
dataset.class_weights)
# Backward pass
grads = tape.gradient(loss, network.trainable_variables)
opt.apply_gradients(zip(grads, network.trainable_variables))
return loss
