annotation_files += annotation_val_files
#Know the number steps to take before decaying the learning rate and batches per epoch
num_batches_per_epoch = len(image_files) / batch_size
num_steps_per_epoch = num_batches_per_epoch
decay_steps = int(num_epochs_before_decay * num_steps_per_epoch)
#=================CLASS WEIGHTS===============================
#Median frequency balancing class_weights
if weighting == "MFB":
class_weights = median_frequency_balancing()
print "========= Median Frequency Balancing Class Weights =========\n", class_weights
#Inverse weighing probability class weights
elif weighting == "ENET":
class_weights = ENet_weighing()
print "========= ENet Class Weights =========\n", class_weights
#============= TRAINING =================
def weighted_cross_entropy(onehot_labels, logits, class_weights):
'''
A quick wrapper to compute weighted cross entropy.
------------------
Technical Details
------------------
The class_weights list can be multiplied by onehot_labels directly because the last dimension
of onehot_labels is 12 and class_weights (length 12) can broadcast across that dimension, which is what we want.
Then we collapse the last dimension for the class_weights to get a shape of (batch_size, height, width, 1)
to get a mask with each pixel's value representing the class_weight.
if combine_dataset:
image_train_files += image_val_files
gt_train_files += gt_val_files
# In[ ]:
if weighting == 'MFB':
class_weights = median_frequency_balancing(gt_train_files,
num_classes=num_class)
print('Median_frequency_balancing class weights is')
print(class_weights)
elif weighting == 'ENet':
if os.path.isfile('enet_class_weights.npy'):
class_weights = np.load('enet_class_weights.npy')
else:
class_weights = ENet_weighing(gt_train_files, num_classes=num_class)
np.save('enet_class_weights.npy', class_weights)
print('Save Enet_class_weights')
print('ENet class weights is')
#class_weights[19] = 0
print('Id Class Weight')
for i in range(int(class_weights.shape[0])):
print('{:2} {:15} {:>10.6}'.format(i, trainId2label[i].name,
class_weights[i]))
# In[ ]:
def input_pipeline(image_train_files, gt_train_files, image_val_files,
gt_val_files, image_size):
#Know the number steps to take before decaying the learning rate and batches per epoch
num_batches_per_epoch = len(image_files) / batch_size
num_steps_per_epoch = num_batches_per_epoch
decay_steps = int(num_epochs_before_decay * num_steps_per_epoch)
#=================CLASS WEIGHTS===============================
#Median frequency balancing class_weights
if weighting == "MFB":
class_weights = median_frequency_balancing(annotation_files, num_classes)
print("========= Median Frequency Balancing Class Weights =========\n",
class_weights)
#Inverse weighing probability class weights
elif weighting == "ENET":
class_weights = ENet_weighing(annotation_files, num_classes)
print("========= ENet Class Weights =========\n", class_weights)
#============= TRAINING =================
def weighted_cross_entropy(onehot_labels, logits, class_weights):
'''
A quick wrapper to compute weighted cross entropy.
------------------
Technical Details
------------------
The class_weights list can be multiplied by onehot_labels directly because the last dimension
of onehot_labels is 12 and class_weights (length 12) can broadcast across that dimension, which is what we want.
Then we collapse the last dimension for the class_weights to get a shape of (batch_size, height, width, 1)
to get a mask with each pixel's value representing the class_weight.
decay_steps = FLAGS.decay_steps
else:
decay_steps = int(num_epochs_before_decay * num_steps_per_epoch)
#=================CLASS WEIGHTS===============================
#Median frequency balancing class_weights
if weighting == "MFB":
class_weights = median_frequency_balancing(image_dir=os.path.join(
dataset_dir, dataset_name, 'trainannot'),
num_classes=num_classes)
print "========= Median Frequency Balancing Class Weights =========\n", class_weights
#Inverse weighing probability class weights
elif weighting == "ENET":
class_weights = ENet_weighing(image_dir=os.path.join(
dataset_dir, dataset_name, 'trainannot'),
num_classes=num_classes)
print "========= ENet Class Weights =========\n", class_weights
#Inverse weighing probability class weights
elif weighting == "CVPR":
class_weights = CVPR_weighing()
print "========= CVPR Class Weights =========\n", class_weights
#============= TRAINING =================
def weighted_cross_entropy(onehot_labels, logits, class_weights):
'''
A quick wrapper to compute weighted cross entropy.
------------------