return tf.reduce_mean(loss)
if FLAGS.dataset == 'train':
print('training on train set')
MAX_STEPS = FLAGS.max_epoches * FLAGS.train_samples // FLAGS.batch_size
train_data = input_data.read_train_data(rgb_mean=FLAGS.rgb_mean,
crop_height=FLAGS.crop_height,
crop_width=FLAGS.crop_width,
classes=FLAGS.classes,
ignore_label=FLAGS.ignore_label,
scales=FLAGS.scales)
val_data = input_data.read_val_data(rgb_mean=FLAGS.rgb_mean,
crop_height=FLAGS.crop_height,
crop_width=FLAGS.crop_width,
classes=FLAGS.classes,
ignore_label=FLAGS.ignore_label,
scales=FLAGS.scales)
elif FLAGS.dataset == 'trainval':
print('training on trainval set')
MAX_STEPS = FLAGS.max_epoches * FLAGS.trainval_samples // FLAGS.batch_size
trainval_data = input_data.read_trainval_data(
rgb_mean=FLAGS.rgb_mean,
crop_height=FLAGS.crop_height,
crop_width=FLAGS.crop_width,
classes=FLAGS.classes,
ignore_label=FLAGS.ignore_label,
scales=FLAGS.scales)
else:
raise Exception('train or trainval is needed')
initial_lr_slow = 1e-5
end_lr_slow = 1e-6
decay_steps_slow = 30000
flags.DEFINE_integer('output_stride', 16, 'output stride used in the resnet model')
if FLAGS.output_stride == 16:
MAX_STEPS = MAX_STEPS_FAST
initial_lr = initial_lr_fast
end_lr = end_lr_fast
decay_steps = decay_steps_fast
BATCH_SIZE = BATCH_SIZE_OS_16
train_data = input_data.read_train_data()
val_data = input_data.read_val_data()
elif FLAGS.output_stride == 8:
MAX_STEPS = MAX_STEPS_SLOW
initial_lr = initial_lr_slow
end_lr = end_lr_slow
decay_steps = decay_steps_slow
BATCH_SIZE = BATCH_SIZE_OS_8
train_data = input_data.read_train_raw_data()
val_data = input_data.read_val_data()
# for saved path
saved_ckpt_path = './checkpoint/'
saved_summary_train_path = './summary/train/'
saved_summary_test_path = './summary/test/'
