with tf.name_scope("loss"):
# Get loss
loss = residual_decoder.get_loss(predict_val=logits, real_val=y)
tf.summary.histogram("loss", loss)
# Prepare optimizer
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
optimizer = tf.train.AdamOptimizer().minimize(loss, global_step=global_step, name='optimizer')
#optimizer = tf.train.AdamOptimizer(0.0001).minimize(loss)
merged = tf.summary.merge_all()
image_batch, anno_batch, filename = input_data.read_batch(BATCH_SIZE, type = 'train')
image_batch_val, anno_batch_val, filename_val = input_data.read_batch(BATCH_SIZE, type = 'val')
with tf.Session() as sess:
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
sess.run(tf.local_variables_initializer())
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
#if os.path.exists(saved_ckpt_path):
ckpt = tf.train.get_checkpoint_state(saved_ckpt_path)
if ckpt and ckpt.model_checkpoint_path:
[64, 0, 128], [64, 64, 0], [0, 128, 192], [0, 0, 0]])
def color_gray(image):
height, width = image.shape
return_img = np.zeros([height, width, 3], np.uint8)
for i in range(height):
for j in range(width):
return_img[i, j, :] = cmap[image[i, j]]
return return_img
image_batch, anno_batch, filename = input_data.read_batch(BATCH_SIZE,
type=prediction_on,
shuffle=False)
with tf.name_scope("input"):
x = tf.placeholder(tf.float32, [BATCH_SIZE, HEIGHT, WIDTH, 3],
name='x_input')
y = tf.placeholder(tf.int32, [BATCH_SIZE, HEIGHT, WIDTH],
name='ground_truth')
keep_prob = tf.placeholder(dtype=tf.float32, name='keep_prob')
logits = denseASPP.denseASPP(x, keep_prob, train=False)
with tf.name_scope('prediction_and_miou'):
prediction = tf.argmax(logits, axis=-1, name='predictions')
softmax = tf.nn.softmax(logits, axis=-1)
predictions = tf.argmax(softmax, axis=-1, name='predictions')
train_mIoU = tf.Variable(0, dtype=tf.float32, trainable=False)
tf.summary.scalar('train_mIoU', train_mIoU)
test_mIoU = tf.Variable(0, dtype=tf.float32, trainable=False)
tf.summary.scalar('test_mIoU', test_mIoU)
merged = tf.summary.merge_all()
image_batch_0, image_batch, anno_batch, filename = input_data.read_batch(
FLAGS.batch_size,
FLAGS.height,
FLAGS.width,
FLAGS.crop_height,
FLAGS.crop_width,
FLAGS.train_random_scales,
FLAGS.scales,
FLAGS.train_random_mirror,
FLAGS.rgb_mean,
type='train')
_, image_batch_test, anno_batch_test, filename_test = input_data.read_batch(
FLAGS.batch_size,
FLAGS.height,
FLAGS.width,
FLAGS.crop_height,
FLAGS.crop_width,
FLAGS.val_random_scales,
FLAGS.scales,
FLAGS.val_random_mirror,
tf.summary.scalar('learning_rate', lr)
optimizer = tf.train.AdamOptimizer(lr).minimize(loss_all)
with tf.name_scope("mIoU"):
softmax = tf.nn.softmax(logits, axis=-1)
predictions = tf.argmax(logits, axis=-1, name='predictions')
train_mIoU = tf.Variable(0, dtype=tf.float32)
tf.summary.scalar('train_mIoU', train_mIoU)
test_mIoU = tf.Variable(0, dtype=tf.float32)
tf.summary.scalar('test_mIoU', test_mIoU)
merged = tf.summary.merge_all()
_, image_batch, anno_batch, filename = input_data.read_batch(BATCH_SIZE,
type='train')
_, image_batch_test, anno_batch_test, filename_test = input_data.read_batch(
BATCH_SIZE, type='val')
with tf.Session() as sess:
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
sess.run(tf.local_variables_initializer())
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
# if os.path.exists(saved_ckpt_path):
ckpt = tf.train.get_checkpoint_state(saved_ckpt_path)
cmap = Labels.trainId2Color
cmap[19] = (0, 0, 0) # add ignore class color
cmap[255] = (0, 0, 0)
def color_gray(image):
height, width = image.shape
return_img = np.zeros([height, width, 3], np.uint8)
for i in range(height):
for j in range(width):
return_img[i, j, :] = cmap[image[i, j]]
return return_img
image_batch_0, image_batch, anno_batch, filename = input_data.read_batch(BATCH_SIZE, type=prediction_on)
with tf.name_scope("input"):
x = tf.placeholder(tf.float32, [BATCH_SIZE, HEIGHT, WIDTH, 3], name='x_input')
y = tf.placeholder(tf.int32, [BATCH_SIZE, HEIGHT, WIDTH], name='ground_truth')
_, logits = PSPNet.PSPNet(x, is_training=False, output_stride=8, pre_trained_model=PRETRAINED_MODEL_PATH)
with tf.name_scope('prediction_and_miou'):
prediction = tf.argmax(logits, axis=-1, name='predictions')