def main():
args = get_args()
logger = get_logger()
logger.info(vars(args))
with tf.device("/gpu:0"):
if args.densenet == True:
model = DenseNet()
else:
model = CnnLstm()
tfrecord = TFRecord()
tfrecord.make_iterator(args.tfr_fname, training=False)
total = sum(1 for _ in tf.python_io.tf_record_iterator(args.tfr_fname)) // tfrecord.batch_size
saver = tf.train.Saver(tf.global_variables())
with tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) as sess:
saver.restore(sess, args.model_fname)
sess.run(tfrecord.init_op)
spec = tfrecord.load(sess, training=False)
predict = model.predict(sess, spec, args.proba)
progress_bar = tqdm(total=total, desc="[PREDICT]", unit="batch", leave=False)
while True:
try:
spec = tfrecord.load(sess, training=False)
if args.proba:
predict = np.vstack([predict, model.predict(sess, spec, args.proba)])
else:
predict = np.hstack([predict, model.predict(sess, spec, args.proba)])
progress_bar.update(1)
except tf.errors.OutOfRangeError:
break
make_submission(predict, args.sample_fname, args.output_fname, args.proba)
logger.info(f"{args.output_fname} is created.")
im = np.array(im)
im = np.delete(im, [1, 2], axis=2)
im = np.array(im) / 255.0
im = np.expand_dims(im, axis=0)
print(im.shape)
model = DenseNet(dense_blocks=5,
dense_layers=-1,
growth_rate=8,
dropout_rate=0.2,
bottleneck=True,
compression=1.0,
weight_decay=1e-4,
depth=40)
model.load_weights("outputs/model-230.h5")
lmarks = model.predict(im)
print(lmarks)
lmarks = lmarks[0]
lmarks[0:8:2] = lmarks[0:8:2] * im.shape[2]
lmarks[1:8:2] = lmarks[1:8:2] * im.shape[1]
print(lmarks)
#print(lmarks)
im = im[0] * 255
im = np.squeeze(im, axis=(2, ))
print(im.shape)
for m in range(0, 8, 2):
cv2.circle(im, (int(lmarks[m]), int(lmarks[m + 1])), 5,
(255, 255, 255), -1)
plt.figure(1, figsize=(25, 25))