def test_flow_warp():
net = Net(2)
imga = Image.open(IMG1)
imgb = Image.open(IMG2)
imga = np.asarray(imga).astype('float32') / 255
imgb = np.asarray(imgb).astype('float32') / 255
imga = imga[..., [2, 1, 0]]
imgb = imgb[..., [2, 1, 0]]
with tf.Session() as sess:
net.restore(sess, checkpoint)
imga = np.expand_dims(imga, 0)
imgb = np.expand_dims(imgb, 0)
print(imga.shape, imgb.shape)
flow = net.flow_results.eval(feed_dict={
'flow/inputa:0': imga,
'flow/inputb:0': imgb,
})
x = flow
u = x[..., 0]
v = x[..., 1]
image = tf.constant(imgb[..., [2, 1, 0]])
wp = warp(image, u, v, additive_warp=True).eval()[0]
wp = (wp * 255).astype('uint8')
Image.fromarray(wp, 'RGB').show()
wp = flow_warp(image, flow).eval()[0]
wp = (wp * 255).astype('uint8')
Image.fromarray(wp, 'RGB').show()
def test_output_flow():
"""calc flow"""
net = Net(2)
imga = Image.open(IMG1)
imgb = Image.open(IMG2)
imga = np.asarray(imga).astype('float32') / 255
imgb = np.asarray(imgb).astype('float32') / 255
imga = imga[..., [2, 1, 0]]
imgb = imgb[..., [2, 1, 0]]
with tf.Session() as sess:
net.restore(sess, checkpoint)
imga = np.expand_dims(imga, 0)
imgb = np.expand_dims(imgb, 0)
print(imga.shape, imgb.shape)
flow = net.flow_results.eval(feed_dict={
'flow/inputa:0': imga,
'flow/inputb:0': imgb,
})
x = flow[0]
x = _viz_flow(x[..., 0], x[..., 1])
Image.fromarray(x, 'RGB').save('flow.png')
def test_reuse():
net = Net(2)
x1 = tf.ones([1, 128, 128, 3])
x2 = tf.ones([1, 128, 128, 3])
y1 = tf.ones([1, 128, 128, 3])
y2 = tf.ones([1, 128, 128, 3])
f1 = net(x1, x2)
f2 = net(y1, y2)
with tf.Session() as sess:
net.restore2(sess, checkpoint, 'FlowNet2')
f1 = f1.eval()
f2 = f2.eval()
assert np.all(np.abs(f1 - f2) < 1.e-6)
def main():
# Create a new network
net = FlowNet2(mode=Mode.TEST)
# Train on the data
net.test(
checkpoint='./checkpoints/FlowNet2/flownet-2.ckpt-0',
input_a_path=FLAGS.input_a,
input_b_path=FLAGS.input_b,
out_path=FLAGS.out,
)
def main():
flag = parser.parse_args()
if flag.name.lower() == 'flownet2':
net = FlowNet2(2)
ckpt = 'checkpoints/FlowNet2/flownet-2.ckpt-0'
elif flag.name.lower() == 'flownets':
net = FlowNetS(2)
ckpt = 'checkpoints/FlowNetS/flownet-S.ckpt-0'
elif flag.name.lower() == 'flownetsd':
net = FlowNetSD(2)
ckpt = 'checkpoints/FlowNetSD/flownet-SD.ckpt-0'
elif flag.name.lower() == 'flownetc':
net = FlowNetC(2)
ckpt = 'checkpoints/FlowNetC/flownet-C.ckpt-0'
elif flag.name.lower() == 'flownetcs':
net = FlowNetCS(2)
ckpt = 'checkpoints/FlowNetCS/flownet-CS.ckpt-0'
elif flag.name.lower() == 'flownetcss':
net = FlowNetCSS(2)
ckpt = 'checkpoints/FlowNetCSS/flownet-CSS.ckpt-0'
else:
return
with tf.Session() as sess:
net.restore(sess, ckpt)
tf.logging.info("exporting to {}/{}".format(flag.dir, flag.name))
builder = tf.saved_model.builder.SavedModelBuilder(flag.dir + '/' +
flag.name)
# build the signature_def_map
input_a = sess.graph.get_tensor_by_name('flow/inputa:0')
input_b = sess.graph.get_tensor_by_name('flow/inputb:0')
output = net.flow_results
inputs = {
'input_a': tf.saved_model.utils.build_tensor_info(input_a),
'input_b': tf.saved_model.utils.build_tensor_info(input_b)
}
outputs = {
'output_flow': tf.saved_model.utils.build_tensor_info(output)
}
sig = tf.saved_model.signature_def_utils.build_signature_def(
inputs=inputs,
outputs=outputs,
method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)
builder.add_meta_graph_and_variables(
sess, [tf.saved_model.tag_constants.SERVING],
signature_def_map={
tf.saved_model.signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY:
sig
},
strip_default_attrs=True)
builder.save()
from src.flowlib import flow_to_image
from src.flownet2.flownet2 import FlowNet2
from src.training_schedules import LONG_SCHEDULE
img_dir = [
y for x in os.walk('./test') for y in glob(os.path.join(x[0], '*.png'))
]
img_dir.sort()
a = cv2.imread(img_dir[0]) / 255.0
b = cv2.imread(img_dir[1]) / 255.0
inputs = {
'input_a': tf.expand_dims(tf.constant(a, dtype=tf.float32), 0),
'input_b': tf.expand_dims(tf.constant(b, dtype=tf.float32), 0)
}
model = FlowNet2(mode=Mode.TEST, debug=False)
predictions = model(inputs, LONG_SCHEDULE)
pred_flow = predictions['flow']
saver = tf.train.Saver()
with tf.Session() as sess:
#saver = tf.train.import_meta_graph('./checkpoints/FlowNet2/flownet-2.ckpt-0.meta')
saver.restore(sess, './checkpoints/FlowNet2/flownet-2.ckpt-0')
pred = sess.run(pred_flow)
print(pred.keys())
pred_flow = pred['flow']
fimg = flow_to_image(pred_flow)
cv2.imwrite('aaa.png', fimg)
def test_restore_checkpoint():
"""restore from checkpoint"""
net = Net(2) # Test Mode
with tf.Session() as sess:
net.restore(sess, checkpoint)
import os
from src.net import Mode
from src.flownet2.flownet2 import FlowNet2
from src.dataloader import load_batch
from src.dataset_configs import FLYING_CHAIRS_DATASET_CONFIG
import keras
import tensorflow as tf
net = FlowNet2()
input_a, input_b, flow = load_batch(FLYING_CHAIRS_DATASET_CONFIG, 'sample', net.global_step)
a = 0
"""
with tf.Session() as sess:
saver = tf.train.import_meta_graph('./checkpoints/FlowNet2/flownet-2.ckpt-0.meta')
saver.restore(sess, './checkpoints/FlowNet2/flownet-2.ckpt-0')
saver.
"""