def predict():
global w, h, offset, f
def parser(serialized_example):
fs = tf.parse_single_example(serialized_example,
features={
"r": tf.FixedLenFeature([9],
tf.float32),
"t": tf.FixedLenFeature([3],
tf.float32),
})
fs["r"] = tf.reshape(fs["r"], [3, 3])
fs["t"] = tf.reshape(fs["t"], [3, 1])
return fs
lod_in = tf.compat.v1.placeholder(tf.float32, shape=[], name='lod_in')
rot = tf.compat.v1.placeholder(tf.float32, shape=[3, 3], name='rotation')
tra = tf.compat.v1.placeholder(tf.float32,
shape=[3, 1],
name='translation')
testset = tf.data.TFRecordDataset(
["datasets/" + FLAGS.dataset + "/" + FLAGS.input + ".test"])
testset = testset.map(parser).repeat().batch(1).make_one_shot_iterator()
features = testset.get_next()
bg = tf.compat.v1.get_variable("Net_bg",
initializer=np.array([1, 1, 1],
dtype=np.float32),
trainable=True)
bg = tf.sigmoid(bg)
depth_init = np.random.uniform(
-5, 0,
[num_mpi,
int(h / FLAGS.subscale),
int(w / FLAGS.subscale), sub_sam]).astype(np.float32)
mpi = np.zeros([num_mpi, h, w, 4], dtype=np.float32)
mpi[0] = [1., 0., 0., .95]
mpi[1] = [1., .5, 0., .95]
mpi[2] = [1., 1., 0., .95]
mpi[3] = [.5, 1., 0., .95]
depth = tf.get_variable("Net_depth",
initializer=depth_init,
trainable=True)
depth = tf.sigmoid(depth)
#depth = tf.image.resize(depth, [h, w], align_corners=True)
with tf.compat.v1.variable_scope("Net"):
mpi = tf.compat.v1.get_variable("mpi",
initializer=mpi,
trainable=False)
mpi = tf.sigmoid(mpi)
#depth = gen_depth(mpi,False)
depth = tf.image.resize(depth, [h, w], align_corners=True)
img_out, shifts, sss = network(mpi, depth, bg, rot, tra,
FLAGS.index % mpi_max, False)
long = tf.concat(shifts, 1)
with tf.compat.v1.variable_scope("post"):
image_out = tf.clip_by_value(img_out, 0.0, 1.0)
longs = tf.clip_by_value(long, 0.0, 1.0)
config = ConfigProto()
config.gpu_options.allow_growth = True
#config = ConfigProto(device_count = {'GPU': 0})
sess = tf.compat.v1.Session(config=config)
sess.run(tf.compat.v1.global_variables_initializer())
variables_to_restore = slim.get_variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
#localpp = getLocalPath("/home2/suttisak",'./model/' + FLAGS.dataset +'/'+ FLAGS.input+str((FLAGS.index)%mpi_max))
localpp = './model/' + FLAGS.dataset + '/tem' + str(
(FLAGS.index) % mpi_max)
ckpt = tf.train.latest_checkpoint(localpp)
saver.restore(sess, ckpt)
if True:
for i in range(0, 300, 1):
try:
out0 = cv2.imread("result/frame/" + FLAGS.dataset +
"_%04d.png" %
((300 * (FLAGS.index % mpi_max) + i - 300) %
(300 * mpi_max)))
out0 = cv2.cvtColor(out0, cv2.COLOR_BGR2RGB) / 255.0
feed = sess.run(features)
if (i % 50 == 0):
print(i, ((300 * (FLAGS.index % mpi_max) + i - 300) %
(300 * mpi_max)))
out = sess.run(image_out,
feed_dict={
rot: feed["r"][0],
tra: feed["t"][0]
})
ii = i / 300
out = np.clip(out * ii + out0 * (1 - ii), 0.0, 1.0)
plt.imsave(
"result/frame/" + FLAGS.dataset + "_%04d.png" %
((300 * (FLAGS.index % mpi_max) + i - 300) %
(300 * mpi_max)), out)
except:
feed = sess.run(features)
if (i % 50 == 0):
print(i, ((300 * (FLAGS.index % mpi_max) + i - 300) %
(300 * mpi_max)))
out = sess.run(image_out,
feed_dict={
rot: feed["r"][0],
tra: feed["t"][0]
})
plt.imsave(
"result/frame/" + FLAGS.dataset + "_%04d.png" %
((300 * (FLAGS.index % mpi_max) + i - 300) %
(300 * mpi_max)), out)
if (not FLAGS.endvdo):
for i in range(0, 300, 1):
feed = sess.run(features)
if (i % 50 == 0): print(i, (300 * (FLAGS.index % mpi_max) + i))
out = sess.run(image_out,
feed_dict={
rot: feed["r"][0],
tra: feed["t"][0]
})
plt.imsave(
"result/frame/" + FLAGS.dataset + "_%04d.png" %
(300 * (FLAGS.index % mpi_max) + i), out)
cmd = 'ffmpeg -y -i ' + 'result/frame/' + FLAGS.dataset + '_%04d.png -c:v libx264 -vf "pad=ceil(iw/2)*2:ceil(ih/2)*2" -pix_fmt yuv420p result/' + FLAGS.dataset + '_moving.mp4'
print(cmd)
os.system(cmd)
if True:
webpath = "webpath/" #"/var/www/html/orbiter/"
if not os.path.exists(webpath + FLAGS.dataset):
os.system("mkdir " + webpath + FLAGS.dataset)
for ii in range(1):
ref_rt = np.array(ref_r[FLAGS.index % mpi_max +
ii:FLAGS.index % mpi_max + ii + 1])
ref_tt = np.array(ref_t[FLAGS.index % mpi_max +
ii:FLAGS.index % mpi_max + ii + 1])
print(ref_rt.shape)
ret, sublay = sess.run([mpi, sss],
feed_dict={
features['r']: ref_rt,
features['t']: ref_tt
})
sublayers = []
mpis = []
sublayers_combined = []
print("sublay", sublay[0].shape, len(sublay))
for i in range(num_mpi):
mpis.append(ret[i, :, :, :4])
ls = []
for j in range(sub_sam):
ls.append(sublay[sub_sam * i + j])
ls = np.concatenate(ls, 0)
#ls = np.expand_dims(ls,-1)
ls = np.tile(ls, (1, 1, 3))
sublayers.append(ls)
ls = np.reshape(
ls, (sub_sam, sublay[0].shape[0], sublay[0].shape[1], 3))
ls = np.clip(np.sum(ls, 0), 0.0, 1.0)
sublayers_combined.append(ls)
#out = np.rot90(out,1)
mpis = np.concatenate(mpis, 1)
plt.imsave(
webpath + FLAGS.dataset + "/mpi%02d.png" %
(FLAGS.index % mpi_max + ii), mpis)
#plt.imsave(webpath + FLAGS.dataset+ "/mpi_alpha"+str(ii)+".png", np.tile(mpis[:, :, 3:], (1, 1, 3)))
sublayers = np.concatenate(sublayers, 1)
sublayers = np.clip(sublayers, 0, 1)
plt.imsave(
webpath + FLAGS.dataset + "/sublayer%02d.png" %
(FLAGS.index % mpi_max + ii), sublayers)
#sublayers_combined = np.concatenate(sublayers_combined, 1)
#plt.imsave(webpath + FLAGS.dataset+ "/sublayers_combined.png", sublayers_combined)
plt.imsave(webpath + FLAGS.dataset + "/mpi.png", mpis)
plt.imsave(webpath + FLAGS.dataset + "/mpi_alpha.png",
np.tile(mpis[:, :, 3:], (1, 1, 3)))
namelist = "["
for ii in range(FLAGS.index % mpi_max + 1):
namelist += "\"%02d\"," % (ii)
namelist += "]"
print(namelist)
with open(
webpath + FLAGS.dataset + "/extrinsics%02d.txt" %
(FLAGS.index % mpi_max), "w") as fo:
for i in range(3):
for j in range(3):
fo.write(str(ref_r[FLAGS.index % mpi_max][i, j]) + " ")
fo.write(" ".join(
[str(x)
for x in np.nditer(ref_t[(FLAGS.index % mpi_max)])]) + "\n")
generateWebGL(webpath + FLAGS.dataset + "/index.html", w, h,
getPlanes(FLAGS.index % mpi_max), namelist, sub_sam, f,
px, py)
def predict():
global w, h, offset, f
def parser(serialized_example):
fs = tf.parse_single_example(serialized_example,
features={
"r": tf.FixedLenFeature([9],
tf.float32),
"t": tf.FixedLenFeature([3],
tf.float32),
})
fs["r"] = tf.reshape(fs["r"], [3, 3])
fs["t"] = tf.reshape(fs["t"], [3, 1])
return fs
lod_in = tf.compat.v1.placeholder(tf.float32, shape=[], name='lod_in')
rot = tf.compat.v1.placeholder(tf.float32, shape=[3, 3], name='rotation')
tra = tf.compat.v1.placeholder(tf.float32,
shape=[3, 1],
name='translation')
testset = tf.data.TFRecordDataset(
["datasets/" + FLAGS.dataset + "/" + FLAGS.input + ".test"])
testset = testset.map(parser).repeat().batch(1).make_one_shot_iterator()
features = testset.get_next()
nh = h + offset * 2
nw = w + offset * 2
depth_init = np.random.uniform(
-5, 0,
[num_mpi,
int(nh / FLAGS.subscale),
int(nw / FLAGS.subscale), sub_sam]).astype(np.float32)
int_noise = np.random.uniform(
-1, -1, [num_mpi, int(nh / 2), int(nw / 2), 4]).astype(np.float32)
mpi = np.zeros([num_mpi, nh, nw, 4], dtype=np.float32)
mpi[0] = [1., 0., 0., .95]
mpi[1] = [1., .5, 0., .95]
mpi[2] = [1., 1., 0., .95]
mpi[3] = [.5, 1., 0., .95]
with tf.compat.v1.variable_scope("Net%d" % (FLAGS.index)):
if is_gen_mpi:
noise = tf.compat.v1.get_variable("noise",
initializer=int_noise,
trainable=False)
mpic = gen_mpi(noise, lod_in, is_train=True)
else:
mpic = tf.compat.v1.get_variable("mpi_c",
initializer=mpi[:, :, :, :3],
trainable=False)
mpia1 = tf.compat.v1.get_variable("mpi_a",
initializer=mpi[:num_mpi - 1, :, :,
3:4],
trainable=False)
mpia2 = tf.compat.v1.get_variable("mpi_la",
initializer=mpi[:1, :, :, 3:4] * 0 +
5,
trainable=False)
mpia = tf.concat([mpia1, mpia2], 0)
mpi = tf.concat([mpic, mpia], -1)
mpi = tf.sigmoid(mpi)
if is_gen_depth:
if FLAGS.sublayers < 1:
depth = tf.get_variable("Net_depth",
initializer=np.ones(
(num_mpi, 3, 3, 1)),
trainable=False)
else:
depth0 = gen_depth(mpi, True)
else:
depth1 = tf.get_variable("Net_depth",
initializer=depth_init[:num_mpi - 1],
trainable=True)
depth2 = tf.get_variable(
"depth2",
initializer=depth_init[num_mpi - 1:, :, :, :sub_sam - 1],
trainable=True)
depth3 = tf.get_variable(
"depth3",
initializer=depth_init[num_mpi - 1:, :, :, :1] * 0 + 5,
trainable=False)
depth = tf.concat([depth2, depth3], -1)
depth = tf.concat([depth1, depth], 0)
#depth = tf.get_variable("Net_depth", initializer=depth_init, trainable=True)
depth0 = tf.sigmoid(depth)
depth = tf.image.resize(depth0, [nh, nw], align_corners=True)
img_out, allalpha, mask = network(mpi, depth, rot, tra, False)
with tf.compat.v1.variable_scope("post%d" % (FLAGS.index)):
image_out = tf.clip_by_value(img_out, 0.0, 1.0)
config = ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.compat.v1.Session(config=config)
sess.run(tf.compat.v1.global_variables_initializer())
variables_to_restore = slim.get_variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
localpp = './model/' + FLAGS.dataset + "/s%02d" % FLAGS.subscale
ckpt = tf.train.latest_checkpoint(localpp)
saver.restore(sess, ckpt)
if True: # make sample picture and video
webpath = "webpath/" #"/var/www/html/orbiter/"
if not os.path.exists(webpath + FLAGS.dataset +
"_s%02d" % FLAGS.subscale):
os.system("mkdir " + webpath + FLAGS.dataset +
"_s%02d" % FLAGS.subscale)
for i in range(0, 300, 1):
feed = sess.run(features)
out = sess.run(image_out,
feed_dict={
lod_in: 0,
rot: feed["r"][0],
tra: feed["t"][0]
})
if (i % 50 == 0):
print(i)
plt.imsave(
"webpath/" + FLAGS.dataset + "_s%02d" % FLAGS.subscale +
"/%04d.png" % (i), out)
plt.imsave("result/frame/" + FLAGS.dataset + "_%04d.png" % (i),
out)
cmd = 'ffmpeg -y -i ' + 'result/frame/' + FLAGS.dataset + '_%04d.png -c:v libx264 -vf "pad=ceil(iw/2)*2:ceil(ih/2)*2" -pix_fmt yuv420p webpath/' + FLAGS.dataset + "_s%02d" % FLAGS.subscale + '/moving.mp4'
print(cmd)
os.system(cmd)
if True: # make web viewer
webpath = "webpath/" #"/var/www/html/orbiter/"
if not os.path.exists(webpath + FLAGS.dataset +
"_s%02d" % FLAGS.subscale):
os.system("mkdir " + webpath + FLAGS.dataset +
"_s%02d" % FLAGS.subscale)
ret, sublay = sess.run([mpi, depth0], feed_dict={lod_in: 0})
sublayers = []
mpis = []
sublayers_combined = []
for i in range(num_mpi):
mpis.append(ret[i, :, :, :4])
ls = []
for j in range(sub_sam):
ls.append(sublay[i, :, :, j:j + 1])
ls = np.concatenate(ls, 0)
ls = np.tile(ls, (1, 1, 3))
sublayers.append(ls)
ls = np.reshape(
ls, (sub_sam, sublay[0].shape[0], sublay[0].shape[1], 3))
ls = np.clip(np.sum(ls, 0), 0.0, 1.0)
sublayers_combined.append(ls)
mpis = np.concatenate(mpis, 1)
plt.imsave(
webpath + FLAGS.dataset + "_s%02d" % FLAGS.subscale +
"/mpi%02d.png" % (FLAGS.index), mpis)
sublayers = np.concatenate(sublayers, 1)
sublayers = np.clip(sublayers, 0, 1)
plt.imsave(
webpath + FLAGS.dataset + "_s%02d" % FLAGS.subscale +
"/sublayer%02d.png" % (FLAGS.index), sublayers)
plt.imsave(
webpath + FLAGS.dataset + "_s%02d" % FLAGS.subscale + "/mpi.png",
mpis)
plt.imsave(
webpath + FLAGS.dataset + "_s%02d" % FLAGS.subscale +
"/mpi_alpha.png", np.tile(mpis[:, :, 3:], (1, 1, 3)))
namelist = "["
for ii in range(FLAGS.index + 1):
namelist += "\"%02d\"," % (ii)
namelist += "]"
print(namelist)
with open(
webpath + FLAGS.dataset + "_s%02d" % FLAGS.subscale +
"/extrinsics%02d.txt" % (FLAGS.index), "w") as fo:
for i in range(3):
for j in range(3):
fo.write(str(ref_r[i, j]) + " ")
fo.write(" ".join([str(x) for x in np.nditer(ref_t)]) + "\n")
generateWebGL(
webpath + FLAGS.dataset + "_s%02d" % FLAGS.subscale +
"/index.html", w, h, getPlanes(), namelist, sub_sam, f, px, py)
generateConfigGL(
webpath + FLAGS.dataset + "_s%02d" % FLAGS.subscale + "/config.js",
w, h, getPlanes(), namelist, sub_sam, f, px, py)
def predict():
global w, h, offset, f
def parser(serialized_example):
fs = tf.parse_single_example(
serialized_example,
features={
"r": tf.FixedLenFeature([9], tf.float32),
"t": tf.FixedLenFeature([3], tf.float32),
})
fs["r"] = tf.reshape(fs["r"], [3, 3])
fs["t"] = tf.reshape(fs["t"], [3, 1])
return fs
lod_in = tf.placeholder(tf.float32, shape=[], name='lod_in')
testset = tf.data.TFRecordDataset(["datasets/" + FLAGS.dataset + "/" + FLAGS.input + ".test"])
testset = testset.map(parser).repeat().batch(1).make_one_shot_iterator()
features = testset.get_next()
#features = {}
#features["r"] = tf.convert_to_tensor(np.array([ref_r]))
#features["t"] = tf.convert_to_tensor(np.array([ref_t]))
#features = tf.convert_to_tensor(features)
latent = np.zeros([laten_d, laten_h, laten_w, 1],dtype=np.float32)
latent[-2:laten_d,:,:,0] = 0.5
for i in range(laten_d):
ch, cw = int(laten_h/2-i/2), int(laten_w/2-i/4)
latent[i,ch-10:ch+10,cw-5:cw+5,0] = 1.0
mpi = np.zeros([num_mpi*reuse_fac, size_h, size_w, 4],dtype=np.float32)
mpi[0] = [1.,0.,0.,.95]
mpi[1] = [1.,.5,0.,.95]
mpi[2] = [1.,1.,0.,.95]
mpi[3] = [.5,1.,0.,.95]
mpi_c = np.random.uniform(-3,1,[num_mpi, size_h, size_w, 3]).astype(np.float32)
mpi_a = np.random.uniform(-3,1,[num_mpi*reuse_fac, size_h, size_w, 1]).astype(np.float32)
bg = tf.get_variable("bg", initializer=np.array([5,5,5],dtype=np.float32), trainable=True)
bg = tf.sigmoid(bg)
latent = tf.get_variable("depth", initializer=latent, trainable=False)
latent = tf.sigmoid(latent)
if (reuse_fac == 1):
#mpi = tf.get_variable("mpi", initializer=mpi, trainable=True)
mpi_c = tf.get_variable("mpic", initializer=mpi_c, trainable=True)
mpi_a = tf.get_variable("mpia", initializer=mpi_a, trainable=True)
#mpi_a = tf.round(mpi_a)
#mpi_a = tf.where(tf.greater(mpi_a,0.5),5.*tf.ones_like(mpi_a),0.*tf.ones_like(mpi_a)) + tf.where(tf.greater(mpi_a,-1.5),0.*tf.ones_like(mpi_a),-5.*tf.ones_like(mpi_a))
mpi = tf.concat([mpi_c,mpi_a],-1)
else:
mpi_c = tf.get_variable("mpic", initializer=mpi_c, trainable=True)
mpi_c = tf.tile(mpi_c, [2, 1, 1, 1])
mpi_a = tf.get_variable("mpia", initializer=mpi_a, trainable=True)
mpi = tf.concat([mpi_c,mpi_a],-1)
mpi = tf.sigmoid(mpi)
img_out, shifts, sss = network(mpi, latent, bg, features, False)
long = tf.concat(shifts, 1)
img_out = tf.clip_by_value(img_out,0.0,1.0)
long = tf.clip_by_value(long,0.0,1.0)
config = ConfigProto()
config.gpu_options.allow_growth = True
#config = ConfigProto(device_count = {'GPU': 0})
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
variables_to_restore = slim.get_variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
ckpt = tf.train.latest_checkpoint('./model/' + FLAGS.dataset +'/'+ FLAGS.input )
saver.restore(sess, ckpt)
if False:
#bug = sess.run(long)
#print(bug.shape)
#out, bug = sess.run([img_out,latent])
#print(bug.shape)
#plt.imshow(bug[:,:,:3])
#tt = np.concatenate([bug[1,:,:,0],bug[20,:,:,0],bug[40,:,:,0],bug[59,:,:,0]],1)
#plt.matshow(tt,cmap='gray')
#plt.imshow(out)
#plt.show()
print("!!")
else:
#ff = sess.run(features)
#print(fff)
if os.path.exists("result/frame"):
os.system("rm -rf result/frame")
os.makedirs("result/frame")
for i in range(200):
print(i)
out = sess.run(img_out)
#out, bug = sess.run([img_out,long])
#out = np.rot90(out,1)
#plt.imsave("result/%04d.png"%(i),bug[:,:,:3])
plt.imsave("result/frame/%04d.png"%(i),out)
cmd = 'ffmpeg -y -i ' + 'result/frame/\%04d.png -c:v libx264 -vf "pad=ceil(iw/2)*2:ceil(ih/2)*2" -pix_fmt yuv420p result/' +FLAGS.input+FLAGS.version+ '.mp4'
print(cmd)
os.system(cmd)
if True:
webpath = "/var/www/html/orbiter/"
if not os.path.exists(webpath + FLAGS.input+FLAGS.version):
os.system("mkdir " + webpath + FLAGS.input+FLAGS.version)
ret, sublay = sess.run([mpi,sss],feed_dict={features['r']:np.array([ref_r]),features['t']:np.array([ref_t])})
sublayers = []
mpis = []
sublayers_combined = []
print("sublay",sublay[0].shape,len(sublay))
for i in range(num_mpi*reuse_fac):
mpis.append(ret[i,:,:,:4])
ls = []
for j in range(sub_sam):
ls.append(sublay[sub_sam*i+j])
ls = np.concatenate(ls,0)
#ls = np.expand_dims(ls,-1)
ls = np.tile(ls,(1,1,3))
sublayers.append(ls)
ls = np.reshape(ls, (sub_sam,sublay[0].shape[0],sublay[0].shape[1],3))
ls = np.clip(np.sum(ls, 0), 0.0, 1.0)
sublayers_combined.append(ls)
#out = np.rot90(out,1)
mpis = np.concatenate(mpis, 1)
plt.imsave(webpath + FLAGS.input+FLAGS.version + "/mpi.png", mpis)
plt.imsave(webpath + FLAGS.input+FLAGS.version + "/mpi_alpha.png", np.tile(mpis[:, :, 3:], (1, 1, 3)))
sublayers = np.concatenate(sublayers, 1)
sublayers = np.clip(sublayers, 0, 1)
plt.imsave(webpath + FLAGS.input+FLAGS.version + "/sublayer.png", sublayers)
sublayers_combined = np.concatenate(sublayers_combined, 1)
plt.imsave(webpath + FLAGS.input+FLAGS.version + "/sublayers_combined.png", sublayers_combined)
with open(webpath + FLAGS.input+FLAGS.version + "/extrinsics.txt", "w") as fo:
for i in range(3):
for j in range(3):
fo.write(str(ref_r[i, j]) + " ")
fo.write(" ".join([str(x) for x in np.nditer(ref_t)]) + "\n")
generateWebGL(webpath + FLAGS.input+FLAGS.version + "/index.html", w, h, getPlanes(),sub_sam, f, px, py)
def predict():
global w, h, offset, f
def parser(serialized_example):
fs = tf.parse_single_example(serialized_example,
features={
"r": tf.FixedLenFeature([9],
tf.float32),
"t": tf.FixedLenFeature([3],
tf.float32),
})
fs["r"] = tf.reshape(fs["r"], [3, 3])
fs["t"] = tf.reshape(fs["t"], [3, 1])
return fs
lod_in = tf.placeholder(tf.float32, shape=[], name='lod_in')
rot = tf.placeholder(tf.float32, shape=[3, 3], name='rotation')
tra = tf.placeholder(tf.float32, shape=[3, 1], name='translation')
testset = tf.data.TFRecordDataset([
"datasets/" + FLAGS.dataset + "/tem" + str(FLAGS.index % mpi_max) +
".test"
])
testset = testset.map(parser).repeat().batch(1).make_one_shot_iterator()
features = testset.get_next()
bg = tf.get_variable("Net_bg",
initializer=np.array([1, 1, 1], dtype=np.float32),
trainable=True)
bg = tf.sigmoid(bg)
latent = np.zeros([laten_d, laten_h, laten_w, 1], dtype=np.float32)
latent[-2:laten_d, :, :, 0] = 0.5
for i in range(laten_d):
ch, cw = int(laten_h / 2 - i / 2), int(laten_w / 2 - i / 4)
latent[i, ch - 10:ch + 10, cw - 5:cw + 5, 0] = 1.0
mpi = np.zeros([num_mpi, h, w, 4], dtype=np.float32)
mpi[0] = [1., 0., 0., .95]
mpi[1] = [1., .5, 0., .95]
mpi[2] = [1., 1., 0., .95]
mpi[3] = [.5, 1., 0., .95]
latent = tf.get_variable("Net_depth", initializer=latent, trainable=False)
latent = tf.sigmoid(latent)
mpis = []
image_out = []
longs = []
ssss = []
for i in range(num_):
with tf.compat.v1.variable_scope("Net%d" %
((FLAGS.index + i) % mpi_max)):
mpi = tf.get_variable("mpi", initializer=mpi, trainable=True)
mpi = tf.sigmoid(mpi)
mpis.append(mpi)
img_out, shifts, sss = network(mpis[i], latent, bg, rot, tra,
(FLAGS.index + i) % mpi_max, False)
ssss.append(sss)
long = tf.concat(shifts, 1)
with tf.compat.v1.variable_scope("post%d" %
((FLAGS.index + i) % mpi_max)):
image_out.append(tf.clip_by_value(img_out, 0.0, 1.0))
longs.append(tf.clip_by_value(long, 0.0, 1.0))
config = ConfigProto()
config.gpu_options.allow_growth = True
#config = ConfigProto(device_count = {'GPU': 0})
sess = tf.Session(config=config)
sess.run(tf.global_variables_initializer())
variables_to_restore = slim.get_variables_to_restore()
saver = tf.train.Saver(variables_to_restore)
#localpp = getLocalPath("/home2/suttisak",'./model/' + FLAGS.dataset +'/'+ FLAGS.input+str((FLAGS.index)%mpi_max))
localpp = './model/' + FLAGS.dataset + '/' + FLAGS.input + str(
(FLAGS.index) % mpi_max)
ckpt = tf.train.latest_checkpoint(localpp)
saver.restore(sess, ckpt)
if FLAGS.FromFuture:
t_vars = slim.get_variables_to_restore()
vars_to_restore = [
var for var in t_vars
if 'Net%d' % ((FLAGS.index + 1) % mpi_max) in var.name
and 'Adam' not in var.name
]
#print(vars_to_restore)
saver = tf.train.Saver(vars_to_restore)
#localpp0 = getLocalPath("/home2/suttisak",'./model/' + FLAGS.dataset +'/'+ FLAGS.input+str((FLAGS.index+1)%mpi_max))
localpp0 = './model/' + FLAGS.dataset + '/' + FLAGS.input + str(
(FLAGS.index + 1) % mpi_max)
ckpt = tf.train.latest_checkpoint(localpp0)
saver.restore(sess, ckpt)
if False:
a = 0
#out, bug = sess.run([img_out,latent])
#tt = np.concatenate([bug[1,:,:,0],bug[20,:,:,0],bug[40,:,:,0],bug[59,:,:,0]],1)
#plt.matshow(tt,cmap='gray')
#plt.show()
else:
for i in range(0, 300, 1):
feed = sess.run(features)
if (i % 50 == 0): print(i)
#out0, out1 = sess.run([image_out[0],image_out[1]])
out0 = sess.run(image_out[0],
feed_dict={
rot: feed["r"][0],
tra: feed["t"][0]
})
out1 = sess.run(image_out[1],
feed_dict={
rot: feed["r"][0],
tra: feed["t"][0]
})
#out, bug = sess.run([img_out,long])
#out0 = np.rot90(out0)
#out1 = np.rot90(out1)
out = (1 - i / 300) * out0 + (
i / 300) * out1 #np.concatenate((out0,out1),1)
#plt.imsave("result/%04d.png"%(i),bug[:,:,:3])
plt.imsave(
"result/frame/" + FLAGS.dataset + "_%04d.png" %
(300 * (FLAGS.index % mpi_max) + i), out)
cmd = 'ffmpeg -y -i ' + 'result/frame/' + FLAGS.dataset + '_%04d.png -c:v libx264 -vf "pad=ceil(iw/2)*2:ceil(ih/2)*2" -pix_fmt yuv420p result/' + FLAGS.dataset + '_moving.mp4'
print(cmd)
os.system(cmd)
if False:
#webpath = "/var/www/html/orbiter/"
webpath = "webpath/"
if not os.path.exists(webpath + FLAGS.dataset):
os.system("mkdir " + webpath + FLAGS.dataset)
for ii in range(1):
ref_rt = np.array(ref_r[FLAGS.index + ii:FLAGS.index + ii + 1])
ref_tt = np.array(ref_t[FLAGS.index + ii:FLAGS.index + ii + 1])
print(ref_rt.shape)
ret, sublay = sess.run([mpis[ii], ssss[ii]],
feed_dict={
features['r']: ref_rt,
features['t']: ref_tt
})
sublayers = []
mpis = []
sublayers_combined = []
print("sublay", sublay[0].shape, len(sublay))
for i in range(num_mpi):
mpis.append(ret[i, :, :, :4])
ls = []
for j in range(sub_sam):
ls.append(sublay[sub_sam * i + j])
ls = np.concatenate(ls, 0)
#ls = np.expand_dims(ls,-1)
ls = np.tile(ls, (1, 1, 3))
sublayers.append(ls)
ls = np.reshape(
ls, (sub_sam, sublay[0].shape[0], sublay[0].shape[1], 3))
ls = np.clip(np.sum(ls, 0), 0.0, 1.0)
sublayers_combined.append(ls)
#out = np.rot90(out,1)
mpis = np.concatenate(mpis, 1)
plt.imsave(
webpath + FLAGS.dataset + "/mpi%02d.png" % (FLAGS.index + ii),
mpis)
#plt.imsave(webpath + FLAGS.dataset+ "/mpi_alpha"+str(ii)+".png", np.tile(mpis[:, :, 3:], (1, 1, 3)))
sublayers = np.concatenate(sublayers, 1)
sublayers = np.clip(sublayers, 0, 1)
plt.imsave(
webpath + FLAGS.dataset + "/sublayer%02d.png" %
(FLAGS.index + ii), sublayers)
#sublayers_combined = np.concatenate(sublayers_combined, 1)
#plt.imsave(webpath + FLAGS.dataset+ "/sublayers_combined.png", sublayers_combined)
plt.imsave(webpath + FLAGS.dataset + "/mpi.png", mpis)
plt.imsave(webpath + FLAGS.dataset + "/mpi_alpha.png",
np.tile(mpis[:, :, 3:], (1, 1, 3)))
namelist = "["
for ii in range(FLAGS.index + 1):
namelist += "\"%02d\"," % (ii)
namelist += "]"
print(namelist)
with open(
webpath + FLAGS.dataset + "/extrinsics%02d.txt" %
(FLAGS.index), "w") as fo:
for i in range(3):
for j in range(3):
fo.write(str(ref_r[FLAGS.index][i, j]) + " ")
fo.write(
" ".join([str(x)
for x in np.nditer(ref_t[(FLAGS.index)])]) + "\n")
generateWebGL(webpath + FLAGS.dataset + "/index.html", w, h,
getPlanes(), namelist, sub_sam, f, px, py)