def save_generated(self, data):
""" Generate images from batch and save them
"""
self.set_inputs(data["A"])
self.forward()
path = self.opt.checkpoints_dir +f"/{self.opt.name}/generated/{data['A_name']}_fake.png"
save_image(path,tensor_to_image(self.fake)[...,:3])
def _get_qrcode(self):
"""
缓存并展示登录二维码
:return:
"""
url = 'https://qr.m.jd.com/show'
payload = {
'appid': 133,
'size': 147,
't': str(int(time.time() * 1000)),
}
headers = {
'User-Agent': self.spider_session.get_user_agent(),
'Referer': 'https://passport.jd.com/new/login.aspx',
}
resp = self.session.get(url=url, headers=headers, params=payload)
if not response_status(resp):
logger.info('获取二维码失败')
return False
save_image(resp, self.qrcode_img_file)
logger.info('二维码获取成功，请打开京东APP扫描')
open_image(self.qrcode_img_file)
return True
def _get_qrcode(self):
"""
缓存并展示登录二维码
:return:
"""
url = 'https://qr.m.jd.com/show'
payload = {
'appid': 133,
'size': 300,
't': str(int(time.time() * 1000)),
}
headers = {
'User-Agent': self.spider_session.get_user_agent(),
'Referer': 'https://passport.jd.com/new/login.aspx',
}
resp = self.session.get(url=url, headers=headers, params=payload)
if not response_status(resp):
logger.info('获取二维码失败')
return False
save_image(resp, self.qrcode_img_file)
logger.info('二维码获取成功，请打开京东APP扫描')
open_image(add_bg_for_qr(self.qrcode_img_file))
if global_config.getRaw('messenger', 'email_enable') == 'true':
email.send('二维码获取成功，请打开京东APP扫描', "<img src='cid:qr_code.png'>", [email.mail_user], 'qr_code.png')
return True
def save_visuals(self, **kwargs):
""" Save recent visuals, i.e. real_A, fake_B etc., to disc
"""
visuals = self.get_visuals()
for visual, image in visuals.items():
path = self.opt.checkpoints_dir +"/{self.opt.name}/images/{visual}_{self.step}.png"
wandb.log({visual: [wandb.Image(image, caption=visual)]}, step=self.step)
save_image(path, image)
def run(path, namespace):
im = util.open_image(path)
if im is not None:
fit_image = fit(im, namespace.width, namespace.height, namespace.color,
namespace.resample)
util.save_image(fit_image, namespace.path, namespace.save_folder,
namespace.save_as, namespace.mode, "fit",
namespace.optimize, namespace.background)
def main():
args = parse_args()
if args.random_seed is not None:
# fixed random seeds for reproducibility
np.random.seed(args.random_seed)
torch.random.manual_seed(args.random_seed)
# infer target label from image
input_img = load_image(args.input_img, size=64)
labels, confidences = send_query(input_img)
target_idx = np.argmax(confidences)
target_class = labels[target_idx]
# ask user if he wants to continue
print(f'Inferred label: {target_class}, confidence of {np.round(confidences[target_idx], 3)}')
if not query_yes_no('Continue ?'):
print('Please choose an input image which the API classifies as your target class. ')
sys.exit(0)
# generate adversarial image
else:
if not args.color:
target_img = image_to_grayscale(input_img)
else:
target_img = input_img
print('Generating adversarial...')
adversarial, conf, num_queries, conv_images, cppn = generate_adversarial(target_class=target_class,
target_image=target_img,
color=args.color,
target_conf=args.target_conf,
max_queries=args.max_queries,
init=args.init)
if conf < args.target_conf:
print(f'Failed to generate an adversarial image after {args.max_queries} queries.')
# write_to_log('log.tsv', f'{target_class}\t{conf}\t{num_queries}\t{args.color}\t{args.init}')
sys.exit(0)
print(f'Found an adversarial image with > {args.target_conf} API confidence after {num_queries} queries.')
output_dir = Path(args.output_dir)
print(f'\tSaving results in: {output_dir}/')
# save adversarial image
adversarial_fname = str(output_dir / f'adversarial_{clean_filename(target_class)}_{conf}')
save_image(adversarial_fname + '.png', adversarial)
if args.high_res:
cppn.set_img_size(2000)
adversarial_high_res = cppn.render_image()
save_image(adversarial_fname + '_HD.png', adversarial_high_res)
# save convergence gif
if not args.no_gif:
conv_gif_fname = str(output_dir / f'convergence_{clean_filename(target_class)}_{conf}.gif')
save_gif_from_images(conv_gif_fname, conv_images)
# write_to_log('log.tsv', f'{target_class}\t{conf}\t{num_queries}\t{args.color}\t{args.init}')
print('Finished.')
def run(path, namespace):
im = util.open_image(path)
if im is not None:
inverted_im = invert(im)
if inverted_im is None:
return
util.save_image(inverted_im, namespace.path, namespace.save_folder,
namespace.save_as, namespace.mode, "inverted",
namespace.optimize, namespace.background)
def store_prediction(self, tst_batch, name, size=5):
pred = self.model(tst_batch, training=False)
pred = tf.math.argmax(pred, axis=-1)
img = combine_preds(tst_batch,
pred,
self.n_class,
bs=size,
resize_out=self.store_resize_f)
save_image(img, "%s/%s.png" % (self.pred_path, name))
def run(path, namespace):
im = util.open_image(path)
if im is not None:
resized_image = resize(im, namespace.width, namespace.height,
namespace.resample)
util.save_image(resized_image, path, namespace.save_folder,
namespace.save_as, namespace.mode, "resized",
namespace.optimize, namespace.background)
def save_image(result_dir, image, image_name, aspect_ratio=1.0):
im = util.tensor2im(image)
save_path = os.path.join(result_dir, image_name)
h, w, _ = im.shape
if aspect_ratio > 1.0:
im = imresize(im, (h, int(w * aspect_ratio)), interp='bicubic')
if aspect_ratio < 1.0:
im = imresize(im, (int(h / aspect_ratio), w), interp='bicubic')
util.save_image(im, save_path)
print('save in path: ', save_path)
def train(self, x, y, learning_rate=0.01):
ds = SupervisedDataSet( len(x[0]), len(y[0]) )
util.save_image(util.array_to_image(x[0]), 'out/real.png')
util.save_image(util.array_to_image(x[-1]), 'out/fake.png')
for i in range( len(x) ):
ds.addSample( x[i], y[i] )
trainer = BackpropTrainer(self.network, ds, learningrate=learning_rate)
return trainer.train()
def run(path, namespace):
im = util.open_image(path)
if im is not None:
util.save_image(im,
path,
namespace.save_folder,
namespace.save_as,
namespace.mode,
"optimized",
optimize=True,
background=namespace.background)
def train(epoch):
model.train()
train_loss = 0
train_likelihood = 0
train_state_kld = 0
for batch_idx in range(epoch_size // args.batch_size):
(states, actions) = next(train_iterator)
states = states.to(device).float()
actions = actions.to(device).float()
input_states = states[:, 0, :3]
pred_states, state_mu, state_logvar = model(input_states)
likelihood, state_kld = loss_function(pred_states, input_states,
state_mu, state_logvar)
loss = likelihood + kl_schedule(epoch) * (args.state_kl * state_kld)
optimizer.zero_grad()
loss.backward()
optimizer.step()
train_loss += loss.item()
train_likelihood += likelihood.item()
train_state_kld += state_kld.item()
if batch_idx > 0 and batch_idx % args.log_interval == 0:
print(('Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f}').format(
epoch, batch_idx * len(states), epoch_size,
100. * batch_idx / len(train_loader),
loss.item() / len(states)))
# render reconstructions
for i, inputs, pred in zip(range(min(5, len(pred_states))), input_states,
pred_states):
inputs = list(inputs.reshape([-1, 3, 64, 64]))
img_path = path.join(render_path, "recon{}_{}.png".format(epoch, i))
util.save_image(img_path, torch.cat([*inputs, pred], dim=2).detach())
# render some example observations
if epoch == 1:
for i in range(min(5, len(states))):
imgs = states[i]
img_path = path.join(render_path,
"observations{}_{}.png".format(epoch, i))
imgs = imgs.reshape([-1, 3, 64, 64])
tvu.save_image(imgs, img_path, nrow=4)
print(('====> Epoch: {} Average loss: {:.4f}'
'\tLL: {:.6f}\tstate KLD: {:.6f}').format(
epoch,
train_loss / epoch_size,
train_likelihood / epoch_size,
train_state_kld / epoch_size,
))
def test_img_save_and_load(self):
test_img = util.load_image(test_img_dir / '2.png', size=64)
self.assertEqual(type(test_img), np.ndarray)
self.assertEqual(test_img.shape, (64, 64, 3))
self.assertEqual(test_img.dtype, np.float64)
util.save_image(test_img_dir / 'test.png', test_img)
file = (test_img_dir / 'test.png')
self.assertTrue(file.exists())
self.assertTrue(file.is_file())
self.assertTrue(np.allclose(test_img, util.load_image(file)))
file.unlink()
def run(path, namespace):
## NOTE: By default, the image is grayscaled and saved as the original.
# image's mode. To save as 'L' or 'LA' the user must pass them explicitily
# with 'mode' optional argument.
im = util.open_image(path)
if im is not None:
dessaturated_im = dessaturate(im)
if dessaturated_im is None:
return
util.save_image(dessaturated_im, path, namespace.save_folder,
namespace.save_as, namespace.mode, "dessaturated",
namespace.optimize, namespace.background)
def load_image(self, shape, path, output_dir='', save=True):
''' loads an image in bgr format '''
images = np.zeros(shape, dtype='float32')
image_size = shape[2:]
in_image = scipy.misc.imread(path)
in_image = scipy.misc.imresize(in_image, (image_size[0], image_size[1]))
images[0] = np.transpose(in_image, (2, 0, 1)) # convert to (3, 227, 227) format
data = images[:,::-1] # convert from RGB to BGR
if save:
name = "%s/samples/%s.jpg" % (output_dir, 'start')
util.save_image(data, name)
return data
def get_boxes():
with tempfile.TemporaryDirectory("box-session") as tmpdir:
if request.mimetype == "image/jpeg":
image_path = os.path.join(tmpdir, "image.jpeg")
data = request.get_data()
elif request.mimetype == "application/json":
json_data = json.loads(request.get_data())
data = base64.decodebytes(json_data["data"].encode())
extension = json_data["type"]
image_path = os.path.join(tmpdir, "image.png")
else:
return "Unsupported MIME type: `{}`".format(request.mimetype)
util.save_image(image_path, data)
return flask.jsonify(google_query.get_boxes(image_path))
def main(args):
"""
see https://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_ml/py_kmeans/py_kmeans_opencv/py_kmeans_opencv.html
"""
if args.like is not None:
_, newcolors = cluster_image(args.likefile, args.k)
else:
newcolors = color_selector(args.colors, args.colorfile)
K = len(newcolors)
labels, _ = cluster_image(args.infile, len(newcolors))
res = newcolors[labels.flatten()]
img = load_image(args.infile)
save_image(args.outfile, res, img)
def validate(self, sess, test_size, x_provider, y_provider, n_epoch):
self.net.is_training = False
mean_dice = 0
mean_iou = 0
for val_index in range(test_size):
val_x, val_y = sess.run([x_provider, y_provider])
prediction, loss = sess.run([self.net.predicter, self.net.loss],
feed_dict={
self.net.x: val_x,
self.net.y: val_y,
})
img_predict = prediction.reshape((self.nx, self.ny, 4))
img_predict = self.remove_minor_cc(img_predict, 0.3, 4)
img_y = val_y.reshape((self.nx * self.ny, 4))
img_predict = img_predict.reshape((self.nx * self.ny, 4))
dice, iou = sess.run([self.net.dice, self.net.iou],
feed_dict={
self.net.img_pred: img_predict,
self.net.img_y: img_y
})
dice1 = dice[0]
dice2 = dice[1]
dice3 = dice[2]
dice = (dice1 + dice2 + dice3) / 3
iou1 = iou[0]
iou2 = iou[1]
iou3 = iou[2]
iou = (iou1 + iou2 + iou3) / 3
mean_dice += dice
mean_iou += iou
if val_index == 0:
img = util.combine_img_prediction(val_x, img_y[:, 1],
img_predict[:, 1],
img_y[:, 2], img_predict[:,
2],
img_y[:, 3], img_predict[:,
3])
util.save_image(
img,
"%s/%s.jpg" % (self.prediction_path, "epoch_%s" % n_epoch +
'_%s' % val_index + '_%s' % dice))
mean_dice = mean_dice / test_size
mean_iou = mean_iou / test_size
logging.info(
"Validation Loss: {:.4f}, Dice: {:.4f}, IoU: {:.4f}".format(
loss, mean_dice, mean_iou))
return mean_dice, mean_iou
def run_kmeans(data_path):
img = imread(data_path)
img2 = img.reshape(img.shape[0] * img.shape[1], img.shape[2])
# Cluster for K = 2,3,...,16 save the image to see differences
for cluster_count in range(2, 17):
km = KMeans(cluster_count)
V, cmap = km.fit(img2)
if cluster_count in [2, 4, 6, 8, 16]:
save_image(
V,
cmap,
img.shape,
"result/kmeans_{0}_clusters.png".format(cluster_count),
)
def store_prediction(self, sess, batch_x, batch_y, name):
prediction = sess.run(self.net.predicter, feed_dict={self.net.x: batch_x,
self.net.y: batch_y,
self.net.keep_prob: 1.})
pred_shape = prediction.shape
loss = sess.run(self.net.cost, feed_dict={self.net.x: batch_x,
self.net.y: batch_y,
self.net.keep_prob: 1.})
logging.info("Validation loss={:.4f}".format(loss))
img = util.combine_img_prediction(batch_x, batch_y, prediction)
util.save_image(img, "%s/%s.jpg" % (self.prediction_path, name))
return pred_shape
def main():
model_name = 'generator_e_59'
model_name = os.path.join(pp.MODEL_SAVES, model_name)
model = Generator()
chainer.serializers.load_npz(model_name, model)
num_features = util.get_number_of_features(pp.CELEB_FACES_FC6_TEST)
all_names = np.array(util.get_names_h5_file(pp.FC6_TEST_H5))
y_tmp = np.zeros((num_features, 32 * 32 * 3), dtype=np.float32)
target_tmp = np.zeros((num_features, 32 * 32 * 3), dtype=np.float32)
save_list_names = os.listdir('/home/gabi/Documents/temp_datasets/test_celeba_reconstruction_m99')
save_list_names = [i.split('_')[0]+'.jpg' for i in save_list_names]
# save_list = random.sample(xrange(num_features), 100)
# save_list_names = [''] * 100
cnt = 0
# for i in save_list:
# save_list_names[cnt] = util.sed_line(pp.CELEB_FACES_FC6_TEST, i).strip().split(',')[0]
# cnt += 1
cnt = 0
for i in all_names:
features = util.get_features_h5_in_batches([i], train=False)
features = util.to_correct_input(features)
labels = util.get_labels([i])
labels = np.asarray(labels, dtype=np.float32)
target_tmp[cnt] = labels
with chainer.using_config('train', False):
f = np.expand_dims(features[0], 0)
prediction = model(f)
y_tmp[cnt] = prediction.data[0]
if i in save_list_names:
util.save_image(prediction, i, epoch=0)
print("image '%s' saved" % i)
cnt += 1
# calculate validation loss
y_tmp.astype(np.float32)
target_tmp.astype(np.float32)
loss = chainer.functions.mean_absolute_error(y_tmp, target_tmp)
print('model: ', model_name, ' loss model: ', loss)
def set_question_values(manipulated_question):
manipulated_question["title"] = request.form["title"]
manipulated_question["submission_time"] = util.get_new_timestamp()
manipulated_question["message"] = request.form["description"]
manipulated_question["user_id"] = data_manager.get_user_data(session['username'])['id']
if 'file' in request.files:
file = request.files['file']
manipulated_question["image"] = util.save_image(file, data_manager.UPLOAD_FOLDER, "question")
return manipulated_question
def save_images(self, webpage, visuals, image_path):
image_dir = webpage.get_image_dir()
name = ntpath.basename(image_path)
webpage.add_header(name)
ims = []
txts = []
links = []
for label, image_numpy in visuals.items():
image_name = '%s_%s.png' % (name, label)
save_path = os.path.join(image_dir, image_name)
util.save_image(image_numpy, save_path)
ims.append(image_name)
txts.append(label)
links.append(image_name)
webpage.add_images(ims, txts, links, width=self.win_size)
def render(self,
optimization_tensor,
input_img,
output_filename='test.png',
iterations=20):
t_score = tf.reduce_mean(
optimization_tensor) # defining the optimization objective
t_grad = tf.gradients(t_score, self.input_tensor)[
0] # behold the power of automatic differentiation!
img = input_img.copy()
img = self.image_optimize(img, iterations, t_grad)
if img is not None:
img = self.norm_visualize(img)
img = np.clip(img, -1, 1)
save_image(img, output_filename, color=self.save_color)
def estimate_foregrounds(directory):
from pymatting import estimate_foreground_ml
# Add your own method here (method name, estimate_foreground function)
fg_methods = [
("multilevel", estimate_foreground_multilevel),
#("naive", estimate_foreground_naive),
]
alpha_methods = [
"gt",
"cf",
"idx",
"ifm",
"knn",
]
print("Running foreground estimation")
for index in range(1, 28):
name = "GT%02d" % index
path = f"{directory}/converted/image/{name}.bmp"
image = util.load_image(path)
for alpha_method in alpha_methods:
path = f"{directory}/alpha/{alpha_method}/{name}.png"
alpha = util.load_image(path, "gray")
for fg_method, estimate_foreground in fg_methods:
foreground = estimate_foreground(image, alpha)
print(
f'Processing image {name} with foreground estimation '
f'method {fg_method:10} and alpha matte {alpha_method:3}')
path = f"{directory}/fg_methods/{fg_method}/{alpha_method}/{name}.bmp"
util.save_image(path, foreground)
def output_minibatch_stats(self, sess, summary_writer, step, batch_x,
batch_y, epoch):
# Calculate batch loss and accuracy
summary_str, loss, acc, predictions = sess.run([
self.summary_op, self.net.cost, self.net.accuracy,
self.net.predicter
],
feed_dict={
self.net.x: batch_x,
self.net.y: batch_y,
self.net.keep_prob:
1.
})
img_pred = util.to_rgb(np.asarray(predictions[0], 'float'))
util.save_image(img_pred, "test_mini/%s_%s_pred.jpg" % (epoch, step))
summary_writer.add_summary(summary_str, step)
summary_writer.flush()
logging.info(
"Iter {:}, Minibatch Loss= {:.4f}, Training Accuracy= {:.4f}%, Minibatch error= {:.1f}%"
.format(step, loss, acc * 100.0, error_rate(predictions, batch_y)))
def evaluate(self, net, iteration, noise_func):
avg_psnr = 0.0
for idx in range(len(self.images)):
orig_img = self.images[idx]
w = orig_img.shape[2]
h = orig_img.shape[1]
noisy_img = noise_func(orig_img)
pred255 = util.infer_image(net, noisy_img)
orig255 = util.clip_to_uint8(orig_img)
assert (pred255.shape[2] == w and pred255.shape[1] == h)
sqerr = np.square(
orig255.astype(np.float32) - pred255.astype(np.float32))
s = np.sum(sqerr)
cur_psnr = 10.0 * np.log10((255 * 255) / (s / (w * h * 3)))
avg_psnr += cur_psnr
util.save_image(self.submit_config, pred255,
"img_{0}_val_{1}_pred.png".format(iteration, idx))
if iteration == 0:
util.save_image(
self.submit_config, orig_img,
"img_{0}_val_{1}_orig.png".format(iteration, idx))
util.save_image(
self.submit_config, noisy_img,
"img_{0}_val_{1}_noisy.png".format(iteration, idx))
avg_psnr /= len(self.images)
print('Average PSNR: %.2f' % autosummary('PSNR_avg_psnr', avg_psnr))
def test2d(config):
prepare_dirs_and_logger(config)
tf.set_random_seed(config.random_seed)
batch_manager = BatchManager(config)
# thread test
sess_config = tf.ConfigProto()
sess_config.gpu_options.allow_growth = True
sess_config.allow_soft_placement = True
sess_config.log_device_placement = False
sess = tf.Session(config=sess_config)
batch_manager.start_thread(sess)
x, y = batch_manager.batch() # [-1, 1]
x_ = x.eval(session=sess)
# y_ = y.eval(session=sess)
batch_manager.stop_thread()
x_w = vort_np(x_)
x_w /= np.abs(x_w).max()
x_w = (x_w + 1) * 0.5
x_w = np.uint8(plt.cm.RdBu(x_w[..., 0]) * 255)[..., :3]
x_ = (x_ + 1) * 127.5 # [0, 255]
b_ch = np.ones([config.batch_size, config.res_y, config.res_x, 1]) * 127.5
x_ = np.concatenate((x_, b_ch), axis=-1)
x_ = np.concatenate((x_, x_w), axis=0)
save_image(x_, '{}/x_fixed.png'.format(config.model_dir))
# random pick from parameter space
x, pi, zi = batch_manager.random_list(config.batch_size)
x_w = vort_np(x / 127.5 - 1)
x_w /= np.abs(x_w).max()
x_w = (x_w + 1) * 0.5
x_w = np.uint8(plt.cm.RdBu(x_w[..., 0]) * 255)[..., :3]
x = np.concatenate((x, x_w), axis=0)
save_image(x, '{}/x.png'.format(config.model_dir))
with open('{}/x_p.txt'.format(config.model_dir), 'w') as f:
f.write(str(pi))
f.write(str(zi))
def main():
p = argparse.ArgumentParser(description='Display a kernel.')
p.add_argument('-out', help='output to *.png file instead of viewing')
p.add_argument('k', nargs='*', help='path to kernel(s)')
args = p.parse_args()
out = None
for fn in args.k:
print('Loading', fn)
step, kernel = util.load_kernel(fn)
print(' Step', step)
print(' Kernel shape is', kernel.shape)
print(' Min', np.min(kernel))
print(' Max', np.max(kernel))
print(' Mean', np.mean(kernel))
print(' Sum', np.sum(kernel))
print(' Sum of abs', np.sum(np.abs(kernel)))
print(' RMS', np.sqrt(np.mean(kernel * kernel)))
render = util.vis_hwoi(kernel, doubles=2)
render = util.hstack([render, util.make_label(fn)], 5)
if out is None:
out = render
else:
out = util.vstack([out, render], 5)
out = util.border(out, 5)
if args.out is not None:
util.save_image(args.out, out)
print('Written to', args.out)
else:
print('Press ESC to close window.')
def render_fn():
return out
util.viewer(None, render_fn)
def _get_QRcode(self):
url = 'https://qr.m.jd.com/show'
payload = {
'appid': 133,
'size': 147,
't': str(int(time.time() * 1000)),
}
headers = {
'User-Agent': self.default_user_agent,
'Referer': 'https://passport.jd.com/new/login.aspx',
}
resp = self.session.get(url=url, headers=headers, params=payload)
if not resp.ok:
logger.info('获取二维码失败')
return False
QRCode_file = 'QRcode.png'
save_image(resp, QRCode_file)
logger.info('二维码获取成功，请打开京东APP扫描')
open_image(QRCode_file)
return True
def _run_inference(output_dir=None,
file_extension='png',
depth=True,
egomotion=False,
model_ckpt=None,
input_dir=None,
input_list_file=None,
batch_size=1,
img_height=128,
img_width=416,
seq_length=3,
architecture=nets.RESNET,
imagenet_norm=True,
use_skip=True,
joint_encoder=True,
shuffle=False,
flip_for_depth=False,
inference_mode=INFERENCE_MODE_SINGLE,
inference_crop=INFERENCE_CROP_NONE,
use_masks=False):
"""Runs inference. Refer to flags in inference.py for details."""
inference_model = model.Model(is_training=False,
batch_size=batch_size,
img_height=img_height,
img_width=img_width,
seq_length=seq_length,
architecture=architecture,
imagenet_norm=imagenet_norm,
use_skip=use_skip,
joint_encoder=joint_encoder)
vars_to_restore = util.get_vars_to_save_and_restore(model_ckpt)
saver = tf.train.Saver(vars_to_restore)
sv = tf.train.Supervisor(logdir='/tmp/', saver=None)
with sv.managed_session() as sess:
saver.restore(sess, model_ckpt)
if not gfile.Exists(output_dir):
gfile.MakeDirs(output_dir)
logging.info('Predictions will be saved in %s.', output_dir)
# Collect all images to run inference on.
im_files, basepath_in = collect_input_images(input_dir, input_list_file,
file_extension)
if shuffle:
logging.info('Shuffling data...')
np.random.shuffle(im_files)
logging.info('Running inference on %d files.', len(im_files))
# Create missing output folders and pre-compute target directories.
output_dirs = create_output_dirs(im_files, basepath_in, output_dir)
# Run depth prediction network.
if depth:
im_batch = []
for i in range(len(im_files)):
if i % 100 == 0:
logging.info('%s of %s files processed.', i, len(im_files))
# Read image and run inference.
if inference_mode == INFERENCE_MODE_SINGLE:
if inference_crop == INFERENCE_CROP_NONE:
im = util.load_image(im_files[i], resize=(img_width, img_height))
elif inference_crop == INFERENCE_CROP_CITYSCAPES:
im = util.crop_cityscapes(util.load_image(im_files[i]),
resize=(img_width, img_height))
elif inference_mode == INFERENCE_MODE_TRIPLETS:
im = util.load_image(im_files[i], resize=(img_width * 3, img_height))
im = im[:, img_width:img_width*2]
if flip_for_depth:
im = np.flip(im, axis=1)
im_batch.append(im)
if len(im_batch) == batch_size or i == len(im_files) - 1:
# Call inference on batch.
for _ in range(batch_size - len(im_batch)): # Fill up batch.
im_batch.append(np.zeros(shape=(img_height, img_width, 3),
dtype=np.float32))
im_batch = np.stack(im_batch, axis=0)
est_depth = inference_model.inference_depth(im_batch, sess)
if flip_for_depth:
est_depth = np.flip(est_depth, axis=2)
im_batch = np.flip(im_batch, axis=2)
for j in range(len(im_batch)):
color_map = util.normalize_depth_for_display(
np.squeeze(est_depth[j]))
visualization = np.concatenate((im_batch[j], color_map), axis=0)
# Save raw prediction and color visualization. Extract filename
# without extension from full path: e.g. path/to/input_dir/folder1/
# file1.png -> file1
k = i - len(im_batch) + 1 + j
filename_root = os.path.splitext(os.path.basename(im_files[k]))[0]
pref = '_flip' if flip_for_depth else ''
output_raw = os.path.join(
output_dirs[k], filename_root + pref + '.npy')
output_vis = os.path.join(
output_dirs[k], filename_root + pref + '.png')
with gfile.Open(output_raw, 'wb') as f:
np.save(f, est_depth[j])
util.save_image(output_vis, visualization, file_extension)
im_batch = []
# Run egomotion network.
if egomotion:
if inference_mode == INFERENCE_MODE_SINGLE:
# Run regular egomotion inference loop.
input_image_seq = []
input_seg_seq = []
current_sequence_dir = None
current_output_handle = None
for i in range(len(im_files)):
sequence_dir = os.path.dirname(im_files[i])
if sequence_dir != current_sequence_dir:
# Assume start of a new sequence, since this image lies in a
# different directory than the previous ones.
# Clear egomotion input buffer.
output_filepath = os.path.join(output_dirs[i], 'egomotion.txt')
if current_output_handle is not None:
current_output_handle.close()
current_sequence_dir = sequence_dir
logging.info('Writing egomotion sequence to %s.', output_filepath)
current_output_handle = gfile.Open(output_filepath, 'w')
input_image_seq = []
im = util.load_image(im_files[i], resize=(img_width, img_height))
input_image_seq.append(im)
if use_masks:
im_seg_path = im_files[i].replace('.%s' % file_extension,
'-seg.%s' % file_extension)
if not gfile.Exists(im_seg_path):
raise ValueError('No segmentation mask %s has been found for '
'image %s. If none are available, disable '
'use_masks.' % (im_seg_path, im_files[i]))
input_seg_seq.append(util.load_image(im_seg_path,
resize=(img_width, img_height),
interpolation='nn'))
if len(input_image_seq) < seq_length: # Buffer not filled yet.
continue
if len(input_image_seq) > seq_length: # Remove oldest entry.
del input_image_seq[0]
if use_masks:
del input_seg_seq[0]
input_image_stack = np.concatenate(input_image_seq, axis=2)
input_image_stack = np.expand_dims(input_image_stack, axis=0)
if use_masks:
input_image_stack = mask_image_stack(input_image_stack,
input_seg_seq)
est_egomotion = np.squeeze(inference_model.inference_egomotion(
input_image_stack, sess))
egomotion_str = []
for j in range(seq_length - 1):
egomotion_str.append(','.join([str(d) for d in est_egomotion[j]]))
current_output_handle.write(
str(i) + ' ' + ' '.join(egomotion_str) + '\n')
if current_output_handle is not None:
current_output_handle.close()
elif inference_mode == INFERENCE_MODE_TRIPLETS:
written_before = []
for i in range(len(im_files)):
im = util.load_image(im_files[i], resize=(img_width * 3, img_height))
input_image_stack = np.concatenate(
[im[:, :img_width], im[:, img_width:img_width*2],
im[:, img_width*2:]], axis=2)
input_image_stack = np.expand_dims(input_image_stack, axis=0)
if use_masks:
im_seg_path = im_files[i].replace('.%s' % file_extension,
'-seg.%s' % file_extension)
if not gfile.Exists(im_seg_path):
raise ValueError('No segmentation mask %s has been found for '
'image %s. If none are available, disable '
'use_masks.' % (im_seg_path, im_files[i]))
seg = util.load_image(im_seg_path,
resize=(img_width * 3, img_height),
interpolation='nn')
input_seg_seq = [seg[:, :img_width], seg[:, img_width:img_width*2],
seg[:, img_width*2:]]
input_image_stack = mask_image_stack(input_image_stack,
input_seg_seq)
est_egomotion = inference_model.inference_egomotion(
input_image_stack, sess)
est_egomotion = np.squeeze(est_egomotion)
egomotion_1_2 = ','.join([str(d) for d in est_egomotion[0]])
egomotion_2_3 = ','.join([str(d) for d in est_egomotion[1]])
output_filepath = os.path.join(output_dirs[i], 'egomotion.txt')
file_mode = 'w' if output_filepath not in written_before else 'a'
with gfile.Open(output_filepath, file_mode) as current_output_handle:
current_output_handle.write(str(i) + ' ' + egomotion_1_2 + ' ' +
egomotion_2_3 + '\n')
written_before.append(output_filepath)
logging.info('Done.')
def finetune_inference(train_model, model_ckpt, output_dir):
"""Train model."""
vars_to_restore = None
if model_ckpt is not None:
vars_to_restore = util.get_vars_to_save_and_restore(model_ckpt)
ckpt_path = model_ckpt
pretrain_restorer = tf.train.Saver(vars_to_restore)
sv = tf.train.Supervisor(logdir=None, save_summaries_secs=0, saver=None,
summary_op=None)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
img_nr = 0
failed_heuristic = []
with sv.managed_session(config=config) as sess:
# TODO(casser): Caching the weights would be better to avoid I/O bottleneck.
while True: # Loop terminates when all examples have been processed.
if model_ckpt is not None:
logging.info('Restored weights from %s', ckpt_path)
pretrain_restorer.restore(sess, ckpt_path)
logging.info('Running fine-tuning, image %s...', img_nr)
img_pred_folder = os.path.join(
output_dir, FLAGS.ft_name + 'id_' + str(img_nr))
if not gfile.Exists(img_pred_folder):
gfile.MakeDirs(img_pred_folder)
step = 1
# Run fine-tuning.
while step <= FLAGS.num_steps:
logging.info('Running step %s of %s.', step, FLAGS.num_steps)
fetches = {
'train': train_model.train_op,
'global_step': train_model.global_step,
'incr_global_step': train_model.incr_global_step
}
_ = sess.run(fetches)
if step % SAVE_EVERY == 0:
# Get latest prediction for middle frame, highest scale.
pred = train_model.depth[1][0].eval(session=sess)
if FLAGS.flip:
pred = np.flip(pred, axis=2)
input_img = train_model.image_stack.eval(session=sess)
input_img_prev = input_img[0, :, :, 0:3]
input_img_center = input_img[0, :, :, 3:6]
input_img_next = input_img[0, :, :, 6:]
img_pred_file = os.path.join(
img_pred_folder,
str(step).zfill(10) + ('_flip' if FLAGS.flip else '') + '.npy')
motion = np.squeeze(train_model.egomotion.eval(session=sess))
# motion of shape (seq_length - 1, 6).
motion = np.mean(motion, axis=0) # Average egomotion across frames.
if SAVE_PREVIEWS or step == FLAGS.num_steps:
# Also save preview of depth map.
color_map = util.normalize_depth_for_display(
np.squeeze(pred[0, :, :]))
visualization = np.concatenate(
(input_img_prev, input_img_center, input_img_next, color_map))
motion_s = [str(m) for m in motion]
s_rep = ','.join(motion_s)
with gfile.Open(img_pred_file.replace('.npy', '.txt'), 'w') as f:
f.write(s_rep)
util.save_image(
img_pred_file.replace('.npy', '.%s' % FLAGS.file_extension),
visualization, FLAGS.file_extension)
with gfile.Open(img_pred_file, 'wb') as f:
np.save(f, pred)
# Apply heuristic to not finetune if egomotion magnitude is too low.
ego_magnitude = np.linalg.norm(motion[:3], ord=2)
heuristic = ego_magnitude >= FLAGS.egomotion_threshold
if not heuristic and step == FLAGS.num_steps:
failed_heuristic.append(img_nr)
step += 1
img_nr += 1
return failed_heuristic
return numpy.array(xr), numpy.array(yr)
def log(s):
print(s)
sys.stdout.flush()
if __name__ == '__main__':
log('Loading dataset ...')
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
real_x = mnist.train.images[:1]
real_y = numpy.array([ numpy.array([1.0]) for x in real_x ])
for i in range(1):
image = util.array_to_image(real_x[i])
util.save_image(image, 'out/real_%d.png' % i)
log('Constructing D ...')
#d = cnn.CNN()
#d = evo_mlp.EvoMLP()
d = pybrain_mlp.PyBrainMLP([784, 128, 1])
log('Constructing G ...')
def fitness_func(generator, genome):
x = numpy.array([ util.generate_image(genome).flatten() for x in range(5)])
y = [ v[0] for v in d.forward(x) ]
return sum(y) / float( len(y) )
g = cppn.CPPN('mnist_config', fitness_func)
log('Initializing G ...')
from pybrain_mlp import PyBrainMLP
from tensorflow.examples.tutorials.mnist import input_data
import util
if __name__ == '__main__':
mnist = input_data.read_data_sets('MNIST_data', one_hot=True)
real = mnist.train.images[0]
fake = [ 0.0 for x in real ]
util.save_image(util.array_to_image(real), 'out/real.png')
util.save_image(util.array_to_image(fake), 'out/fake.png')
x = [ real, fake ]
y = [ [1.0], [0.0] ]
mlp = PyBrainMLP([784, 1])
for i in range(10000):
loss = mlp.train(x, y)
if i % 100 == 0:
print("%d: %f" % (i, loss))
if i % 100 == 0:
result = mlp.forward(x)
print('Real: %s' % result[0])
print('Fake: %s' % result[1])