def _compile_into_webpage(batch_dirs, out_html, title=None):
rows, caps, types = [], [], []
# For each batch (which has just one sample)
for batch_dir in batch_dirs:
for metadata_path in sorted(
glob(join(batch_dir, '?_metadata.json'))):
path_prefix = metadata_path[:-len('metadata.json')]
metadata = ioutil.read_json(metadata_path)
metadata = str(metadata)
row = [
metadata, path_prefix + 'base-vs-pred.apng',
path_prefix + 'gt-vs-pred.apng', path_prefix + 'nn.png'
]
rowcaps = [
"Metadata", "Prediction vs. Diffuse Base",
"Prediction vs. Ground Truth", "Nearest Neighbor"
]
rowtypes = ['text', 'image', 'image', 'image']
rows.append(row)
caps.append(rowcaps)
types.append(rowtypes)
n_rows = len(rows)
assert n_rows > 0, "No row"
# Write HTML
html = xm.vis.html.HTML()
if title is not None:
html.add_header(title)
img_table = html.add_table()
for r, rcaps, rtypes in zip(rows, caps, types):
img_table.add_row(r, rtypes, captions=rcaps)
html.save(out_html)
def __init__(self, config, mode, **kwargs):
self.data_root = config.get('DEFAULT', 'data_root')
self.data_paths = ioutil.read_json(
self.data_root.rstrip('/') + '.json')
# Because paths in JSON are relative, prepend data root directory
for _, paths in self.data_paths.items():
for k, v in paths.items():
if k != 'complete':
paths[k] = join(self.data_root, v)
super().__init__(config, mode, **kwargs)
# Trigger init. in a main thread before starting multi-threaded work.
# See http://yaqs/eng/q/6292200559345664 for details
Image.init()
def __init__(self, config, mode, **kwargs):
self.data_root = config.get('DEFAULT', 'data_root')
data_status_path = self.data_root.rstrip('/') + '.json'
if not exists(data_status_path):
raise FileNotFoundError(
("Data status JSON not found at \n\t%s\nRun "
"$REPO/data_gen/postproc.py to generate it") %
data_status_path)
self.data_paths = ioutil.read_json(data_status_path)
# Because paths in JSON are relative, prepend data root directory
for _, paths in self.data_paths.items():
for k, v in paths.items():
if k != 'complete':
paths[k] = join(self.data_root, v)
super().__init__(config, mode, **kwargs)
# Trigger init. in a main thread before starting multi-threaded work.
# See http://yaqs/eng/q/6292200559345664 for details
Image.init()
def _compile_into_video(batch_dirs, out_mp4, fps=12):
frames = {}
for batch_dir in tqdm(batch_dirs, desc="Compiling visualized batches"):
for metadata_path in glob(join(batch_dir, '?_metadata.json')):
path_prefix = metadata_path[:-len('metadata.json')]
pred_path = path_prefix + 'pred.png'
if not exists(pred_path):
logger.warn("Skipping because of missing file:\n\t%s" %
pred_path)
continue
# Metadata
metadata = ioutil.read_json(metadata_path)
id_ = metadata['id']
# Prediction
pred = xm.io.img.load(pred_path, as_array=True)
frame = pred
frames[id_] = frame
# Make video
frames_sorted = [frames[k] for k in sorted(frames)]
ioutil.write_video(frames_sorted, out_mp4, fps=fps)
def _load_data(self, id_):
if isinstance(id_, tf.Tensor):
id_ = id_.numpy().decode()
paths = self.data_paths[id_]
imh = self.config.getint('DEFAULT', 'imh')
imw = self.config.getint('DEFAULT', 'imw')
# Load images
base = xm.io.img.load(paths['diffuse'], as_array=True)[:, :, :3]
cvis = xm.io.img.load(paths['cvis'], as_array=True)
lvis = xm.io.img.load(paths['lvis'], as_array=True)
warp = ioutil.read_npy(paths['uv2cam'])
if self.mode == 'test':
rgb = np.zeros_like(base) # placeholders
rgb_camspc = np.zeros((imh, imw, 3))
else:
rgb = xm.io.img.load(paths['rgb'], as_array=True)[:, :, :3]
rgb_camspc = xm.io.img.load(paths['rgb_camspc'],
as_array=True)[:, :, :3]
# Normalize to [0, 1]
base = xm.img.normalize_uint(base)
cvis = xm.img.normalize_uint(cvis)
lvis = xm.img.normalize_uint(lvis)
if self.mode != 'test':
rgb = xm.img.normalize_uint(rgb)
rgb_camspc = xm.img.normalize_uint(rgb_camspc)
# Resize images
uvh = self.config.getint('DEFAULT', 'uvh')
base = xm.img.resize(base, new_h=uvh)
cvis = xm.img.resize(cvis, new_h=uvh)
lvis = xm.img.resize(lvis, new_h=uvh)
rgb = xm.img.resize(rgb, new_h=uvh)
rgb_camspc = xm.img.resize(rgb_camspc, new_h=imh, new_w=imw)
# NOTE: We didn't resize warp because this introduces artifacts --
# always warp first and then resize
# Neighbor diffuse base and full
nn = ioutil.read_json(paths['nn'])
nn_id = self._get_nn_id(nn)
if nn_id is None:
nn_id = 'incomplete-data_{cam}_{light}'.format(**nn)
# NOTE: When neighbor is missing, simply return black placeholders
nn_base = np.zeros_like(base)
nn_rgb = np.zeros_like(rgb)
nn_rgb_camspc = np.zeros_like(rgb_camspc)
else:
nn_base = xm.io.img.load(self.data_paths[nn_id]['diffuse'],
as_array=True)[:, :, :3]
nn_rgb = xm.io.img.load(self.data_paths[nn_id]['rgb'],
as_array=True)[:, :, :3]
nn_rgb_camspc = xm.io.img.load(
self.data_paths[nn_id]['rgb_camspc'], as_array=True)[:, :, :3]
nn_rgb_camspc = nn_rgb_camspc[:, :, :3] # discards alpha
nn_base = xm.img.normalize_uint(nn_base)
nn_rgb = xm.img.normalize_uint(nn_rgb)
nn_rgb_camspc = xm.img.normalize_uint(nn_rgb_camspc)
nn_base = xm.img.resize(nn_base, new_h=uvh)
nn_rgb = xm.img.resize(nn_rgb, new_h=uvh)
nn_rgb_camspc = xm.img.resize(nn_rgb_camspc, new_h=imh, new_w=imw)
# Return
base = base.astype(np.float32)
cvis = cvis.astype(np.float32)[:, :, None] # HxWx1
lvis = lvis.astype(np.float32)[:, :, None]
warp = warp.astype(np.float32)
rgb = rgb.astype(np.float32)
rgb_camspc = rgb_camspc.astype(np.float32)
nn_base = nn_base.astype(np.float32)
nn_rgb = nn_rgb.astype(np.float32)
nn_rgb_camspc = nn_rgb_camspc.astype(np.float32)
return \
id_, base, cvis, lvis, warp, rgb, rgb_camspc, nn_id, nn_base, \
nn_rgb, nn_rgb_camspc