def __init__(self, args, init_train_info={}, sub_dir=None):
self.args = args
misc.ensure_dir(args.logdir)
sub_dir = args.continue_from or sub_dir or misc.datetimestr()
self.logdir = os.path.join(args.logdir, sub_dir)
misc.ensure_dir(self.logdir)
self._setup_log_file()
self._create_train_info(args, init_train_info)
def transfer(data_16000_dir, data_dir, args):
misc.ensure_dir(data_dir, erase_old=True)
for file in os.listdir(data_16000_dir):
in_folder = os.path.join(data_16000_dir, file)
if not os.path.isdir(in_folder): continue
out_folder = os.path.join(data_dir, file + '.pth')
misc.ensure_dir(out_folder, erase_old=True)
transfer_folder(in_folder, out_folder, args)
def transfer(raw_data_dir, data_dir, sample_rate):
import misc
misc.ensure_dir(data_dir, erase_old=True)
for file in os.listdir(raw_data_dir):
in_folder = os.path.join(raw_data_dir, file)
if not os.path.isdir(in_folder): continue
out_folder = os.path.join(data_dir, file)
misc.ensure_dir(out_folder, erase_old=True)
transfer_folder(in_folder, out_folder, sample_rate)
def save_errors(_error_sign, _scene_errs):
# Save the calculated errors to a JSON file.
errors_path = p['out_errors_tpath'].format(
eval_path=p['eval_path'],
result_name=result_name,
error_sign=_error_sign,
scene_id=scene_id)
misc.ensure_dir(os.path.dirname(errors_path))
misc.log('Saving errors to: {}'.format(errors_path))
inout.save_json(errors_path, _scene_errs)
def end_log_game_action(self, game_name):
import json
if game_name not in self._game_info: return
game = self._game_info.get(game_name)
if 'actions' in game:
file_path = os.path.join(self.logdir, 'game_info')
misc.ensure_dir(file_path)
file_path = os.path.join(file_path, game_name + '.actions.json')
misc.save_file(file_path, json.dumps(game['actions']))
del self._game_info[game_name]
def create_manifest(manifest_file, data_files):
data = sort_files(data_files)
print 'dumping: "%s"...' % manifest_file
with io.FileIO(manifest_file, "w") as file:
for d in tqdm(data):
label = int(d[0].split('/')[-1].split('-')[1])
line = (d[0] + ',{},{}\n').format(label, d[1])
file.write(line.encode('utf-8'))
if __name__ == '__main__':
from train import parse_arguments
import os, misc
args = parse_arguments()
wav_files = [
os.path.join(dirpath, f)
for dirpath, dirnames, files in os.walk(args.data_dir)
for f in fnmatch.filter(files, '*.wav.pth')
]
print 'Number of files: {}'.format(len(wav_files))
random.shuffle(wav_files)
train_size = len(wav_files) - args.eval_size - args.test_size
train_data = wav_files[0:train_size]
eval_data = wav_files[train_size:train_size + args.eval_size]
test_data = wav_files[train_size + args.eval_size:]
misc.ensure_dir(args.manifest_dir, erase_old=True)
create_manifest(os.path.join(args.manifest_dir, 'train.csv'), train_data)
create_manifest(os.path.join(args.manifest_dir, 'eval.csv'), eval_data)
create_manifest(os.path.join(args.manifest_dir, 'test.csv'), test_data)
'--local_model',
default=
"C:\\Users\\fcalcagno\\Documents\\pytorch-playground_local\\svhn\\log\\best-90.pth",
help='Where the local model is located')
args = parser.parse_args()
args.logdir = os.path.join(os.path.dirname(__file__), args.logdir)
misc.logger.init(args.logdir, 'train_log')
print = misc.logger.info
# select gpu
args.gpu = 1
args.ngpu = 1
# logger
misc.ensure_dir(args.logdir)
print("=================FLAGS==================")
for k, v in args.__dict__.items():
print('{}: {}'.format(k, v))
print("========================================")
# seed
args.cuda = torch.cuda.is_available()
torch.manual_seed(args.seed)
if args.cuda:
torch.cuda.manual_seed(args.seed)
# data loader and model
train_loader, test_loader = dataset_digits.get(batch_size=args.batch_size,
csv_path=args.csv_path,
data_root=args.data_root,
scene_ids = dataset_params.get_present_scene_ids(dp_split)
for scene_id in scene_ids:
# Load scene GT.
scene_gt_path = dp_split['scene_gt_tpath'].format(scene_id=scene_id)
scene_gt = inout.load_scene_gt(scene_gt_path)
# Load scene camera.
scene_camera_path = dp_split['scene_camera_tpath'].format(
scene_id=scene_id)
scene_camera = inout.load_scene_camera(scene_camera_path)
# Create folders for the output masks (if they do not exist yet).
mask_dir_path = os.path.dirname(dp_split['mask_tpath'].format(
scene_id=scene_id, im_id=0, gt_id=0))
misc.ensure_dir(mask_dir_path)
mask_visib_dir_path = os.path.dirname(dp_split['mask_visib_tpath'].format(
scene_id=scene_id, im_id=0, gt_id=0))
misc.ensure_dir(mask_visib_dir_path)
# Initialize a renderer.
misc.log('Initializing renderer...')
width, height = dp_split['im_size']
ren = renderer.create_renderer(width,
height,
renderer_type=p['renderer_type'],
mode='depth')
# Add object models.
for obj_id in dp_model['obj_ids']:
'meshlab_script_path':
os.path.join(os.path.dirname(os.path.realpath(__file__)),
'meshlab_scripts', r'remesh_for_eval_cell=0.25.mlx'),
}
################################################################################
# Load dataset parameters.
dp_model_in = dataset_params.get_model_params(p['datasets_path'], p['dataset'],
p['model_in_type'])
dp_model_out = dataset_params.get_model_params(p['datasets_path'],
p['dataset'],
p['model_out_type'])
# Attributes to save for the output models.
attrs_to_save = []
# Process models of all objects in the selected dataset.
for obj_id in dp_model_in['obj_ids']:
misc.log('\n\n\nProcessing model of object {}...\n'.format(obj_id))
model_in_path = dp_model_in['model_tpath'].format(obj_id=obj_id)
model_out_path = dp_model_out['model_tpath'].format(obj_id=obj_id)
misc.ensure_dir(os.path.dirname(model_out_path))
misc.run_meshlab_script(p['meshlab_server_path'], p['meshlab_script_path'],
model_in_path, model_out_path, attrs_to_save)
misc.log('Done.')
# Output path templates.
out_rgb_tpath =\
os.path.join('{out_path}', '{obj_id:06d}', 'rgb', '{im_id:06d}.png')
out_depth_tpath =\
os.path.join('{out_path}', '{obj_id:06d}', 'depth', '{im_id:06d}.png')
out_scene_camera_tpath =\
os.path.join('{out_path}', '{obj_id:06d}', 'scene_camera.json')
out_scene_gt_tpath =\
os.path.join('{out_path}', '{obj_id:06d}', 'scene_gt.json')
out_views_vis_tpath =\
os.path.join('{out_path}', '{obj_id:06d}', 'views_radius={radius}.ply')
################################################################################
out_path = out_tpath.format(dataset=dataset)
misc.ensure_dir(out_path)
# Load dataset parameters.
dp_split_test = dataset_params.get_split_params(datasets_path, dataset, 'test')
dp_model = dataset_params.get_model_params(datasets_path, dataset, model_type)
dp_camera = dataset_params.get_camera_params(datasets_path, dataset, cam_type)
if not obj_ids:
obj_ids = dp_model['obj_ids']
# Image size and K for the RGB image (potentially with SSAA).
im_size_rgb = [int(round(x * float(ssaa_fact))) for x in dp_camera['im_size']]
K_rgb = dp_camera['K'] * ssaa_fact
# Intrinsic parameters for RGB rendering.
fx_rgb, fy_rgb, cx_rgb, cy_rgb =\
})
# Visualization of the visibility mask.
if p['vis_visibility_masks']:
depth_im_vis = visualization.depth_for_vis(depth, 0.2, 1.0)
depth_im_vis = np.dstack([depth_im_vis] * 3)
visib_gt_vis = visib_gt.astype(np.float)
zero_ch = np.zeros(visib_gt_vis.shape)
visib_gt_vis = np.dstack([zero_ch, visib_gt_vis, zero_ch])
vis = 0.5 * depth_im_vis + 0.5 * visib_gt_vis
vis[vis > 1] = 1
vis_path = p['vis_mask_visib_tpath'].format(
delta=p['delta'],
dataset=p['dataset'],
split=p['dataset_split'],
scene_id=scene_id,
im_id=im_id,
gt_id=gt_id)
misc.ensure_dir(os.path.dirname(vis_path))
inout.save_im(vis_path, vis)
# Save the info for the current scene.
scene_gt_info_path = dp_split['scene_gt_info_tpath'].format(
scene_id=scene_id)
misc.ensure_dir(os.path.dirname(scene_gt_info_path))
inout.save_json(scene_gt_info_path, scene_gt_info)
for obj_id in dp_model['obj_ids']:
# Load object model.
misc.log('Loading 3D model of object {}...'.format(obj_id))
model_path = dp_model['model_tpath'].format(obj_id=obj_id)
ren.add_object(obj_id, model_path)
poses = misc.get_symmetry_transformations(models_info[obj_id],
p['max_sym_disc_step'])
for pose_id, pose in enumerate(poses):
for view_id, view in enumerate(p['views']):
R = view['R'].dot(pose['R'])
t = view['R'].dot(pose['t']) + view['t']
vis_rgb = ren.render_object(obj_id, R, t, fx, fy, cx, cy)['rgb']
# Path to the output RGB visualization.
vis_rgb_path = p['vis_rgb_tpath'].format(vis_path=p['vis_path'],
dataset=p['dataset'],
obj_id=obj_id,
view_id=view_id,
pose_id=pose_id)
misc.ensure_dir(os.path.dirname(vis_rgb_path))
inout.save_im(vis_rgb_path, vis_rgb)
misc.log('Done.')
def vis_object_poses(poses,
K,
renderer,
rgb=None,
depth=None,
vis_rgb_path=None,
vis_depth_diff_path=None,
vis_rgb_resolve_visib=False):
"""Visualizes 3D object models in specified poses in a single image.
Two visualizations are created:
1. An RGB visualization (if vis_rgb_path is not None).
2. A Depth-difference visualization (if vis_depth_diff_path is not None).
:param poses: List of dictionaries, each with info about one pose:
- 'obj_id': Object ID.
- 'R': 3x3 ndarray with a rotation matrix.
- 't': 3x1 ndarray with a translation vector.
- 'text_info': Info to write at the object (see write_text_on_image).
:param K: 3x3 ndarray with an intrinsic camera matrix.
:param renderer: Instance of the Renderer class (see renderer.py).
:param rgb: ndarray with the RGB input image.
:param depth: ndarray with the depth input image.
:param vis_rgb_path: Path to the output RGB visualization.
:param vis_depth_diff_path: Path to the output depth-difference visualization.
:param vis_rgb_resolve_visib: Whether to resolve visibility of the objects
(i.e. only the closest object is visualized at each pixel).
"""
fx, fy, cx, cy = K[0, 0], K[1, 1], K[0, 2], K[1, 2]
# Indicators of visualization types.
vis_rgb = vis_rgb_path is not None
vis_depth_diff = vis_depth_diff_path is not None
if vis_rgb and rgb is None:
raise ValueError(
'RGB visualization triggered but RGB image not provided.')
if (vis_depth_diff or
(vis_rgb and vis_rgb_resolve_visib)) and depth is None:
raise ValueError(
'Depth visualization triggered but D image not provided.')
# Prepare images for rendering.
im_size = None
ren_rgb = None
ren_rgb_info = None
ren_depth = None
if vis_rgb:
im_size = (rgb.shape[1], rgb.shape[0])
ren_rgb = np.zeros(rgb.shape, np.uint8)
ren_rgb_info = np.zeros(rgb.shape, np.uint8)
if vis_depth_diff:
if im_size and im_size != (depth.shape[1], depth.shape[0]):
raise ValueError('The RGB and D images must have the same size.')
else:
im_size = (depth.shape[1], depth.shape[0])
if vis_depth_diff or (vis_rgb and vis_rgb_resolve_visib):
ren_depth = np.zeros((im_size[1], im_size[0]), np.float32)
# Render the pose estimates one by one.
for pose in poses:
# Rendering.
ren_out = renderer.render_object(pose['obj_id'], pose['R'], pose['t'],
fx, fy, cx, cy)
m_rgb = None
if vis_rgb:
m_rgb = ren_out['rgb']
m_mask = None
if vis_depth_diff or (vis_rgb and vis_rgb_resolve_visib):
m_depth = ren_out['depth']
# Get mask of the surface parts that are closer than the
# surfaces rendered before.
visible_mask = np.logical_or(ren_depth == 0, m_depth < ren_depth)
m_mask = np.logical_and(m_depth != 0, visible_mask)
ren_depth[m_mask] = m_depth[m_mask].astype(ren_depth.dtype)
# Combine the RGB renderings.
if vis_rgb:
if vis_rgb_resolve_visib:
ren_rgb[m_mask] = m_rgb[m_mask].astype(ren_rgb.dtype)
else:
ren_rgb_f = ren_rgb.astype(np.float32) + m_rgb.astype(
np.float32)
ren_rgb_f[ren_rgb_f > 255] = 255
ren_rgb = ren_rgb_f.astype(np.uint8)
# Draw 2D bounding box and write text info.
obj_mask = np.sum(m_rgb > 0, axis=2)
ys, xs = obj_mask.nonzero()
if len(ys):
# bbox_color = model_color
# text_color = model_color
bbox_color = (0.3, 0.3, 0.3)
text_color = (1.0, 1.0, 1.0)
text_size = 11
bbox = misc.calc_2d_bbox(xs, ys, im_size)
im_size = (obj_mask.shape[1], obj_mask.shape[0])
ren_rgb_info = draw_rect(ren_rgb_info, bbox, bbox_color)
if 'text_info' in pose:
text_loc = (bbox[0] + 2, bbox[1])
ren_rgb_info = write_text_on_image(ren_rgb_info,
pose['text_info'],
text_loc,
color=text_color,
size=text_size)
# Blend and save the RGB visualization.
if vis_rgb:
misc.ensure_dir(os.path.dirname(vis_rgb_path))
vis_im_rgb = 0.5 * rgb.astype(np.float32) + \
0.5 * ren_rgb.astype(np.float32) + \
1.0 * ren_rgb_info.astype(np.float32)
vis_im_rgb[vis_im_rgb > 255] = 255
inout.save_im(vis_rgb_path,
vis_im_rgb.astype(np.uint8),
jpg_quality=95)
# Save the image of depth differences.
if vis_depth_diff:
misc.ensure_dir(os.path.dirname(vis_depth_diff_path))
# Calculate the depth difference at pixels where both depth maps are valid.
valid_mask = (depth > 0) * (ren_depth > 0)
depth_diff = valid_mask * (ren_depth.astype(np.float32) - depth)
delta = 15
below_delta = valid_mask * (depth_diff < delta)
below_delta_vis = (255 * below_delta).astype(np.uint8)
depth_diff_vis = 255 * depth_for_vis(depth_diff - depth_diff.min())
depth_diff_vis = np.dstack(
[below_delta_vis, depth_diff_vis, depth_diff_vis]).astype(np.uint8)
depth_diff_vis[np.logical_not(valid_mask)] = 0
depth_diff_valid = depth_diff[valid_mask]
depth_info = [
{
'name': 'min diff',
'fmt': ':.3f',
'val': np.min(depth_diff_valid)
},
{
'name': 'max diff',
'fmt': ':.3f',
'val': np.max(depth_diff_valid)
},
{
'name': 'mean diff',
'fmt': ':.3f',
'val': np.mean(depth_diff_valid)
},
]
depth_diff_vis = write_text_on_image(depth_diff_vis, depth_info)
inout.save_im(vis_depth_diff_path, depth_diff_vis)