def compute_average_shape(self, dataset_path):
filelist = fp.dir(dataset_path, '.pts')
im_filelist = fp.dir(dataset_path, '.jpg')
if len(filelist) == 0:
ValueError('dataset_path contains no pts file!')
n = 0
with open(filelist[0]) as f:
lines = f.readlines()
n = int(lines[1].split(':')[1])
P = np.zeros([n, 2], np.float32)
p0 = np.zeros([1, 2], np.float32)
for k, filename in enumerate(filelist):
im_filename = im_filelist[k]
im = cv2.imread(im_filename)
pts = self.read_pts(filename)
row = np.array([im.shape[1], im.shape[0]])
pts = pts / row
p0 += pts[0]
pts -= pts[0]
P += pts
p0 /= len(filelist)
P /= len(filelist)
P += p0
P[:, 1], P[:, 0] = P[:, 0], P[:, 1].copy()
self.save_pts(self.template_filename, P)
print(self.template_filename + ' saved!')
def ComputeEvalForMethod(fd_rst, fd_gt, NumOfThresholds):
fnames_rst = fp.dir(fd_rst, ['.jpg', '.png', '.tiff'])
fnames_gt = fp.dir(fd_gt, ['.jpg', '.png', '.tiff'])
assert len(fnames_rst) == len(fnames_gt), 'len(fnames_rst) == len(fnames_gt)'
NumOfSamples = len(fnames_rst)
EvalMatrix = np.zeros([NumOfSamples, 5, NumOfThresholds], np.float32)
thresholds = np.linspace(0, 1, NumOfThresholds)
for k in tqdm.tqdm(range(NumOfSamples)):
fname_rst = fnames_rst[k]
fname_gt = fnames_gt[k]
rst = cv2.imread(fname_rst, cv2.IMREAD_GRAYSCALE)
gt = cv2.imread(fname_gt, cv2.IMREAD_GRAYSCALE)
rst = rst.astype(np.float32) / 255.0
gt = gt.astype(np.float32) / 255.0
for j, thres in enumerate(thresholds):
Precion, Recall, IoU, FPR, F1 = SegEval(rst >= thres, gt >= 0.01)
EvalMatrix[k,:,j] = Precion, Recall, IoU, FPR, F1
Mu = np.mean(EvalMatrix, axis=0)
Var = np.mean(EvalMatrix, axis=0)
np.save(os.path.join(fd_rst, 'EvalMatrix.npy'), EvalMatrix)
np.save(os.path.join(fd_rst, 'Mu.mu'), Mu)
np.save(os.path.join(fd_rst, 'Var.var'), Var)
def preload(self, fd_npy):
fnames_npy = fp.dir(fd_npy, '.npy')
data_npy = [np.load(fname, 'r') for fname in fnames_npy]
Process = psutil.Process(os.getpid())
#tr = tqdm.trange(len(fnames_npy), desc='Memory Cost', leave=True)
#for d in data_npy:
# logger.info(d.shape)
Ns = [d.shape[0] for d in data_npy]
N = sum(Ns)
return N, data_npy, data_npy[0].shape[1:]
def _get_filenames_helper(fd, title2folder, exts):
filenames = {}
if len(title2folder) == 0:
return filenames
for title, path in title2folder.items():
fullpath = os.path.join(fd, path)
fnames = [
os.path.join('./', path, fp.stem(x))
for x in fp.dir(fullpath, exts, case_sensitive=True)
]
filenames[title] = fnames
flens = [len(fnames) for k, fnames in filenames.items()]
minlen = min(flens)
filenames = {k: fnames[:minlen] for k, fnames in filenames.items()}
return filenames
def copy_specified_files(fd, fd_out, ext, indices, suffix):
fp.mkdir(fd_out)
sub_fds = fp.subdirs(fd)
for sub_fd in sub_fds:
fnames = fp.dir(sub_fd, ext)
fnames = [fnames[idx] for idx in indices]
path, name, ext_ = fp.fileparts(sub_fd)
sub_fd_out = os.path.join(fd_out, name + ext_ + suffix)
fp.mkdir(sub_fd_out)
cmd = 'cp '
for fname in fnames:
cmd += fname + ' '
cmd += sub_fd_out
print(cmd)
os.system(cmd)
def FindBestModel(self):
fname_ckpt_orig = os.path.join(self.FLAGS.log_path, 'checkpoint')
fname_ckpt_backup = os.path.join(self.FLAGS.log_path,
'checkpoint-backup')
try:
cmd = 'cp {} {}'.format(fname_ckpt_orig, fname_ckpt_backup)
os.system(cmd)
except:
logger.warning('The original checkpoint file not backup-ed!')
return
try:
#self.restore_best_model()
fname_ckpt_models = fp.dir(self.FLAGS.log_path,
'.meta',
case_sensitive=True)
fname_ckpt_models = fname_ckpt_models[::-1]
for fname in fname_ckpt_models:
if 'temp_model' in fname:
continue
fname = fname.rstrip('.meta')
fname = fname.replace('\\', '/')
print(fname)
with open(fname_ckpt_orig, 'w') as f:
f.write('model_checkpoint_path: "{}"\n'.format(fname))
f.write('all_model_checkpoint_paths: "{}"\n'.format(fname))
eval_loss_vals = self.Evaluate()
[
logger.info(line)
for line in pyutils.dict_to_string(eval_loss_vals, 3)
]
self.record_best_model(self.best_model_table, eval_loss_vals,
fname)
print(fname + ' END')
except Exception as e:
traceback.print_exc()
cmd = 'cp {} {}'.format(fname_ckpt_backup, fname_ckpt_orig)
os.system(cmd)
logger.error('ERROR')
else:
cmd = 'cp {} {}'.format(fname_ckpt_backup, fname_ckpt_orig)
os.system(cmd)
def preload(self, fd_npy):
fnames_npy = fp.dir(fd_npy, '.npy')
data_npy = [np.load(f, 'r') for f in fnames_npy]
Ns = [d.shape[0] for d in data_npy]
N = sum(Ns)
return N, data_npy, data_npy[0].shape[1:]