def test():
# reconstruction documents of the dataset D2
docs = ['datasets/D1/mechanical/D{:03}'.format(i) for i in range(1, 62) if i != 3]
docs += ['datasets/D2/mechanical/D{:03}'.format(i) for i in range(1, 21)]
docs += ['datasets/D3/mechanical/D{:03}'.format(i) for i in range(1, 101)]
# load strips
strips_list = [Strips(path=doc, filter_blanks=True) for doc in docs]
#random.shuffle(docs)
#docs = docs[: MAX_NUM_DOCS]
for doc, strips in zip(docs, strips_list):
map_doc_strips[doc] = strips
# each individual strip of a document is assigned an unique ID.
# map document filename into the ids of such document strips
# map_doc_ids = {}
# cum = 0
# map_doc_ids[doc] = list(range(cum, cum + size))
# cum += size
size = len(strips.strips)
# instances
instance = {
'docs': [doc],
'accuracy': None,
'sizes': [size],
'opt_time': None
}
result = solve(instance)
print('doc={} accuracy={:.2f}% opt_time={:.2f}s'.format(
doc, 100 * result['accuracy'], result['opt_time']
))
def test():
# reconstruction documents of the dataset D2
docs = ['datasets/D2/mechanical/D{:03}'.format(i) for i in range(1, 21)]
# pick MAX_NUM_DOCS documents
random.shuffle(docs)
docs = docs[: MAX_NUM_DOCS]
# load strips
strips_list = [Strips(path=doc, filter_blanks=True) for doc in docs]
# each individual strip of a document is assigned an unique ID.
# map document filename into the ids of such document strips
map_doc_ids = {}
cum = 0
for doc, strips in zip(docs, strips_list):
size = len(strips.strips)
map_doc_strips[doc] = strips # map_doc_strips is a global variable
map_doc_ids[doc] = list(range(cum, cum + size))
cum += size
# main loop
start = time()
for k in range(1, MAX_NUM_DOCS + 1):
# picking k documents
picked_docs = docs[ : k]
# union of ids of the picked docs' strips
sizes_k = []
for doc in picked_docs:
sizes_k.append(len(map_doc_ids[doc]))
# instances
instance = {
'docs': picked_docs,
'accuracy': None,
'sizes': sizes_k,
'opt_time': None
}
result = solve(instance)
print('k={} accuracy={:.2f}% opt_time={:.2f}s'.format(
k, 100 * result['accuracy'], result['opt_time']
))
def test(args, radius=15, hdisp=3, disp_noise=2, pcont=0.3):
docs = glob.glob('{}/**/*.tif'.format(ISRI_DATASET_DIR), recursive=True)
for f, fname in enumerate(docs, 1):
print('Processing document {}/{}'.format(f, len(docs)))
if os.path.basename(fname).replace('.tif', '') in ignore_images:
print('{} is no considered to compose the dataset.'.format(fname))
continue
# generate temporary strips
print(' => Shredding')
if os.path.exists(TEMP_DIR):
shutil.rmtree(TEMP_DIR)
os.makedirs('{}/strips'.format(TEMP_DIR))
image = cv2.imread(fname)
h, w, c = image.shape
acc = 0
for i in range(args.num_strips):
dw = int((w - acc) / (args.num_strips - i))
strip = image[:, acc:acc + dw]
noise_left = np.random.randint(0, 255,
(h, disp_noise)).astype(np.uint8)
noise_right = np.random.randint(0, 255,
(h, disp_noise)).astype(np.uint8)
for j in range(c): # for each channel
strip[:, :disp_noise, j] = cv2.add(strip[:, :disp_noise, j],
noise_left)
strip[:, -disp_noise:, j] = cv2.add(strip[:, -disp_noise:, j],
noise_right)
cv2.imwrite('{}/strips/D001{:02}.jpg'.format(TEMP_DIR, i + 1),
strip)
acc += dw
print(' => Load strips object', end='')
strips = Strips(path=TEMP_DIR, filter_blanks=True)
print('done!')
import sys
import matplotlib.pyplot as plt
import time
sys.path.append('.')
from docrec.strips.strips import Strips
# segmentation
print('=> Segmentation')
t0 = time.time()
strips = Strips(path='datasets/D2/mechanical/D002', filter_blanks=True)
strips.plot()
plt.show()
print('Strips elapsed time={:.2f} seconds'.format(time.time() - t0))
N = len(strips.strips)
fig = plt.figure(figsize=(8, 8), dpi=150)
for i in range(N):
for j in range(N):
if i + 1 == j:
t0 = time.time()
print(i, j, N)
image = strips.pair(i, j, filled=True)
print('Pairing time={:.2f} seconds'.format(time.time() - t0))
plt.clf()
plt.imshow(image)
#plt.savefig('/home/thiagopx/temo/{}-{}.pdf'.format(i, j))
plt.show()
#strips.plot()
parser.add_argument(
'-j', '--j', action='store', dest='j', required=False, type=int,
default=3, help='Sj.'
)
args = parser.parse_args()
# model
images_ph = tf.placeholder(tf.float32, name='images_ph', shape=(None, 3, input_size, input_size)) # channels first
images_adjust_op = tf.image.convert_image_dtype(images_ph, tf.float32)
logits_op = squeezenet(images_ph, 'val', 2, channels_first=True)
probs_op = tf.nn.softmax(logits_op)
predictions_op = tf.argmax(logits_op, 1)
# pair
i, j = args.i, args.j
strips = Strips(path=args.doc, filter_blanks=True)
si, sj = strips.strips[i], strips.strips[j]
hi, wi, _ = si.image.shape
hj, wj, _ = sj.image.shape
min_y = radius_search + radius_feat
max_y = min(hi, hj) - 1 - radius_search - radius_feat
smi = np.correlate(si.offsets_r, [0.05, 0.1, 0.7, 0.1, 0.05], mode='same')
smj = np.correlate(sj.offsets_l, [0.05, 0.1, 0.7, 0.1, 0.05], mode='same')
support = np.hstack([si.filled_image(), sj.filled_image()])
hs, ws, _ = support.shape
blank = np.zeros((hs, 31, 3), dtype=np.uint8)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
params_fname = open('best_model.txt').read()
load(params_fname, sess, model_scope='SqueezeNet')
def test():
global solver
global extra_args
# parameters processing
parser = argparse.ArgumentParser(
description='Testing reconstruction of mixed documents.')
parser.add_argument('-d',
'--dataset',
action='store',
dest='dataset',
required=False,
type=str,
default='cdip',
help='Dataset [D1, D2, or cdip].')
parser.add_argument('-np',
'--nproc',
action='store',
dest='nproc',
required=False,
type=int,
default=10,
help='Number of processes.')
args = parser.parse_args()
assert args.dataset in ['D1', 'D2', 'cdip']
# reconstruction instances
if args.dataset == 'D1':
docs = [
'datasets/D1/mechanical/D{:03}'.format(i) for i in range(1, 62)
if i != 3
]
elif args.dataset == 'D2':
docs = [
'datasets/D2/mechanical/D{:03}'.format(i) for i in range(1, 21)
]
else:
docs = [
'datasets/D3/mechanical/D{:03}'.format(i) for i in range(1, 101)
] # cdip
# shuffle documents
random.shuffle(docs) # this should be considered in other scripts
ndocs = len(docs)
# global compatibility matrix
compatibilities = json.load(
open('results/exp1_proposed/{}_matrix.json'.format(args.dataset), 'r'))
compatibilities = np.array(
result_glob['compatibilities']
) # generated with permutated documents (same seed)
# strips ids for each document
cum = 0
ids_strips = {}
for doc in docs:
size = len(Strips(path=doc, filter_blanks=True).strips)
ids_strips[doc] = list(range(cum, cum + size))
cum += size
# results / initial configuration
ndocs_per_iter = [1, 2, 3, 4, 5] + list(range(10, ndocs + 1, 5))
ndocs_per_iter.reverse() # process in reverve order
results = {
'matrix_id': '{}_matrix'.format(args.dataset),
# 'k_init': 0, # index of ndocs_per_iter where the process should start
'backup_iter': 0, # last iteration
'state': None, # random number generator state
'data': {str(k): []
for k in ndocs_per_iter} # experimental results data
}
results_fname = 'results/exp1_proposed/{}.json'.format(
dir_name, args.dataset)
if os.path.exists(results_fname):
results = json.load(open(results_fname))
it = results['backup_iter']
state = results['state']
if state is not None:
state = (state[0], tuple(state[1]), state[2])
random.setstate(state)
# main loop
with Pool(processes=args.nproc) as pool:
start = time()
total = sum([ndocs - k + 1 for k in ndocs_per_iter])
combs = [(k, offset) for k in ndocs_per_iter
for offset in range(ndocs - k + 1)]
instances = []
for k, offset in combs[it:]:
# picking documents
picked_docs = docs[offset:offset + k]
# union of ids
ids_strips_k = []
sizes_k = []
for doc in sorted(picked_docs):
ids_strips_k += ids_strips[doc]
sizes_k.append(len(ids_strips[doc]))
# crop compatibilities
compatibilities_k = compatibilities[ids_strips_k][:, ids_strips_k]
# shuffle strips
N = compatibilities_k.shape[0]
init_perm_k = list(range(N))
random.shuffle(init_perm_k)
# shuffle compatibilites
compatibilities_k = compatibilities_k[init_perm_k][:, init_perm_k]
# update instances
instance = {
'k': str(k),
'offset': offset,
'docs': picked_docs,
'solution': None,
'accuracy': None,
'init_perm': init_perm_k,
'sizes': sizes_k,
'compatibilities': compatibilities_k,
'opt_time': None
}
instances.append(instance)
# run when the buffer size is greater then args.nproc or when the last iteration is achieved
if (len(instances) == args.nproc) or (it + len(instances)
== total):
print('Iterations {}-{}/{}'.format(it + 1, it + len(instances),
total),
end=' ')
results_buffer = pool.map(solve, instances)
it += len(instances)
instances = [] # reset instances
elapsed = time() - start
predicted = elapsed * (total - it) / it
print(':: elapsed={:.2f}s :: predicted={:.2f}s'.format(
elapsed, predicted))
for result in results_buffer:
results['data'][result['k']].append(result)
print(
' => k={} offset={} accuracy={:.2f}% opt_time={:.2f}s'
.format(result['k'], result['offset'],
100 * result['accuracy'], result['opt_time']))
# dump results and current state
results['backup_iter'] = it
results['state'] = random.getstate()
json.dump(results, open(results_fname, 'w'))
logits_op = model.output
conv10_op = model.view
probs_op = tf.nn.softmax(logits_op)
comp_op = tf.reduce_max(probs_op[:, 1])
disp_op = tf.argmax(probs_op[:, 1]) - args.vshift
sess.run(tf.global_variables_initializer())
best_epoch = json.load(open('traindata/{}/info.json'.format(args.arch), 'r'))['best_epoch']
weights_path = 'traindata/{}/model/{}.{}'.format(args.arch, best_epoch, model_file_ext)
model.load_weights(weights_path)
base_path = 'illustration/heatmap/{}'.format(args.arch)
os.makedirs(base_path, exist_ok=True)
strips_regular = Strips(path=args.doc, filter_blanks=True)
strips_shuffled = strips_regular.copy().shuffle()
N = strips_regular.size()
for strips, strips_type in zip([strips_regular, strips_shuffled], ['regular', 'shuffled']):
# features
features = []
for strip in strips.strips:
left, right = extract_features(strip, (input_size_h, input_size_w))
features.append((left, right))
# converto to BGR
for strip in strips.strips:
strip.image = strip.image[..., :: - 1]
from docrec.strips.strips import Strips
def features(points):
feat = []
for (x3, y3), (x2, y2), (x1, y1) in zip(points[2:], points[1:-1],
points[:-2]):
val = (y3 - y2) * (y2 - y1) + (x3 - x2) * (x2 - x1)
feat.append(val)
return np.array(feat)
doc = 'datasets/D2/mechanical/D002'
i = 21
j = 22
strips = Strips(path=doc, filter_blanks=True)
si, sj = strips.strips[i], strips.strips[j]
hi, wi, _ = si.image.shape
hj, wj, _ = sj.image.shape
offset = wi
stride_feat = 5
radius_feat = 20
radius_search = 5
num_points = 100
#num_feat = 2 * int(radius_feat / stride_feat)
min_y = radius_search + radius_feat
max_y = min(hi, hj) - 1 - radius_search - radius_feat
stride = int((max_y - min_y) / (num_points - 1))
smi = np.correlate(si.offsets_r, [0.05, 0.1, 0.7, 0.1, 0.05], mode='same')
smj = np.correlate(sj.offsets_l, [0.05, 0.1, 0.7, 0.1, 0.05], mode='same')
Pipeline(sleit, solvers_min)
]
# reconstruction instances
docs1 = [
'datasets/D1/mechanical/D{:03}'.format(i) for i in range(1, 62) if i != 3
]
docs2 = ['datasets/D2/mechanical/D{:03}'.format(i) for i in range(1, 21)]
docs = docs1 + docs2
processed = 1
total = len(docs) * len(pipelines)
results = dict()
for doc in docs:
t0 = time.time()
strips = Strips(path=doc, filter_blanks=True)
strips.shuffle()
init_permutation = strips.permutation()
t_load = time.time() - t0
results[doc] = dict(init_permutation=init_permutation,
time=t_load,
algorithms=dict())
for pipeline in pipelines:
print('[{:.2f}%] algorithm={} doc={} ::'.format(
100 * processed / total, pipeline.algorithm.id(), doc),
end='')
processed += 1
d = 2 if pipeline.algorithm.id() == 'marques' else 0
pipeline.run(strips, d)
results[doc]['algorithms'][pipeline.algorithm.id()] = dict()
results[doc]['algorithms'][
weights_path,
10, (3000, width),
num_classes=NUM_CLASSES,
verbose=False,
thresh=args.thresh)
if args.solver == 'concorde':
solver = SolverConcorde(maximize=True, max_precision=2)
elif args.solver == 'kbh':
solver = SolverKBH(maximize=True)
else:
solver = SolverLS(maximize=True)
pipeline = Pipeline(algorithm, [solver])
# load strips and shuffle the strips
print('1) Load strips')
strips = Strips(path=args.doc, filter_blanks=True)
strips.shuffle()
init_permutation = strips.permutation()
print('Shuffled order: ' + str(init_permutation))
print('2) Results')
pipeline.run(strips)
# matrix -> list (displacements according the neighobors strips in solution)
compatibilities = pipeline.algorithm.compatibilities
displacements = pipeline.algorithm.displacements
solution = pipeline.solvers[0].solution
displacements = [
displacements[prev][curr]
for prev, curr in zip(solution[:-1], solution[1:])
]
corrected = [init_permutation[idx] for idx in solution]
print('Processing document {}/{}'.format(f, len(docs)))
if os.path.basename(fname).replace('.tif', '') in ignore_images:
print('{} is no considered to compose the dataset.'.format(fname))
continue
# generate temporary strips
print(' => Shredding')
if os.path.exists(TEMP_DIR):
shutil.rmtree(TEMP_DIR)
os.makedirs('{}/strips'.format(TEMP_DIR))
image = cv2.imread(fname)
h, w, c = image.shape
acc = 0
for i in range(30):
dw = int((w - acc) / (30 - i))
strip = image[:, acc:acc + dw]
noise_left = np.random.randint(0, 255,
(h, disp_noise)).astype(np.uint8)
noise_right = np.random.randint(0, 255,
(h, disp_noise)).astype(np.uint8)
for j in range(c): # for each channel
strip[:, :disp_noise, j] = cv2.add(strip[:, :disp_noise, j],
noise_left)
strip[:, -disp_noise:, j] = cv2.add(strip[:, -disp_noise:, j],
noise_right)
cv2.imwrite('{}/strips/D001{:02}.jpg'.format(TEMP_DIR, i + 1), strip)
acc += dw
print(' => Load strips object')
strips = Strips(path=TEMP_DIR, filter_blanks=True)
# solver = [SolverLS(maximize=True)]
solvers = [
SolverConcorde(maximize=False, max_precision=2),
SolverKBH(maximize=False)
]
# reconstruction instances
docs1 = [
'datasets/D1/mechanical/D{:03}'.format(i) for i in range(1, 62) if i != 3
]
docs2 = ['datasets/D2/mechanical/D{:03}'.format(i) for i in range(1, 21)]
docs = docs1 + docs2
# reconstruction instances
strips_all = {
doc: Strips(path=doc, filter_blanks=True).shuffle()
for doc in docs
}
processed = 1
total = len(docs) * len(algorithms)
records = []
for algorithm in algorithms:
d = 2 if algorithm.id() == 'marques' else 0
for doc, strips in strips_all.items():
print('[{:.2f}%] algorithm={} doc={}'.format(100 * processed / total,
algorithm.id(), doc))
processed += 1
init_permutation = strips.permutation()
compatibilities = algorithm(strips=strips, d=d).compatibilities
qc = Qc(compatibilities,
predictions_op = tf.argmax(logits_op, 1)
# segmentation
print('Processing document {}'.format(args.doc))
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
params_fname = open('best_model.txt').read()
load(params_fname, sess, model_scope='SqueezeNet')
classes = ['negative', 'positive']
t0_global = time.time()
print('=> Segmentation')
t0 = time.time()
#strips = StripsText(path=args.doc, filter_blanks=True)
strips = Strips(path=args.doc, filter_blanks=True)
print('Segmentation elapsed time={:.2f} seconds'.format(time.time() - t0))
N = len(strips.strips)
pcont = 0.2
scores = np.zeros((N, N), dtype=np.int32)
for i in range(N - 1):
border = strips.strips[i].offsets_r
j = i + 1
print('=> Scoring [{}][{}]'.format(i, j))
image = strips.pair(i, j, accurate=True, filled=True)
h, w, _ = image.shape
_, image_bin = cv2.threshold(cv2.cvtColor(image, cv2.COLOR_RGB2GRAY),
0, 1, cv2.THRESH_BINARY +
cv2.THRESH_OTSU) # range [0, 1]
import sys
import os
import cv2
sys.path.append('.')
from docrec.strips.strips import Strips
docs = ['datasets/D2/mechanical/D{:03}'.format(i) for i in range(1, 21)][:1]
for doc in docs:
print('Processing document {}'.format(doc))
strips = Strips(path=doc, filter_blanks=True)
image = strips.reconstruction_image()
cv2.imwrite('test/test_recimage/{}.jpg'.format(os.path.basename(doc)),
image)
def main():
# parameters processing
parser = argparse.ArgumentParser(
description='Testing reconstruction of mixed documents.')
parser.add_argument('-d',
'--dataset',
action='store',
dest='dataset',
required=False,
type=str,
default='cdip',
help='Dataset [D1, D2, or cdip].')
parser.add_argument('-m',
'--model-id',
action='store',
dest='model_id',
required=False,
type=str,
default=None,
help='Model identifier (tag).')
args = parser.parse_args()
assert args.dataset in ['D1', 'D2', 'cdip']
# save directory
save_dir = 'results/exp1_proposed'
os.makedirs(save_dir, exist_ok=True)
results_fname = '{}/{}_matrix.json'.format(save_dir, args.dataset)
# system setup
weights_path = json.load(
open('traindata/{}/info.json'.format(args.model_id),
'r'))['best_model']
algorithm = Proposed('sn',
weights_path,
10, (3000, 32),
num_classes=2,
verbose=True,
thresh_method='sauvola',
seed=SEED)
# reconstruction instances
if args.dataset == 'D1':
docs = [
'datasets/D1/mechanical/D{:03}'.format(i) for i in range(1, 62)
if i != 3
]
elif args.dataset == 'D2':
docs = [
'datasets/D2/mechanical/D{:03}'.format(i) for i in range(1, 21)
]
else:
docs = [
'datasets/D3/mechanical/D{:03}'.format(i) for i in range(1, 101)
] # cdip
# shuffle documents
random.shuffle(docs)
ndocs = len(docs)
# build all
strips_list = [Strips(path=doc, filter_blanks=True) for doc in docs]
strips = MixedStrips(strips_list, shuffle=False)
t0 = time()
algorithm.run(strips, 0)
comp_time = time() - t0
print('ndocs={} inf_time={:.2f}% comp_time={:.2f}s'.format(
ndocs, algorithm.inference_time, comp_time))
results = {
'init_perm': strips.init_perm,
'sizes': strips.sizes,
'compatibilities': algorithm.compatibilities.tolist(),
'displacements': algorithm.displacements.tolist(),
'inf_time': algorithm.inference_time,
'comp_time': comp_time
}
json.dump(results, open(results_fname, 'w'))