def write_corpus(fname, corpus, progress_cnt=1000, index=False, num_terms=None, metadata=False):
"""
Save the vector space representation of an entire corpus to disk.
Note that the documents are processed one at a time, so the whole corpus
is allowed to be larger than the available RAM.
"""
mw = MmWriter(fname)
# write empty headers to the file (with enough space to be overwritten later)
mw.write_headers(-1, -1, -1) # will print 50 spaces followed by newline on the stats line
# calculate necessary header info (nnz elements, num terms, num docs) while writing out vectors
_num_terms, num_nnz = 0, 0
docno, poslast = -1, -1
offsets = []
if hasattr(corpus, 'metadata'):
orig_metadata = corpus.metadata
corpus.metadata = metadata
if metadata:
docno2metadata = {}
else:
metadata = False
for docno, doc in enumerate(corpus):
if metadata:
bow, data = doc
docno2metadata[docno] = data
else:
bow = doc
if docno % progress_cnt == 0:
logger.info("PROGRESS: saving document #%i" % docno)
if index:
posnow = mw.fout.tell()
if posnow == poslast:
offsets[-1] = -1
offsets.append(posnow)
poslast = posnow
max_id, veclen = mw.write_vector(docno, bow)
_num_terms = max(_num_terms, 1 + max_id)
num_nnz += veclen
if metadata:
utils.pickle(docno2metadata, fname + '.metadata.cpickle')
corpus.metadata = orig_metadata
num_docs = docno + 1
num_terms = num_terms or _num_terms
if num_docs * num_terms != 0:
logger.info("saved %ix%i matrix, density=%.3f%% (%i/%i)" % (
num_docs, num_terms,
100.0 * num_nnz / (num_docs * num_terms),
num_nnz,
num_docs * num_terms))
# now write proper headers, by seeking and overwriting the spaces written earlier
mw.fake_headers(num_docs, num_terms, num_nnz)
mw.close()
if index:
return offsets
def _calibrate(images_path=IMAGES_PATH,
chessboard_rows=CHESSBOARD_ROWS,
chessboard_cols=CHESSBOARD_COLS,
image_size=CALIBRATION_IMAGE_SIZE,
calibration_pickle_file=CALIBRATION_PICKLE_FILE):
obj = np.zeros((chessboard_rows * chessboard_cols, 3), np.float32)
obj[:, :2] = np.mgrid[:chessboard_cols, :chessboard_rows].T.reshape(
-1, 2)
points_object = []
points_image = []
images = glob.glob(images_path)
for image in images:
image_array = imread(image)
if image_array.shape != image_size:
image_array = imresize(image_array, image_size)
gray = cv2.cvtColor(image_array, cv2.COLOR_RGB2GRAY)
ret, corners = cv2.findChessboardCorners(
gray, (chessboard_cols, chessboard_rows), None)
if ret:
points_object.append(obj)
points_image.append(corners)
calibration = (points_object, points_image)
utils.pickle(calibration, calibration_pickle_file)
return calibration
def gaussMixture(testX, goodSample,
data=None, train=False, plot=False):
if train==True:
n_classes = 3
covar_type = 'full'
est = GMM(n_components=n_classes, covariance_type=covar_type)
est.fit(data)
utils.pickle(est, 'SrcTeam/capsuleData/capsule_gauss')
else:
est = utils.unpickle('SrcTeam/capsuleData/capsule_gauss')
numMatch = 0.0
numGood = goodSample.shape[0]
testData = np.reshape(testX,(1,testX.size))
predLabel = est.predict(testData)
for i in range(numGood):
if est.predict(goodSample[i:i+1,:]) == predLabel:
numMatch += 1
if plot==True:
fig = pl.figure()
pl.clf()
ax = Axes3D(fig)
labels = est.predict(data)
ax.scatter(data[:,0],data[:,1],data[:,2],c=labels.astype(np.float))
pl.show()
return float(numMatch) / numGood
def k_means(testX, goodSample,
data=None, train=False, plot=False):
if train==True:
n_clusters = 3
est = KMeans(n_clusters)
est.fit(data)
centers = est.cluster_centers_
utils.pickle(est, 'SrcTeam/capsuleData/capsule_k_means')
else:
est = utils.unpickle('SrcTeam/capsuleData/capsule_k_means')
numMatch = 0.0
numGood = goodSample.shape[0]
#sampleLabel = clusterLabel(centers, sample)
testLabel = est.predict(testX)
for i in range(numGood):
if est.predict(goodSample[i,:]) == testLabel:
numMatch += 1
if plot==True:
fig = pl.figure()
pl.clf()
ax = Axes3D(fig)
labels = est.labels_
ax.scatter(data[:,0],data[:,1],data[:,2],c=labels.astype(np.float))
pl.show()
return float(numMatch) / numGood
def gaussMixture(sample, goodSample,
data=None, train=True, plot=False):
if train==True:
n_classes = 3
covar_type = 'full'
classifier = GMM(n_components=n_classes, covariance_type=covar_type)
classifier.fit(data)
utils.pickle(classifier, 'data/capsule_gauss')
else:
utils.classifier = unpickle('data/capsule_gauss')
numMatch = 0.0
numGood = goodSample.shape[0]
sampleLabel = classifier.predict(sample)
for i in range(numGood):
if classifier.predict(goodSample[i,:]) == sampleLabel:
numMatch += 1
if plot==True:
fig = pl.figure()
pl.clf()
ax = Axes3D(fig)
labels = est.predict(data)
ax.scatter(data[:,0],data[:,1],data[:,2],c=labels.astype(np.float))
pl.show()
return float(numMatch) / numGood
def main(args):
files, clean_labels = parse(osp.join(args.data_root, 'clean_train.txt'))
files, noisy_labels = parse(osp.join(args.data_root, 'noisy_train.txt'))
matrix_c = compute_matrix_c(clean_labels, noisy_labels)
write_matrix(matrix_c, osp.join(args.data_root, 'matrix_c.txt'))
pickle(matrix_c, osp.join(args.data_root, 'matrix_c.pkl'))
noise_types = get_noise_types(clean_labels, noisy_labels, matrix_c)
make_data(files, noise_types, args.data_root)
def main(args):
files, clean_labels = parse(osp.join(args.data_root, 'clean_train.txt'))
files, noisy_labels = parse(osp.join(args.data_root, 'noisy_train.txt'))
matrix_c = compute_matrix_c(clean_labels, noisy_labels)
write_matrix(matrix_c, osp.join(args.data_root, 'matrix_c.txt'))
pickle(matrix_c, osp.join(args.data_root, 'matrix_c.pkl'))
noise_types = get_noise_types(clean_labels, noisy_labels, matrix_c)
make_data(files, noise_types, args.data_root)
def main(args):
size = args.size
q = generate_matrix_q(args.level)
write_matrix(
q, osp.join(args.data_root, 'matrix_q' + repr(args.level) + '.txt'))
pickle(q, osp.join(args.data_root, 'matrix_q' + repr(args.level) + '.pkl'))
files, labels = parse(osp.join(args.data_root, 'train.txt'))
noisy_labels = corrupt(labels, q)
write_file_label_list(files[:size], labels[:size],
osp.join(args.data_root, 'clean.txt'))
write_file_label_list(
files[:size], noisy_labels[:size],
osp.join(args.data_root, 'noisy_' + repr(args.level) + '.txt'))
write_list(
noisy_labels[:size],
osp.join(args.data_root, 'labels_noisy_' + repr(args.level) + '.txt'))
write_list([f + ' -1' for f in files[:size]],
osp.join(args.data_root, 'images.txt'))
def main(args):
mkdir_if_missing(args.output_dir)
# training data
data = []
labels = []
for i in xrange(1, 6):
dic = unpickle(osp.join(args.data_root, 'data_batch_{}'.format(i)))
data.append(dic['data'])
labels = np.r_[labels, dic['labels']]
data = np.vstack(data)
make_data(data, labels, args.output_dir, 'train')
# test data
dic = unpickle(osp.join(args.data_root, 'test_batch'))
make_data(dic['data'], dic['labels'], args.output_dir, 'test')
# Identity for confusion initialization
matrix_I = np.identity(10)
write_matrix(matrix_I, osp.join(args.output_dir, 'identity.txt'))
pickle(matrix_I, osp.join(args.output_dir, 'identity.pkl'))
def main(args):
q = generate_matrix_q(args.level)
write_matrix(q, osp.join(args.data_root, 'matrix_q.txt'))
pickle(q, osp.join(args.data_root, 'matrix_q.pkl'))
files, labels = parse(osp.join(args.data_root, 'train.txt'))
noisy_labels = corrupt(labels, q)
write_file_label_list(files[:10000], labels[:10000],
osp.join(args.data_root, 'clean_train.txt'))
write_file_label_list(files[:10000], noisy_labels[:10000],
osp.join(args.data_root, 'noisy_train.txt'))
noisy_as_clean_labels = labels[:10000] + noisy_labels[10000:]
noisy_as_none_labels = labels[:10000] + [-1] * 40000
clean_as_none_labels = [-1] * 10000 + noisy_labels[10000:]
merged = zip(files, noisy_as_clean_labels, noisy_as_none_labels, clean_as_none_labels)
np.random.shuffle(merged)
files, nacl, nanl, canl = zip(*merged)
write_file_label_list(files, nacl,
osp.join(args.data_root, 'mixed_train.txt'))
write_list([f + ' -1' for f in files],
osp.join(args.data_root, 'mixed_train_images.txt'))
write_list(nanl, osp.join(args.data_root, 'mixed_train_label_clean.txt'))
write_list(canl, osp.join(args.data_root, 'mixed_train_label_noisy.txt'))
def chooseAction(self, observedState):
"""
Here, choose pacman's next action based on the current state of the game.
This is where all the action happens.
"""
# calculate reward (score delta) for last action
current_score = observedState.score
last_score = self.score
reward = current_score - last_score
if self.chatter: print reward
# pass reward to learner
self.learner.reward_callback(reward)
# apply basis function to calculate new state
state = self.basis(observedState)
# ask learner to plan new state
allowed_action_codes = [self.actionCodes[a] for a in self.actionBasis.allowedActions(self, observedState)]
action_code = self.learner.action_callback(state,allowed_action_codes)
# update score
self.score = current_score
# update number of actions taken
self.action_count += 1
# save results
if((self.save_every > 0) and (self.action_count % self.save_every == 0)):
if self.chatter: print "Saving..."
utils.pickle(self.learner, self.learn_file)
# take action
if self.chatter: print state, self.actions[action_code],
action = self.actionBasis(observedState, self.actions[action_code])
return action
def main(args):
if args.cfg_file is not None:
cfg_from_file(args.cfg_file)
if args.set_cfgs is not None:
cfg_from_list(args.set_cfgs)
# parse gpus
gpus = map(int, args.gpus.split(','))
assert len(gpus) >= mpi_size, "Number of GPUs must be >= MPI size"
cfg.GPU_ID = gpus[mpi_rank]
# parse feature blob names
blob_names = args.blob_names.split(',')
print('Using config:')
pprint.pprint(cfg)
while not osp.exists(args.caffemodel) and args.wait:
print('Waiting for {} to exist...'.format(args.caffemodel))
time.sleep(10)
# load imdb
imdb = get_imdb(args.imdb_name)
root_dir = imdb._root_dir
images_dir = imdb._data_path
output_dir = get_output_dir(imdb.name,
osp.splitext(osp.basename(args.caffemodel))[0])
if args.eval_only:
def _load(fname):
fpath = osp.join(output_dir, fname)
assert osp.isfile(fpath), "Must have extracted detections and " \
"features first before evaluation"
return unpickle(fpath)
if mpi_rank == 0:
gboxes = _load('gallery_detections.pkl')
gfeatures = _load('gallery_features.pkl')
pfeatures = _load('probe_features.pkl')
else:
# setup caffe
caffe.mpi_init()
caffe.set_mode_gpu()
caffe.set_device(cfg.GPU_ID)
# 1. Detect and extract features from all the gallery images in the imdb
start, end = mpi_dispatch(len(imdb.image_index), mpi_size, mpi_rank)
if args.use_gt:
net = caffe.Net(args.probe_def, args.caffemodel, caffe.TEST)
gboxes, gfeatures = usegt_and_exfeat(net, imdb,
start=start, end=end, blob_names=blob_names)
else:
net = caffe.Net(args.gallery_def, args.caffemodel, caffe.TEST)
gboxes, gfeatures = detect_and_exfeat(net, imdb,
start=start, end=end, blob_names=blob_names)
gboxes = mpi_collect(mpi_comm, mpi_rank, gboxes)
gfeatures = mpi_collect(mpi_comm, mpi_rank, gfeatures)
del net # to release the cudnn conv static workspace
# 2. Only extract features from given probe rois
start, end = mpi_dispatch(len(imdb.probes), mpi_size, mpi_rank)
net = caffe.Net(args.probe_def, args.caffemodel, caffe.TEST)
pfeatures = exfeat(net, imdb.probes,
start=start, end=end, blob_names=blob_names)
pfeatures = mpi_collect(mpi_comm, mpi_rank, pfeatures)
del net
# Save
if mpi_rank == 0:
pickle(gboxes, osp.join(output_dir, 'gallery_detections.pkl'))
pickle(gfeatures, osp.join(output_dir, 'gallery_features.pkl'))
pickle(pfeatures, osp.join(output_dir, 'probe_features.pkl'))
# Evaluate
if mpi_rank == 0:
imdb.evaluate_detections(gboxes, det_thresh=args.det_thresh)
imdb.evaluate_detections(gboxes, det_thresh=args.det_thresh,
labeled_only=True)
imdb.evaluate_search(gboxes, gfeatures['feat'], pfeatures['feat'],
det_thresh=args.det_thresh,
gallery_size=args.gallery_size,
dump_json=osp.join(output_dir, 'results.json'))
def write_corpus(fname,
corpus,
progress_cnt=1000,
index=False,
num_terms=None,
metadata=False):
"""
Save the vector space representation of an entire corpus to disk.
Note that the documents are processed one at a time, so the whole corpus
is allowed to be larger than the available RAM.
"""
mw = MmWriter(fname)
# write empty headers to the file (with enough space to be overwritten later)
mw.write_headers(
-1, -1,
-1) # will print 50 spaces followed by newline on the stats line
# calculate necessary header info (nnz elements, num terms, num docs) while writing out vectors
_num_terms, num_nnz = 0, 0
docno, poslast = -1, -1
offsets = []
if hasattr(corpus, 'metadata'):
orig_metadata = corpus.metadata
corpus.metadata = metadata
if metadata:
docno2metadata = {}
else:
metadata = False
for docno, doc in enumerate(corpus):
if metadata:
bow, data = doc
docno2metadata[docno] = data
else:
bow = doc
if docno % progress_cnt == 0:
logger.info("PROGRESS: saving document #%i" % docno)
if index:
posnow = mw.fout.tell()
if posnow == poslast:
offsets[-1] = -1
offsets.append(posnow)
poslast = posnow
max_id, veclen = mw.write_vector(docno, bow)
_num_terms = max(_num_terms, 1 + max_id)
num_nnz += veclen
if metadata:
utils.pickle(docno2metadata, fname + '.metadata.cpickle')
corpus.metadata = orig_metadata
num_docs = docno + 1
num_terms = num_terms or _num_terms
if num_docs * num_terms != 0:
logger.info(
"saved %ix%i matrix, density=%.3f%% (%i/%i)" %
(num_docs, num_terms, 100.0 * num_nnz /
(num_docs * num_terms), num_nnz, num_docs * num_terms))
# now write proper headers, by seeking and overwriting the spaces written earlier
mw.fake_headers(num_docs, num_terms, num_nnz)
mw.close()
if index:
return offsets
def serialize(self):
'''Pickle the trained recognizer to the models/ directory of the dataset.'''
if not os.path.exists(self._datadir + '/models/'):
os.mkdir(self._datadir + '/models/')
pickle(self, self._datadir + '/models/' + self._modelname)
def main(args):
mkdir_if_missing(args.output_dir)
matrix_I = np.identity(2)
write_matrix(matrix_I, osp.join(args.output_dir, 'identity.txt'))
pickle(matrix_I, osp.join(args.output_dir, 'identity.pkl'))
def main(args):
q = generate_matrix_q(args.level)
q = np.transpose(q)
write_matrix(
q, osp.join(args.data_root, 'matrix_q' + repr(args.level) + '.txt'))
pickle(q, osp.join(args.data_root, 'matrix_q' + repr(args.level) + '.pkl'))
def gt_roidb(self):
cache_file = osp.join(self.cache_path, self.name + '_gt_roidb.pkl')
if osp.isfile(cache_file):
roidb = unpickle(cache_file)
return roidb
# Load all images and build a dict from image to boxes
all_imgs = loadmat(osp.join(self._root_dir, 'annotation', 'Images.mat'))
all_imgs = all_imgs['Img'].squeeze()
name_to_boxes = {}
name_to_pids = {}
for im_name, __, boxes in all_imgs:
im_name = str(im_name[0])
boxes = np.asarray([b[0] for b in boxes[0]])
boxes = boxes.reshape(boxes.shape[0], 4)
valid_index = np.where((boxes[:, 2] > 0) & (boxes[:, 3] > 0))[0]
assert valid_index.size > 0, \
'Warning: {} has no valid boxes.'.format(im_name)
boxes = boxes[valid_index]
name_to_boxes[im_name] = boxes.astype(np.int32)
name_to_pids[im_name] = -1 * np.ones(boxes.shape[0], dtype=np.int32)
def _set_box_pid(boxes, box, pids, pid):
for i in xrange(boxes.shape[0]):
if np.all(boxes[i] == box):
pids[i] = pid
return
print 'Warning: person {} box {} cannot find in Images'.format(pid, box)
# Load all the train / test persons and label their pids from 0 to N-1
# Assign pid = -1 for unlabeled background people
if self._image_set == 'train':
train = loadmat(osp.join(self._root_dir,
'annotation/test/train_test/Train.mat'))
train = train['Train'].squeeze()
for index, item in enumerate(train):
scenes = item[0, 0][2].squeeze()
for im_name, box, __ in scenes:
im_name = str(im_name[0])
box = box.squeeze().astype(np.int32)
_set_box_pid(name_to_boxes[im_name], box,
name_to_pids[im_name], index)
else:
test = loadmat(osp.join(self._root_dir,
'annotation/test/train_test/TestG50.mat'))
test = test['TestG50'].squeeze()
for index, item in enumerate(test):
# query
im_name = str(item['Query'][0,0][0][0])
box = item['Query'][0,0][1].squeeze().astype(np.int32)
_set_box_pid(name_to_boxes[im_name], box,
name_to_pids[im_name], index)
# gallery
gallery = item['Gallery'].squeeze()
for im_name, box, __ in gallery:
im_name = str(im_name[0])
if box.size == 0: break
box = box.squeeze().astype(np.int32)
_set_box_pid(name_to_boxes[im_name], box,
name_to_pids[im_name], index)
# Construct the gt_roidb
gt_roidb = []
for im_name in self.image_index:
boxes = name_to_boxes[im_name]
boxes[:, 2] += boxes[:, 0]
boxes[:, 3] += boxes[:, 1]
pids = name_to_pids[im_name]
num_objs = len(boxes)
gt_classes = np.ones((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
overlaps[:, 1] = 1.0
overlaps = csr_matrix(overlaps)
gt_roidb.append({
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'gt_pids': pids,
'flipped': False})
pickle(gt_roidb, cache_file)
print "wrote gt roidb to {}".format(cache_file)
return gt_roidb
import os.path as osp
import numpy as np
from argparse import ArgumentParser
from utils import pickle
import caffe
def write_matrix(mat, file_path):
content = [' '.join(map(str, r)) for r in mat]
with open(file_path, 'w') as f:
f.write('\n'.join(content))
fp = 0.95
fn = 0.95
tp = 0.05
tn = 0.05
data = np.array([[fn, tn], [tp, fp]], dtype=np.float)
write_matrix(data, '../data/infogain/Q17.txt')
pickle(data, '../data/infogain/Q17.pkl')
shape = data.shape
shape = (1, ) * (4 - len(shape)) + shape
data = data.reshape(shape)
blob = caffe.proto.caffe_pb2.BlobProto()
blob.num, blob.channels, blob.height, blob.width = data.shape
blob.data.extend(list(data.ravel().astype(float)))
with open('../data/infogain/Q17.binaryproto', 'wb') as f:
f.write(blob.SerializeToString())
def gt_roidb(self):
cache_file = osp.join(self.cache_path, self.name + '_gt_roidb.pkl')
if osp.isfile(cache_file):
roidb = unpickle(cache_file)
return roidb
# Load all images and build a dict from image to boxes
all_imgs = loadmat(osp.join(self._root_dir, 'annotation',
'Images.mat'))
all_imgs = all_imgs['Img'].squeeze()
name_to_boxes = {}
name_to_pids = {}
for im_name, __, boxes in all_imgs:
im_name = str(im_name[0])
boxes = np.asarray([b[0] for b in boxes[0]])
boxes = boxes.reshape(boxes.shape[0], 4)
valid_index = np.where((boxes[:, 2] > 0) & (boxes[:, 3] > 0))[0]
assert valid_index.size > 0, \
'Warning: {} has no valid boxes.'.format(im_name)
boxes = boxes[valid_index]
name_to_boxes[im_name] = boxes.astype(np.int32)
name_to_pids[im_name] = -1 * np.ones(boxes.shape[0],
dtype=np.int32)
def _set_box_pid(boxes, box, pids, pid):
for i in xrange(boxes.shape[0]):
if np.all(boxes[i] == box):
pids[i] = pid
return
print 'Warning: person {} box {} cannot find in Images'.format(
pid, box)
# Load all the train / test persons and label their pids from 0 to N-1
# Assign pid = -1 for unlabeled background people
if self._image_set == 'train':
train = loadmat(
osp.join(self._root_dir,
'annotation/test/train_test/Train.mat'))
train = train['Train'].squeeze()
for index, item in enumerate(train):
scenes = item[0, 0][2].squeeze()
for im_name, box, __ in scenes:
im_name = str(im_name[0])
box = box.squeeze().astype(np.int32)
_set_box_pid(name_to_boxes[im_name], box,
name_to_pids[im_name], index)
else:
test = loadmat(
osp.join(self._root_dir,
'annotation/test/train_test/TestG50.mat'))
test = test['TestG50'].squeeze()
for index, item in enumerate(test):
# query
im_name = str(item['Query'][0, 0][0][0])
box = item['Query'][0, 0][1].squeeze().astype(np.int32)
_set_box_pid(name_to_boxes[im_name], box,
name_to_pids[im_name], index)
# gallery
gallery = item['Gallery'].squeeze()
for im_name, box, __ in gallery:
im_name = str(im_name[0])
if box.size == 0: break
box = box.squeeze().astype(np.int32)
_set_box_pid(name_to_boxes[im_name], box,
name_to_pids[im_name], index)
# Construct the gt_roidb
gt_roidb = []
for im_name in self.image_index:
boxes = name_to_boxes[im_name]
boxes[:, 2] += boxes[:, 0]
boxes[:, 3] += boxes[:, 1]
pids = name_to_pids[im_name]
num_objs = len(boxes)
gt_classes = np.ones((num_objs), dtype=np.int32)
overlaps = np.zeros((num_objs, self.num_classes), dtype=np.float32)
overlaps[:, 1] = 1.0
overlaps = csr_matrix(overlaps)
gt_roidb.append({
'boxes': boxes,
'gt_classes': gt_classes,
'gt_overlaps': overlaps,
'gt_pids': pids,
'flipped': False
})
pickle(gt_roidb, cache_file)
print "wrote gt roidb to {}".format(cache_file)
return gt_roidb
def main(args):
if args.cfg_file is not None:
cfg_from_file(args.cfg_file)
if args.set_cfgs is not None:
cfg_from_list(args.set_cfgs)
# parse gpus
gpus = map(int, args.gpus.split(','))
assert len(gpus) >= mpi_size, "Number of GPUs must be >= MPI size"
cfg.GPU_ID = gpus[mpi_rank]
# parse feature blob names
blob_names = args.blob_names.split(',')
print('Using config:')
pprint.pprint(cfg)
while not osp.exists(args.caffemodel) and args.wait:
print('Waiting for {} to exist...'.format(args.caffemodel))
time.sleep(10)
# load imdb
imdb = get_imdb(args.imdb_name)
root_dir = imdb._root_dir
images_dir = imdb._data_path
output_dir = get_output_dir(imdb.name,
osp.splitext(osp.basename(args.caffemodel))[0])
if args.eval_only:
def _load(fname):
fpath = osp.join(output_dir, fname)
assert osp.isfile(fpath), "Must have extracted detections and " \
"features first before evaluation"
return unpickle(fpath)
if mpi_rank == 0:
gboxes = _load('gallery_detections.pkl')
gfeatures = _load('gallery_features.pkl')
pfeatures = _load('probe_features.pkl')
else:
# setup caffe
caffe.mpi_init()
caffe.set_mode_gpu()
caffe.set_device(cfg.GPU_ID)
# 1. Detect and extract features from all the gallery images in the imdb
start, end = mpi_dispatch(len(imdb.image_index), mpi_size, mpi_rank)
if args.use_gt:
net = caffe.Net(args.probe_def, args.caffemodel, caffe.TEST)
gboxes, gfeatures = usegt_and_exfeat(net,
imdb,
start=start,
end=end,
blob_names=blob_names)
else:
net = caffe.Net(args.gallery_def, args.caffemodel, caffe.TEST)
gboxes, gfeatures = detect_and_exfeat(net,
imdb,
start=start,
end=end,
blob_names=blob_names)
gboxes = mpi_collect(mpi_comm, mpi_rank, gboxes)
gfeatures = mpi_collect(mpi_comm, mpi_rank, gfeatures)
del net # to release the cudnn conv static workspace
# 2. Only extract features from given probe rois
start, end = mpi_dispatch(len(imdb.probes), mpi_size, mpi_rank)
net = caffe.Net(args.probe_def, args.caffemodel, caffe.TEST)
pfeatures = exfeat(net,
imdb.probes,
start=start,
end=end,
blob_names=blob_names)
pfeatures = mpi_collect(mpi_comm, mpi_rank, pfeatures)
del net
# Save
if mpi_rank == 0:
pickle(gboxes, osp.join(output_dir, 'gallery_detections.pkl'))
pickle(gfeatures, osp.join(output_dir, 'gallery_features.pkl'))
pickle(pfeatures, osp.join(output_dir, 'probe_features.pkl'))
# Evaluate
if mpi_rank == 0:
imdb.evaluate_detections(gboxes, det_thresh=args.det_thresh)
imdb.evaluate_detections(gboxes,
det_thresh=args.det_thresh,
labeled_only=True)
imdb.evaluate_search(gboxes,
gfeatures['feat'],
pfeatures['feat'],
det_thresh=args.det_thresh,
gallery_size=args.gallery_size,
dump_json=osp.join(output_dir, 'results.json'))
def _add_to_tfrecord(filename, tfrecord_writer, offset=0):
"""Loads data from the cifar10 pickle files and writes files to a TFRecord.
Args:
filename: The filename of the cifar10 pickle file.
tfrecord_writer: The TFRecord writer to use for writing.
offset: An offset into the absolute number of images previously written.
Returns:
The new offset.
"""
with tf.gfile.Open(filename, 'rb') as f:
if sys.version_info < (3,):
data = cPickle.load(f)
else:
data = cPickle.load(f, encoding='bytes')
images = data[b'data']
num_images = images.shape[0]
images = images.reshape((num_images, 3, 32, 32))
labels = data[b'fine_labels']
coarse_labels = data[b'coarse_labels']
c2f_map = {}
for lb, cl in zip(labels, coarse_labels):
if cl not in c2f_map:
c2f_map[cl] = {lb}
else:
c2f_map[cl].add(lb)
utils.pickle(c2f_map, utils.root_path + '/data/cifar100/c2f_map.pkl')
b2a_map = {}
ind = 0
for c, fs in c2f_map.items():
for f in fs:
b2a_map[f] = ind
ind += 1
utils.pickle(b2a_map, utils.root_path + '/data/cifar100/b2a_map.pkl')
a2b_map = {a: b for b, a in b2a_map.items()}
# labels = [b2a_map[lb] for lb in labels]
with tf.Graph().as_default():
image_placeholder = tf.placeholder(dtype=tf.uint8)
encoded_image = tf.image.encode_png(image_placeholder)
with tf.Session() as sess:
for j in range(num_images):
sys.stdout.write('\r>> Reading file [%s] image %d/%d' % (
filename, offset + j + 1, offset + num_images))
sys.stdout.flush()
image = np.squeeze(images[j]).transpose((1, 2, 0))
lb = labels[j]
png_string = sess.run(encoded_image,
feed_dict={image_placeholder: image})
example = dataset_utils.image_to_tfexample(
png_string, b'png', _IMAGE_SIZE, _IMAGE_SIZE, lb)
tfrecord_writer.write(example.SerializeToString())
return offset + num_images