def gen(self, ham, itr=1000, nbasis=1000, burn=500, thin=1):
_, state = ground(ham)
n = ham.lattice.numel()
try:
# os.makedirs(os.path.join(self.root, self.raw_folder))
os.makedirs(os.path.join(self.root, self.processed_folder))
except OSError as e:
if e.errno == errno.EEXIST:
pass
else:
raise
data = []
basis = []
for _ in range(nbasis):
ops = MBOp(choice([sigmax, sigmay, sigmaz]) for i in range(n))
_state = ops.invtrans(state)
sampler = STMetropolis(proposal=lambda x: abs(_state[bin(x)])**2,
size=ham.size)
data.extend(sampler.sample(itr=itr, burn=burn, thin=thin))
basis.extend(
torch.FloatTensor(ops.params()).resize_(2 * n)
for i in range(len(sampler.collector)))
return data, basis
def print_instances_class_histogram(dataset_dicts, class_names):
"""
Args:
dataset_dicts (list[dict]): list of dataset dicts.
class_names (list[str]): list of class names (zero-indexed).
"""
num_classes = len(class_names)
hist_bins = np.arange(num_classes + 1)
histogram = np.zeros((num_classes,), dtype=np.int)
for entry in dataset_dicts:
annos = entry["annotations"]
classes = [x["category_id"] for x in annos if not x.get("iscrowd", 0)]
histogram += np.histogram(classes, bins=hist_bins)[0]
N_COLS = min(6, len(class_names) * 2)
def short_name(x):
# make long class names shorter. useful for lvis
if len(x) > 13:
return x[:11] + ".."
return x
data = list(
itertools.chain(*[[short_name(class_names[i]), int(v)] for i, v in enumerate(histogram)])
)
total_num_instances = sum(data[1::2])
data.extend([None] * (N_COLS - (len(data) % N_COLS)))
if num_classes > 1:
data.extend(["total", total_num_instances])
data = itertools.zip_longest(*[data[i::N_COLS] for i in range(N_COLS)])
logging.info("Distribution of instances among all {} categories:\n".format(num_classes))
logging.info(pformat(data))
def parse_annotation(ICDAR_path, id, len_max):
# print(id)
texts = list()
labels = list()
f_task1 = open(
os.path.join(ICDAR_path, 'task3-test(347p)/task3-test(347p)/' + id))
f_img = os.path.join(
ICDAR_path,
'task3-test(347p)/task3-test(347p)/' + id.strip('txt') + 'jpg')
im = Image.open(f_img)
im_width = im.size[0]
im_height = im.size[1]
# In ICDAR case, the first line is our ROI coordinate (xmin, ymin)
line_txt = f_task1.readline()
coor = line_txt.split(',')
ROI_x = int(coor[0].strip('\''))
ROI_y = int(coor[1].strip('\''))
final_total = 0
while line_txt:
line_txt = f_task1.readline()
coor = line_txt.split(',')
# print(coor)
if coor[0] != '"\r\n' and coor[0] != '"\n' and coor[0] != '' and coor[
0] != '\r\n':
xmin = float((int(coor[0].strip('\'')) - ROI_x)) / float(im_width)
ymin = float((int(coor[1].strip('\'')) - ROI_y)) / float(im_height)
xmax = float((int(coor[4].strip('\'')) - ROI_x)) / float(im_width)
ymax = float((int(coor[5].strip('\'')) - ROI_y)) / float(im_height)
text = coor[8:]
# 'ori_text' pretains special signs which block the following comparison but are useful in encoding
ori_text = copy.deepcopy(text)
ori_text = ','.join(ori_text)
ori_text = list(ori_text)
text_ascii = [ord(c) for c in ori_text]
data = [xmin, ymin, xmax, ymax]
data.extend(text_ascii)
data_pad = [0] * len_max
if len(data) <= len_max:
data_pad[:len(data)] = data
else:
data_pad = data[:len_max]
# print(label)
texts.append(data_pad)
return {'texts': texts}, [ROI_x, ROI_y, im_width, im_height]
def print_instances_class_histogram(dataset_dicts, class_names):
"""
Args:
dataset_dicts (list[dict]): list of dataset dicts.
class_names (list[str]): list of class names (zero-indexed).
"""
logger = logging.getLogger(__name__)
logger.info("Build instances class histogram")
num_classes = len(class_names)
hist_bins = np.arange(num_classes + 1)
histogram = np.zeros((num_classes, ), dtype=np.int)
for entry in tqdm(dataset_dicts):
annos = entry["annotations"]
classes = np.asarray([x["category_id"] for x in annos], dtype=np.int)
if len(classes):
assert classes.min(
) >= 0, f"Got an invalid category_id={classes.min()}"
assert (
classes.max() < num_classes
), f"Got an invalid category_id={classes.max()} for a dataset of {num_classes} classes"
histogram += np.histogram(classes, bins=hist_bins)[0]
N_COLS = min(6, len(class_names) * 2)
def short_name(x):
# make long class names shorter. useful for lvis
if len(x) > 13:
return x[:11] + ".."
return x
data = list(
itertools.chain(
*[[short_name(class_names[i]), int(v)]
for i, v in enumerate(histogram)]))
total_num_instances = sum(data[1::2])
data.extend([None] * (N_COLS - (len(data) % N_COLS)))
if num_classes > 1:
data.extend(["total", total_num_instances])
data = itertools.zip_longest(*[data[i::N_COLS] for i in range(N_COLS)])
# TODO: Too many classes expected. Need to summarize histogram
table = tabulate(
data,
headers=["category", "#instances"] * (N_COLS // 2),
tablefmt="pipe",
numalign="left",
stralign="center",
)
log_first_n(
logging.INFO,
"Distribution of instances among all {} categories:\n".format(
num_classes) + colored(table, "cyan"),
key="message",
)
def build_data(self, textdata: TextData):
data = []
for n in range(len(textdata.sentences)):
indexes = textdata.get_sentence_indexes(n)
ngram_position_range = range(-self.context_size, len(indexes))
ngrams = [
self.get_ngram_tensors_at_position(position, indexes)
for position in ngram_position_range
]
data.extend(ngrams)
return data
def print_instances_class_histogram(dataset_dicts, class_names):
"""
Args:
dataset_dicts (list[dict]): list of dataset dicts.
class_names (list[str]): list of class names (zero-indexed).
"""
num_classes = len(class_names)
hist_bins = np.arange(num_classes + 1) # [0,1]
histogram = np.zeros((num_classes, ), dtype=np.int)
for entry in dataset_dicts:
annos = entry["annotations"] # iscrowd 为0 的才有效
classes = [x["category_id"] for x in annos if not x.get("iscrowd", 0)]
# np.histogram(a, bin:int or list, range:tuple..)
# a： 待统计数组，bin可以是int,那么要指定range,或者list表示bin的范围，
# return hist:统计好的每个区间的数，bin:区间的划分
histogram += np.histogram(classes, bins=hist_bins)[0]
N_COLS = min(6, len(class_names) * 2) # 一个图cols?
def short_name(x):
# make long class names shorter. useful for lvis
if len(x) > 13:
return x[:11] + ".."
return x
# 改短名字
data = list(
itertools.chain(
*[[short_name(class_names[i]), int(v)]
for i, v in enumerate(histogram)]))
# data = [name1, num1, name2, v2, ...]
total_num_instances = sum(
data[1::2]) # data[1::2] a[i:j:step]表示从index = 1 开始切片，直到末尾，step = 2
data.extend([None] * (N_COLS - (len(data) % N_COLS))) # 补齐为N_COLS 的倍数
if num_classes > 1:
data.extend(["total", total_num_instances])
#data[i::N_COLS],为什么要搞出这么多个呢(6个)
# zip_longest, 跟zip差不多，给定多个列表，a,b,.. ,按照最长的那个枚举，不足的填None
data = itertools.zip_longest(*[data[i::N_COLS] for i in range(N_COLS)])
# 不知道干啥。。。
table = tabulate(
data,
headers=["category", "#instances"] * (N_COLS // 2),
tablefmt="pipe",
numalign="left",
stralign="center",
)
log_first_n(
logging.INFO,
"Distribution of instances among all {} categories:\n".format(
num_classes) + colored(table, "cyan"),
key="message",
)
def build_vocab(self, min_freq=0, max_freq=sys.maxsize):
"""
build vocab + add eos
encode sentence
"""
with open(os.path.join(self.data_dir, 'train.txt'), 'r') as fn:
data = fn.readlines()
if 'lambada' in self.data_dir:
with open(os.path.join(self.data_dir, 'test.txt'), 'r') as fn:
data.extend(fn.readlines())
with open(os.path.join(self.data_dir, 'valid.txt'), 'r') as fn:
data.extend(fn.readlines())
print('building vocab ...')
self.vocab = defaultdict(int)
self.tok2id = {}
self.id2tok = []
for line in tqdm(data):
line = line.strip().split()
for tok in line:
self.vocab[tok] += 1
self.vocab = {
a: self.vocab[a]
for a in self.vocab
if self.vocab[a] >= min_freq and self.vocab[a] <= max_freq
}
# sort vocab in case of using adaptive softmax
self.vocab = list(
sorted(self.vocab.items(), key=lambda a: a[1], reverse=True))
print(self.vocab[:10])
if 'lambada' in self.data_dir:
self.vocab = self.vocab[:60000]
self.vocab.append(('<unk>', 0))
self.id2tok = ['<pad>'] + ['<eos>'] + [a[0] for a in self.vocab]
self.tok2id = {a: i for i, a in enumerate(self.id2tok)}
self.vocab_size = len(self.id2tok)
print('end building vocab ...')
print('vocab size', len(self.tok2id))
with open(os.path.join(self.data_dir, 'vocab.pkl'), 'wb') as fn:
pickle.dump(
{
'id2tok': self.id2tok,
'tok2id': self.tok2id,
'vocab_size': self.vocab_size
}, fn)
def __getitem__(self, index):
assert (index < self._dataset_size)
#if self._is_for_train: # random
# index = random.randint(0, self._dataset_size-1)
# get sample data
data = self._data[index] / 1000
data = list(data)
data_reverse = copy.deepcopy(data)
data_reverse.reverse()
filt_rri1 = list(moving_average(RRi(data), order=1))
filt_rri2 = list(moving_average(RRi(data), order=2))
filt_rri3 = list(moving_average(RRi(data), order=3))
filt_rri1_reverse = copy.deepcopy(filt_rri1)
filt_rri2_reverse = copy.deepcopy(filt_rri2)
filt_rri3_reverse = copy.deepcopy(filt_rri3)
filt_rri1_reverse.reverse()
filt_rri2_reverse.reverse()
filt_rri3_reverse.reverse()
order_data = [filt_rri1, filt_rri2, filt_rri3]
order_data_reverse = [
filt_rri1_reverse, filt_rri2_reverse, filt_rri3_reverse
]
label = int(self._label[index])
subject = self._subject[index]
mean = self._mean_train[index]
sdnn = self._sdnn_train[index]
pnn50 = self._pnn50_train[index]
rmssd = self._rmssd_train[index]
lnrmssd = self._lnrmssd_train[index]
vlf = self._vlf_train[index]
lf = self._lf_train[index]
hf = self._hf_train[index]
rlh = self._rlh_train[index]
features = list(np.stack((mean, sdnn, pnn50, rmssd, lnrmssd, \
vlf, lf, hf, rlh )))
makeup_length = 512 - len(data)
if len(data) > 512:
data = data[:512]
else:
data.extend(0 for _ in range(makeup_length))
return data, data_reverse, order_data, order_data_reverse, label, subject, features
def print_instances_class_histogram(dataset_dicts,
class_names,
attribute='annotations'):
"""
Args:
dataset_dicts (list[dict]): list of data dicts.
class_names (list[str]): list of class names (zero-indexed).
"""
num_classes = len(class_names)
hist_bins = np.arange(num_classes + 1)
histogram = np.zeros((num_classes, ), dtype=np.int)
for entry in dataset_dicts:
classes = []
annos = entry[attribute]
if attribute == 'annotations':
classes = [
x["category_id"] for x in annos if not x.get("iscrowd", 0)
]
elif attribute == 'annotations2':
classes = [x["category2_id"] for x in annos]
histogram += np.histogram(classes, bins=hist_bins)[0]
N_COLS = min(6, len(class_names) * 2)
def short_name(x):
# make long class names shorter. useful for lvis
if len(x) > 13:
return x[:11] + ".."
return x
data = list(
itertools.chain(
*[[short_name(class_names[i]), int(v)]
for i, v in enumerate(histogram)]))
total_num_instances = sum(data[1::2])
data.extend([None] * (N_COLS - (len(data) % N_COLS)))
if num_classes > 1:
data.extend(["total", total_num_instances])
data = itertools.zip_longest(*[data[i::N_COLS] for i in range(N_COLS)])
table = tabulate(
data,
headers=["category", "#instances"] * (N_COLS // 2),
tablefmt="pipe",
numalign="left",
stralign="center",
)
log_first_n(
logging.INFO,
"Distribution of instances among all {} categories:\n".format(
num_classes) + colored(table, "cyan"),
key="message",
)
def generate(self):
data = []
basis = []
for _ in range(self.nbasis):
ops = MBOp(choice([sigmax, sigmay, sigmaz]) for i in range(self.n))
_state = ops.invtrans(self.state)
sampler = STMetropolis(proposal=lambda x: abs(_state[bin(x)])**2,
size=self.size)
data.extend(
sampler.sample(itr=self.itr, burn=self.burn, thin=self.thin))
basis.extend(
torch.FloatTensor(ops.params()).resize_(2 * self.n)
for i in range(len(sampler.collector)))
return data, basis
def load_data(self):
data = []
for idx, model_name in enumerate(os.listdir(self.root)):
model_path = os.path.join(self.root, model_name)
if 'task_' + str(self.task) in model_path:
log('\nExtracting weights from %s' % model_path)
model_weights = torch.load(model_path)['model_state_dict']
vec = self.kernels_to_vector(model_weights)
data.extend([vec])
path = os.path.join(self.root, '../pickles/dataset.pkl')
torch.save(data, path)
return data
def __init__(self, configs):
split = configs['split']
root = configs['root']
self.transform = configs['transform']
self.split = split
data_dict = torch.load(
os.path.join(root, 'data', 'cifar10', split + '.pth'))
labels = []
data = []
for label, data_list in data_dict.items():
n_samples = len(data_list)
labels.extend([label] * n_samples)
data.extend(data_list)
print('Loaded %d data, %d labels' % (len(labels), len(data)))
self.data = np.concatenate([x.reshape(1, -1) for x in data])
print('Concatenated shape:', self.data.shape)
self.data = self.data.reshape((-1, 3, 32, 32))
self.data = self.data.transpose((0, 2, 3, 1)) # convert to HWC
self.labels = labels
def get_cafe_data():
'''
Returns (data, labels) where data is an n x d tensor and labels is an
n x 1 tensor.
'''
images = load_cafe()
# Build data and label tensors
data = []
labels = []
for i, exprList in enumerate(images):
data.extend([transform(image).reshape(-1) for image in exprList])
labels.extend([i] * len(exprList))
labels = torch.tensor(labels)
data = torch.stack(data)
return data, labels
def receive(client_socket, size):
"""
Recevie data to specific socket client.
Args:
client_socket: socket id
size: data size to be received
Returns:
received data.
"""
n = 8 * size
data = bytearray()
while len(data) < n:
packet = client_socket.recv(n - len(data))
if not packet:
continue
data.extend(packet)
assert (n == len(data))
res = decode_float_array(data, size)
return res
def my_collate(batch, frame_size, fixed_frame_num):
'''
Break an utterance to multiple frames.
The last frame is dropped if len(frame) < frame_size.
'''
data, target = [], []
for (mat, label) in batch:
data_shape = np.shape(mat)
# print("mfcc mat shape:", data_shape, "label:", label)
if mat.shape[0] >= frame_size:
mul_mat, mul_label = breakIntoFrames(mat, label, frame_size,
fixed_frame_num)
data.extend(mul_mat)
target.extend(mul_label)
# print(" mul_mat:", len(mul_mat), " mul_label:", len(mul_label))
# print("total data len:", len(data), "total target len:", len(target))
data = torch.LongTensor(data)
target = torch.LongTensor(target)
return [data, target]
def load_json(json_path,
x='feat_length',
y='token_length',
x_range=(1, 9999),
y_range=(1, 999),
rate=(1, 99)):
# json
try:
# json_path is a single file
with open(json_path) as f:
data = json.load(f)
except:
# json_path is a dir where *.json in
data = []
for dir, _, fs in os.walk(json_path): # os.walk获取所有的目录
for f in fs:
if f.endswith('.json'): # 判断是否是".json"结尾
filename = os.path.join(dir, f)
print('loading json file :', filename)
with open(filename) as f:
add = json.load(f)
data.extend(add)
print('loaded {} samples'.format(len(add)))
list_to_pop = []
for i, sample in enumerate(data):
len_x = sample[x]
len_y = sample[y]
if not (x_range[0] <= len_x <= x_range[1]) or \
not (y_range[0] <= len_y <= y_range[1]) or \
not (rate[0] <= (len_x / len_y) <= rate[1]):
list_to_pop.append(i)
# filter
print('filtered {}/{} samples\n'.format(len(list_to_pop), len(data)))
list_to_pop.reverse()
[data.pop(i) for i in list_to_pop]
return data
def __init__(self,
json_path,
reverse=False,
feat_range=(1, 99999),
label_range=(1, 100),
rate_in_out=(4, 999)):
try:
# json_path is a single file
with open(json_path) as f:
data = json.load(f)
except:
# json_path is a dir where *.json in
data = []
for dir, _, fs in os.walk(json_path): # os.walk获取所有的目录
for f in fs:
if f.endswith('.json'): # 判断是否是".json"结尾
filename = os.path.join(dir, f)
print('loading json file :', filename)
with open(filename) as f:
add = json.load(f)
data.extend(add)
print('loaded {} samples'.format(len(add)))
# filter
list_to_pop = []
for i, sample in enumerate(data):
len_x = sample['feat_length']
len_y = sample['token_length']
if not (feat_range[0] <= len_x <= feat_range[1]) or \
not (label_range[0] <= len_y <= label_range[1]) or \
not (rate_in_out[0] <= (len_x / len_y) <= rate_in_out[1]):
list_to_pop.append(i)
print('filtered {}/{} samples\n'.format(len(list_to_pop), len(data)))
list_to_pop.reverse()
[data.pop(i) for i in list_to_pop]
self.data = sorted(data, key=lambda x: float(x["feat_length"]))
if reverse:
self.data.reverse()
def __getitem__(self, index):
data = []
data0, label0 = self.get_item_once(index, self.video_list0)
data1, label0 = self.get_item_once(index, self.video_list1)
data2, label0 = self.get_item_once(index, self.video_list2)
"""
data3, label0 = self.get_item_once(index,self.video_list3)
data4, label0 = self.get_item_once(index,self.video_list4)
data5, label0 = self.get_item_once(index,self.video_list5)
data6, label0 = self.get_item_once(index,self.video_list6)
data7, label0 = self.get_item_once(index,self.video_list7)
"""
data.extend(data0)
data.extend(data1)
data.extend(data2)
"""
data.extend(data3)
data.extend(data4)
data.extend(data5)
data.extend(data6)
data.extend(data7)
"""
process_data = self.transform(data)
return process_data, label0
def __init__(self, scene, data_path, train, transform=None,
target_transform=None, real=False, skip_images=False, seed=7,
undistort=False, vo_lib='stereo', data_dir=None, unsupervise=False,
config=None):
"""
:param scene: e.g. 'full' or 'loop'. collection of sequences.
:param data_path: Root RobotCar data directory.
Usually '../data/deepslam_data/RobotCar'
:param train: flag for training / validation
:param transform: Transform to be applied to images
:param target_transform: Transform to be applied to poses
:param real: it determines load ground truth pose or vo pose
:param skip_images: return None images, only poses
:param seed: random seed
:param undistort: whether to undistort images (slow)
:param vo_lib: Library to use for VO ('stereo' or 'gps')
(`gps` is a misnomer in this code - it just loads the position information
from GPS)
:param data_dir: indicating where to load stats.txt file(to normalize image&pose)
:param unsupervise: load training set as supervise or unsupervise
"""
np.random.seed(seed)
self.train = train
self.transform = transform
self.target_transform = target_transform
self.skip_images = skip_images
self.undistort = undistort
base_dir = osp.expanduser(osp.join(data_path, scene))
self.config = config
# data_dir = osp.join('..', 'data', 'RobotCar', scene)
if self.config.has_key('new_split') and self.config.new_split:
print("use new split dataset")
if train:
split_filename = osp.join(base_dir, 'train_split.txt')
else:
split_filename = osp.join(base_dir, 'test_split.txt')
with open(split_filename, 'r') as f:
seqs = [l.rstrip() for l in f if not l.startswith('#')]
pose_filename = osp.join(base_dir, "dataset_train.txt")
pose_dict = {}
with open(pose_filename, 'r') as f:
data = f.readlines()[3:]
pose_filename = osp.join(base_dir, "dataset_test.txt")
with open(pose_filename, 'r') as f:
data.extend(f.readlines()[3:])
imgs = [
l.split(' ')[0] for l in data if not l.startswith('#')
]
ps = np.asarray([
[float(num) for num in l.split(' ')[1:]]
for l in data if not l.startswith('#')
], dtype=np.float32)
poses = np.zeros((ps.shape[0], 6))
poses[:, :3] = ps[:, :3]
poses[:, 3:] = np_qlog_t(ps[:, 3:])
for idx, img_name in enumerate(imgs):
pose_dict[img_name] = poses[idx, :]
self.poses = np.empty((0, 6))
self.imgs = []
for seq in seqs:
# seq_dir = osp.join(base_dir, seq)
img_names = [img for img in imgs if img.startswith(seq)]
# print(img_names)
poses = np.asarray([
pose_dict[img_name] for img_name in img_names if pose_dict.has_key(img_name)
])
self.imgs.extend([osp.join(base_dir, img_name) for img_name in img_names])
self.poses = np.vstack((self.poses, poses))
else:
if train:
if unsupervise:
split_filename = osp.join(base_dir, 'unsupervised_train_split.txt')
else:
split_filename = osp.join(base_dir, 'dataset_train.txt')
else:
split_filename = osp.join(base_dir, 'dataset_test.txt')
with open(split_filename, 'r') as f:
data = f.readlines()
self.imgs = [
osp.join(
base_dir,
l.split(' ')[0]
) for l in data[3:] if not l.startswith('#')
]
ps = np.asarray([
[float(num) for num in l.split(' ')[1:]]
for l in data[3:] if not l.startswith('#')
], dtype=np.float32)
self.poses = np.zeros((ps.shape[0], 6))
self.poses[:, :3] = ps[:, :3]
self.poses[:, 3:] = np_qlog_t(ps[:, 3:])
self.mask_sampling = self.config.mask_sampling
if self.mask_sampling:
muimg = read_grayscale_image(osp.join(base_dir, self.config.mu_mask_name))
self.muimg = torch.tensor(muimg.transpose(2, 0, 1)).type(torch.FloatTensor)
self.sigmaimg = self.muimg * (1 - self.muimg)
pose_stats_filename = osp.join(data_dir, 'pose_stats.txt')
if train and not real:
mean_t = np.mean(self.poses[:, :3], axis=0)
std_t = np.std(self.poses[:, :3], axis=0)
np.savetxt(pose_stats_filename, np.vstack((mean_t, std_t)), fmt='%8.7f')
print("Saved")
else:
mean_t, std_t = np.loadtxt(pose_stats_filename)
self.poses[:, :3] -= mean_t
self.poses[:, :3] /= std_t
# convert the pose to translation + log quaternion, align, normalize
self.gt_idx = np.asarray(range(len(self.poses)))
# camera model and image loader
self.im_loader = partial(load_image)
def __getitem__(self, index):
count, id_idx, ii, dset, protein_id, seq_length = self.protein_list[
index]
window_size = self.window_size
id_idx = int(id_idx)
win_start = ii - window_size
win_end = ii + window_size
seq_length = int(seq_length)
label_idx = (win_start + win_end) // 2
all_seq_features = []
seq_len = 0
for idx in self.all_sequences[id_idx][:self.max_seq_len]:
acid_one_hot = [0 for i in range(20)]
acid_one_hot[idx] = 1
all_seq_features.append(acid_one_hot)
seq_len += 1
while seq_len < self.max_seq_len:
acid_one_hot = [0 for i in range(20)]
all_seq_features.append(acid_one_hot)
seq_len += 1
all_pssm_features = self.all_pssm[id_idx][:self.max_seq_len]
seq_len = len(all_pssm_features)
while seq_len < self.max_seq_len:
zero_vector = [0 for i in range(20)]
all_pssm_features.append(zero_vector)
seq_len += 1
all_dssp_features = self.all_dssp[id_idx][:self.max_seq_len]
seq_len = len(all_dssp_features)
while seq_len < self.max_seq_len:
zero_vector = [0 for i in range(9)]
all_dssp_features.append(zero_vector)
seq_len += 1
local_features = []
labels = []
while win_start < 0:
data = []
acid_one_hot = [0 for i in range(20)]
data.extend(acid_one_hot)
pssm_zero_vector = [0 for i in range(20)]
data.extend(pssm_zero_vector)
dssp_zero_vector = [0 for i in range(9)]
data.extend(dssp_zero_vector)
local_features.extend(data)
win_start += 1
valid_end = min(win_end, seq_length - 1)
while win_start <= valid_end:
data = []
idx = self.all_sequences[id_idx][win_start]
acid_one_hot = [0 for i in range(20)]
acid_one_hot[idx] = 1
data.extend(acid_one_hot)
pssm_val = self.all_pssm[id_idx][win_start]
data.extend(pssm_val)
try:
dssp_val = self.all_dssp[id_idx][win_start]
except:
dssp_val = [0 for i in range(9)]
data.extend(dssp_val)
local_features.extend(data)
win_start += 1
while win_start <= win_end:
data = []
acid_one_hot = [0 for i in range(20)]
data.extend(acid_one_hot)
pssm_zero_vector = [0 for i in range(20)]
data.extend(pssm_zero_vector)
dssp_zero_vector = [0 for i in range(9)]
data.extend(dssp_zero_vector)
local_features.extend(data)
win_start += 1
label = self.all_label[id_idx][label_idx]
label = np.array(label, dtype=np.float32)
all_seq_features = np.stack(all_seq_features)
all_seq_features = all_seq_features[np.newaxis, :, :]
all_pssm_features = np.stack(all_pssm_features)
all_pssm_features = all_pssm_features[np.newaxis, :, :]
all_dssp_features = np.stack(all_dssp_features)
all_dssp_features = all_dssp_features[np.newaxis, :, :]
local_features = np.stack(local_features)
return all_seq_features, all_pssm_features, all_dssp_features, local_features, label
def __init__(self, hp, mode):
super(Photo2Sketch_Dataset, self).__init__()
self.hp = hp
self.mode = mode
hp.root_dir = '/home/media/On_the_Fly/Code_ALL/Final_Dataset'
hp.dataset_name = 'ShoeV2'
hp.seq_len_threshold = 251
# coordinate_path = os.path.join(hp.root_dir, hp.dataset_name , hp.dataset_name + '_Coordinate')
self.root_dir = os.path.join(hp.root_dir, hp.dataset_name)
with open('./preprocess/ShoeV2_RDP_3', 'rb') as fp:
self.Coordinate = pickle.load(fp)
coordinate_refine = {}
seq_len = []
for key in self.Coordinate.keys():
if len(self.Coordinate[key]) < 81:
coordinate_refine[key] = self.Coordinate[key]
seq_len.append(len(self.Coordinate[key]))
self.Coordinate = coordinate_refine
hp.max_seq_len = max(seq_len)
hp.average_seq_len = int(np.round(np.mean(seq_len) + np.std(seq_len)))
# greater_than_average = 0
# for seq in seq_len:
# if seq > self.hp.average_len:
# greater_than_average +=1
self.Train_Sketch = [
x for x in self.Coordinate
if ('train' in x) and (len(self.Coordinate[x]) < 130)
] # separating trains
self.Test_Sketch = [
x for x in self.Coordinate
if ('test' in x) and (len(self.Coordinate[x]) < 130)
] # separating tests
self.train_transform = get_transform('Train')
self.test_transform = get_transform('Test')
# # seq_len = []
# # for key in self.Coordinate.keys():
# # seq_len += [len(self.Coordinate[key])]
# # plt.hist(seq_len)
# # plt.savefig('histogram of number of Coordinate Points.png')
# # plt.close()
# # hp.max_seq_len = max(seq_len)
# hp.max_seq_len = 130
"""" Preprocess offset coordinates """
self.Offset_Coordinate = {}
for key in self.Coordinate.keys():
self.Offset_Coordinate[key] = to_delXY(self.Coordinate[key])
data = []
for sample in self.Offset_Coordinate.values():
data.extend(sample[:, 0])
data.extend(sample[:, 1])
data = np.array(data)
scale_factor = np.std(data)
for key in self.Coordinate.keys():
self.Offset_Coordinate[key][:, :2] /= scale_factor
"""" <<< Preprocess offset coordinates >>> """
"""" <<< Done >>> """
def make_data(root_path, annotation_path, allow_reverse, mirror, balance,
balance_proportions, subset, sample_duration, required_overlap,
shuffle):
assert (subset in ['train', 'val', 'test'],
'subset "{}" is not "train", "val" or "test"'.format(subset))
changevalue = {'validation': 'val', 'train': 'train'}
subset = changevalue[subset]
frame_format = lambda x: str(x).zfill(4) + '.png'
data = []
action_list = _load_list(annotation_path)
local_track_action_dict = defaultdict(list)
for action in action_list:
local_track_action_dict[action[0]].append(action[1:])
for local_track_path, actions in tqdm(local_track_action_dict.items(),
desc='Creating a dataset'):
# sort the list of action by the starting frame
actions.sort(key=lambda a: a[1])
if subset != actions[0][-2]: continue
frame_dir = os.path.join(root_path, local_track_path)
files = os.listdir(frame_dir)
# split the frames into samples
for i in range(len(files) // sample_duration):
use_sample = True
sample_start = i * sample_duration + 1
sample_end = min((i + 1) * sample_duration, len(files))
# make sure all the frames in the sample are saved properly, if they're not ignore the sample
frame_indices = list(range(sample_start, sample_end + 1))
for index in frame_indices:
frame_path = os.path.join(root_path, local_track_path,
frame_format(index))
if not os.path.exists(frame_path):
use_sample = False
break
if not use_sample: continue
# count the number of frames in sample that are part of the action
action_label = 0
for (label, start, end, frame, subset, hard_data) in actions:
action_overlap = max(
0, end - sample_start +
1 if sample_start > start else sample_end - start + 1)
if action_overlap > sample_duration * required_overlap:
action_label = label
break
sample = {
'path': frame_dir,
'label': action_label,
'frame_indices': frame_indices,
'mirror': False
}
data.append(sample)
reverse_data = []
if allow_reverse:
for sample in tqdm(data, desc='Reversing the data'):
reverse = True if sample['label'] != 0 else random.random() > 0.5
if reverse:
reverse_data.append(_reverse_sample(sample))
data.extend(reverse_data)
mirror_data = []
if mirror:
for sample in tqdm(data, desc='Adding mirror flip'):
mirror_data.append(_mirror_sample(sample))
data.extend(mirror_data)
labels = set([i['label'] for i in data])
count_labels = {
label: sum([1 for i in data if i['label'] == label])
for label in labels
}
print('Dataset is generated. Label counts are:', count_labels)
if balance:
required = min(count_labels.values())
if balance_proportions is not None:
sampling_probabilities = {
label: (required / value) * balance_proportions[label]
for label, value in count_labels.items()
}
else:
sampling_probabilities = {
label: required / value
for label, value in count_labels.items()
}
balanced_data = []
for sample in tqdm(data, desc='Balancing dataset'):
if random.random() < sampling_probabilities[sample['label']]:
balanced_data.append(sample)
data = balanced_data
count_labels = {
label: sum([1 for i in data if i['label'] == label])
for label in labels
}
print('Balanced dataset is generated. Label counts are:', count_labels)
if shuffle:
random.shuffle(data)
return data
def search(dirname, data):
filenames = os.listdir(dirname)
for filename in filenames:
full_filename = os.path.join(dirname, filename, '*.png')
data.extend(Get_Dataset(full_filename))
print(full_filename + ' Load!')
for string in args.model:
if '=' in string:
update = parse_dotted(string)
else:
with open(string, 'r') as f:
update = yaml.safe_load(f)
# If the yaml file contains an entry with key `model` use that one instead
if 'model' in update.keys():
update = update['model']
update_rec(model, update)
# Data from --data args
for path in args.data:
path = Path(path).expanduser().resolve()
if path.is_dir():
data.extend(path.glob('*.pt'))
elif path.is_file() and path.suffix == '.pt':
data.append(path)
else:
raise ValueError(f'Invalid data: {path}')
# Options from --options args
for string in args.options:
if '=' in string:
update = parse_dotted(string)
else:
with open(string, 'r') as f:
update = yaml.safe_load(f)
update_rec(options, update)
# Resolving paths