-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathdata_io.py
170 lines (157 loc) · 8.27 KB
/
data_io.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
import numpy as np
import pandas as pd
import h5py
import re
class DataIO(object):
def __init__(self):
self.h5f = None
self.dset_loc = None # h5 path of the dataset (for partial loading)
self.local_index = None # index of an element within a h5 dataset
self.flattened_labels = None
def load_all(self, h5fn, dset_name_pattern, camera_pos='*', t_lim=None):
"""
Load all data that match the dataset name pattern.
"""
if self.h5f is None:
self.h5f = h5py.File(h5fn, 'r')
full_dset_list = self.h5f.keys()
r = re.compile(dset_name_pattern)
matched_dset_list = list(filter(r.match, full_dset_list))
print('Selected datasets: %s' % matched_dset_list)
images_set = []
labels_m_set = []
labels_s_set = []
labels_t_set = []
labels_cpos_set = []
if isinstance(camera_pos, int):
camera_pos = [camera_pos] # convert to list
elif isinstance(camera_pos, str):
if camera_pos == '*':
camera_pos = range(0, 14)
print('Selected camera positions: %s' % camera_pos)
for dset_name in matched_dset_list:
for cpos in camera_pos:
h5_path = '/%s/images_camera_%02d' % (dset_name, cpos)
print('Loading dataset %s' % h5_path)
images = self.load_dataset(h5_path)
labels_m, labels_s, labels_t = self.get_labels(dset_name, cpos)
labels_cpos = np.ones(labels_m.shape, dtype=np.int) * cpos
if t_lim is not None:
t_low, t_high = np.min(t_lim), np.max(t_lim)
flags = np.logical_and(labels_t>=t_low, labels_t<=t_high)
images = images[flags]
labels_t = labels_t[flags]
labels_m = labels_m[flags]
labels_s = labels_s[flags]
labels_cpos = labels_cpos[flags]
labels_t_ = (labels_t / 5).astype(np.int)
labels_t_ = labels_t_ - np.min(labels_t_)
images_set.append(images)
labels_m_set.append(labels_m)
labels_s_set.append(labels_s)
labels_t_set.append(labels_t_)
labels_cpos_set.append(labels_cpos)
if len(images_set) > 0:
images_set = np.concatenate(images_set, axis=0)
labels_m_set = np.concatenate(labels_m_set, axis=0)
labels_s_set = np.concatenate(labels_s_set, axis=0)
labels_t_set = np.concatenate(labels_t_set, axis=0)
labels_cpos_set = np.concatenate(labels_cpos_set, axis=0)
return images_set, labels_t_set
def load_partial(self, h5fn, dset_name_pattern, camera_pos=0, t_lim=None, hvd_size=None, hvd_rank=None):
"""
Load the data that match the dataset name pattern partially according to the rank.
"""
if hvd_size is None or hvd_rank is None:
return self.load_all(h5fn, dset_name_pattern, camera_pos, t_lim)
else:
self.dset_loc, self.local_index, self.flattened_labels = self.flatten_index(h5fn, dset_name_pattern, camera_pos, t_lim)
n_images_per_proc = int(len(self.dset_loc) / hvd_size)
print('hvd_size = %d, n_images_per_proc = %d, total_images = %d' % (hvd_size, n_images_per_proc, len(self.dset_loc)))
idx_global_start = n_images_per_proc * hvd_rank
idx_global_end = n_images_per_proc * hvd_rank + n_images_per_proc
print('Rank %d gets %d images, from #%d to #%d' % (hvd_rank, n_images_per_proc, idx_global_start, idx_global_end))
selected_elems = self.dset_loc[idx_global_start:idx_global_end]
selected_elems_idx_local = self.local_index[idx_global_start:idx_global_end]
df = pd.DataFrame(data={'path': selected_elems, 'local_index': selected_elems_idx_local})
data_collective = None
for path_unique, local_id in df.groupby('path'):
local_id = local_id['local_index'].to_numpy()
print('[Rank = %d, N_p = %d] loading dataset: %s' % (hvd_rank, hvd_size, path_unique))
if data_collective is None:
data_collective = self.load_dataset(path_unique)[local_id]
else:
data_collective = np.append(data_collective, self.load_dataset(path_unique)[local_id], axis=0)
# data_collective.append(self.load_dataset(path_unique)[local_id])
# if data_collective is None:
# data_collective = self.load_dataset(path_unique)[local_id]
# else:
# print('shape load', self.load_dataset(path_unique).shape)
# data_collective = np.vstack([data_collective, self.load_dataset(path_unique)[local_id]])
# X = np.array(data_collective)
X = data_collective
print('X shape', X.shape, X[0].shape)
# X = X.reshape(X.shape[0]*X.shape[1], X.shape[2], X.shape[3], X.shape[4])
Y = self.flattened_labels[idx_global_start:idx_global_end]
return X, Y
def load_dataset(self, h5_path):
"""
Load a dataset according to the h5_path.
"""
images = self.h5f[h5_path][()]
if images.dtype == np.uint8:
# if uint8, then the range is [0, 255]. Normalize to [0, 1]
# if float32, don't do anything since the range is already [0, 1]
images = images.astype(np.float) / 255
elif images.shape[-1] == 1:
# just gray scale float images. Repeat along the channel
# old_shape = images.shape
# images = np.repeat(images.astype(np.float32), 3, axis=(len(old_shape)-1))
# print('Repeating...', old_shape, images.shape)
pass
return images
def flatten_index(self, h5fn, dset_name_pattern, camera_pos=0, t_lim=None):
"""
Traverse the HDF5 dataset tree, and flatten the structure (not the data, just the structure)
so that it is more easily diviable accroding to the `hvd_size()`.
"""
dset_loc = []
local_index = []
label_flattened = []
if self.h5f is None:
self.h5f = h5py.File(h5fn, 'r')
full_dset_list = self.h5f.keys()
r = re.compile(dset_name_pattern)
matched_dset_list = list(filter(r.match, full_dset_list))
print('Flattening... Selected datasets: %s' % matched_dset_list)
images_set = []
labels_cpos_set = []
if isinstance(camera_pos, int):
camera_pos = [camera_pos] # convert to list
elif isinstance(camera_pos, str):
if camera_pos == '*':
camera_pos = range(0, 14)
print('Selected camera positions: %s' % camera_pos)
for dset_name in matched_dset_list:
for cpos in camera_pos:
dset_path_full = '/%s/images_camera_%02d' % (dset_name, cpos)
label_dset_path_full = '/%s/t_myr_camera_%02d' % (dset_name, cpos)
print('Obtaining the shape of dataset %s' % dset_path_full)
shape = self.h5f[dset_path_full].shape
labels = self.h5f[label_dset_path_full][()]
labels_t_ = (labels / 5).astype(np.int)
labels_t_ = labels_t_ - np.min(labels_t_)
dset_loc = np.append(dset_loc, [dset_path_full]*shape[0])
local_index = np.append(local_index, np.arange(0, shape[0], dtype=np.int))
label_flattened = np.append(label_flattened, labels_t_)
return dset_loc, local_index.astype(np.int), label_flattened
def get_labels(self, dset_name, camera_pos=0):
print('Getting labels...')
size_ratios = np.array([0.25, 0.5, 0.75, 1, 1.25, 1.5])
mass_ratios = np.array([0.25, 0.5, 0.75, 1, 1.25, 1.5])
s = float(dset_name.split('_')[1])
m = float(dset_name.split('_')[3])
cat_t = self.h5f['%s/t_myr_camera_%02d' % (dset_name, camera_pos)].value
cat_s = np.argwhere(size_ratios==s)[0, 0] * np.ones(cat_t.shape, dtype=np.int)
cat_m = np.argwhere(mass_ratios==m)[0, 0] * np.ones(cat_t.shape, dtype=np.int)
return cat_m, cat_s, cat_t