def data_loader(cfg, fname):
dims = cfg['dims']
chunk_size = cfg['batch_size'] * cfg['batches_per_chunk']
xc = np.zeros((
chunk_size,
cfg['n_channels'],
) + dims, dtype=np.float32)
reader = npytar.NpyTarReader(fname)
yc = []
for ix, (x, name) in enumerate(reader):
cix = ix % chunk_size
xc[cix] = x.astype(np.float32)
yc.append(int(name.split('.')[0]) - 1)
if len(yc) == chunk_size:
xc = jitter_chunk(xc, cfg)
yield (2.0 * xc - 1.0, np.asarray(yc, dtype=np.float32))
yc = []
xc.fill(0)
if len(yc) > 0:
# pad to nearest multiple of batch_size
if len(yc) % cfg['batch_size'] != 0:
new_size = int(np.ceil(
len(yc) / float(cfg['batch_size']))) * cfg['batch_size']
xc = xc[:new_size]
xc[len(yc):] = xc[:(new_size - len(yc))]
yc = yc + yc[:(new_size - len(yc))]
xc = jitter_chunk(xc, cfg)
yield (2.0 * xc - 1.0, np.asarray(yc, dtype=np.float32))
def data_loader(cfg, fname):
dims = cfg['dims']
chunk_size = cfg['n_rotations']
xc = np.zeros((chunk_size, cfg['n_channels'],)+dims, dtype=np.float32)
reader = npytar.NpyTarReader(fname)
yc = []
for ix, (x, name) in enumerate(reader):
cix = ix % chunk_size
xc[cix] = x.astype(np.float32)
yc.append(int(name.split('.')[0])-1)
if len(yc) == chunk_size:
yield (2.0*xc - 1.0, np.asarray(yc, dtype=np.float32))
yc = []
xc.fill(0)
assert(len(yc)==0)
def data_loader(fname):
# the number for reading each time
chunk_size = cfg['batch_size']*cfg['batches_per_chunk']
xc = np.zeros((chunk_size, cfg['n_channels'],cfg['n_levels'],cfg['n_rings'],dims), dtype=np.float32)
reader = npytar.NpyTarReader(fname)
yc = []
for ix, (x, name) in enumerate(reader):
cix = ix % chunk_size
xc[cix] = x.astype(np.float32)
yc.append(int(name.split('.')[0])-1)
if len(yc) == chunk_size:
yield (xc, np.asarray(yc, dtype=np.float32))
yc = []
xc.fill(0)
if len(yc) > 0:
# pad to nearest multiple of batch_size
new_size = int(np.ceil(len(yc)/float(cfg['batch_size'])))*cfg['batch_size']
xc = xc[:new_size]
xc[len(yc):] = xc[:(new_size-len(yc))]
yc = yc + yc[:(new_size - len(yc))]
yield (xc, np.asarray(yc, dtype=np.float32))
def main(args):
reader = npytar.NpyTarReader(args.viz_data_fname)
out = np.load(args.viz_out_fname)
yhat = out['yhat']
ygnd = out['ygnd']
css = """
html { margin: 0 }
body {
background:#fff;
color:#000;
font:75%/1.5em Helvetica, "DejaVu Sans", "Liberation sans", "Bitstream Vera Sans", sans-serif;
position:relative;
}
/*dt { font-weight: bold; float: left; clear; left }*/
div { padding: 10px; width: 80%; margin: auto }
img { border: 1px solid #eee }
dl { margin:0 0 1.5em; }
dt { font-weight:700; }
dd { margin-left:1.5em; }
table {
border-collapse:collapse;
border-spacing:0;
margin:0 0 1.5em;
padding:0;
}
td { padding:0.333em;
vertical-align:middle;
}
}"""
with open(args.viz_fname, 'w') as f:
f.write('<html><head><style>')
f.write(css)
f.write('</style></head>')
f.write('<body>')
display_ix = np.random.randint(0, len(ygnd), args.num_instances)
xds, yds = [], []
for ix, (xd, yd) in enumerate(reader):
if ix in display_ix:
dix = ix
iloc = np.round(np.array(xd.shape) / 2)
imgXY = show_png(xd[:, :, iloc[2]]).get_html_embed_code()
imgXZ = show_png(xd[:, iloc[1], :]).get_html_embed_code()
imgYZ = show_png(xd[iloc[0], :, :]).get_html_embed_code()
f.write('<div>')
f.write('<table><tr><td>')
f.write(imgXY)
f.write(imgXZ)
f.write(imgYZ)
f.write('</td>')
f.write('<td>')
f.write('<dl><dt>Instance:</dt><dd>{}</dd>'.format(yd))
f.write('<dt>Predicted label:</dt><dd>{}</dd>'.format(
str(yhat[dix])))
f.write('<dt>True label:</dt><dd>{}</dd></dl>'.format(
str(ygnd[dix])))
f.write('</td></tr></table>')
f.write('</div>')
xds.append(xd)
yds.append(yd)
f.write('</body></html>')
from path import Path
import numpy as np
import voxnet
from voxnet import npytar
from voxnet import isovox
from voxnet.data import shapenet10
parser = argparse.ArgumentParser()
parser.add_argument('out_fname', type=Path)
parser.add_argument('test_fname', type=Path)
parser.add_argument('viz_fname', type=Path)
parser.add_argument('--num-instances', type=int, default=10)
args = parser.parse_args()
reader = npytar.NpyTarReader(args.test_fname)
out = np.load(args.out_fname)
yhat = out['yhat']
ygnd = out['ygnd']
iv = isovox.IsoVox()
css = """
html { margin: 0 }
body {
background:#fff;
color:#000;
font:75%/1.5em Helvetica, "DejaVu Sans", "Liberation sans", "Bitstream Vera Sans", sans-serif;
position:relative;
}
xc[cix] = x.astype(np.float32)
yc.append(int(name.split('.')[0]) - 1)
if ix % num_divide == 0:
xc = jitter_chunk(xc)
xc = 2.0 * xc - 1.0
xc = np.array(xc)
xc = np.reshape(xc, (-1, 32, 32, 32))
data_list.append(xc)
xc.fill(0)
return data_list, yc
if Train:
# Read all dataset
reader = npytar.NpyTarReader('shapenet40_train.tar')
# Set subset for augmentation
xc = np.zeros((num_train_divide, 32, 32, 32))
# After augmentation
# data_list & yc includes whole dataset
data_list, yc = tar_reader(reader, xc, Train)
yc = np.array(yc)
print('Start Saving!')
num_h5 = 9
len_train = int(num_train / num_train_divide) # 36
data_list_divide = int(len_train / num_h5) # 3
label_divide = int(num_train / num_h5) # 9843
for i in range(num_h5):
final_xc = np.stack(data_list[data_list_divide * i:data_list_divide *