def run():
start_num = 2
end_num = 8
dest = 'all_pairs_survey.png'
next_x = 0
for i, n in enumerate(range(start_num, end_num + 1)):
im_dim = [76, 96][n > 5]
next_x += im_dim + 1
result_size = (2 * 96 + 1, next_x - 1)
print(result_size)
result = np.ones((2 * 96 + 1, next_x - 1))
next_y = 0
for target_label in [0, 1]:
next_x = 0
print('target_label', target_label)
for i, n in enumerate(range(start_num, end_num + 1)):
im_dim = [76, 96][n > 5]
dy = [10, 0][n > 5]
spec = SampleSpec(n, n, im_dim=im_dim, min_cell=15, max_cell=18)
vis_input, vis_labels, stats = spec.blocking_generate_with_stats(
200)
sample_index = find_index(vis_labels, target_label)
y = next_y + dy
result[y:y + im_dim,
next_x:next_x + im_dim] = vis_input[sample_index, 0, :, :]
next_x += im_dim + 1
print(stats)
# vis_input = 1.0 - vis_input
next_y += 96 + 1
scipy.misc.imsave(dest, result)
def run():
# get the dataloader we want to work with
data_loader = get_data_loader()
# build main typenet object
typenet = TypeNet(data_loader.img_shp,
data_loader.output_size,
args.num_top_features,
debug=args.debug,
activations=args.activations)
# parallelize across multiple GPU's
if args.ngpu > 1:
typenet = nn.DataParallel(typenet)
print('Devices:', typenet.device_ids)
# push to cuda
if args.cuda:
typenet = typenet.cuda()
# build optimizer
print("training...")
optimizer = optim.Adam(typenet.parameters(), lr=args.lr)
# build the logger object
logger = None
if args.filelog != '':
short_name = get_short_name()
logger = FileLogger(
{
'train': ['epoch', 'loss', 'acc'],
'test': ['epoch', 'loss', 'acc'],
}, [
__file__,
'allpairs/grid_generator.py',
'allpairs/symbol_drawing.py',
],
path=args.filelog,
file_prefix=short_name,
args=args)
logger.set_info('note', args.note)
logger.set_info('uuid', logger.uuid)
logger.set_info('args', str(args))
# main training loop
best_acc = 0.0
for epoch in range(1, args.epochs + 1):
train(epoch, typenet, optimizer, data_loader, logger)
acc = test(epoch, typenet, data_loader, logger)
best_acc = max(best_acc, acc)
print('best acc: {}'.format(best_acc))
# close all the generators, needed to deal with multi-process generators
SampleSpec.close_generators()
def run():
f = open(args.csv, 'w')
spec = SampleSpec(args.num_pairs,
args.num_classes,
im_dim=args.pixels,
min_cell=15,
max_cell=18)
for i in range(args.num):
progress(i, args.num)
images, labels, stats = spec.blocking_generate_with_stats(1)
outpath = os.path.join(args.dest, '{}.png'.format(str(i).zfill(6)))
scipy.misc.imsave(outpath, images[0, 0, :, :])
f.write(str(labels[0]))
f.write('\n')
def set_sample_spec(num_pairs, num_classes, reset_every=None, im_dim=76):
global sample_spec
assert sample_spec.generators is None, 'attempting to redefine spec after it has been used'
sample_spec = SampleSpec(num_pairs=num_pairs,
num_classes=num_classes,
im_dim=im_dim,
min_cell=15,
max_cell=18,
reset_every=reset_every)
def test_images_same(self):
correct = ((
('e11b8020b9f8fabe', '18a3ae95b53eb2a3'),
('b1cbdc248365556e', '3a45220b17d7bf49'),
('7fbfff719fe71393', 'b436c33047a91227'),
('07d1403088b6c624', '6b0a099525d4b709'),
('deb6823a3285233a', 'c6c1a74630ba13ca'),
('e6433a7d3bfb7c5e', 'e7884540c7d4364b'),
('90228d0abfa1c73b', '7ce1c8786cc40475'),
('c4be817429e61d25', '6dcd05d7168be05e'),
('be5a2d302cae9f18', '89659712a3f2c7c7'),
('0a42707c9f1c3c32', '35b3d810a33533c8'),
), (
('d1a12878c42725a5', '1759c3b44205f0fb'),
('041d9ff827beba3e', 'bf9aa5fbcf176355'),
('0162327e25617782', '8ee558aac58bb032'),
('de3571b7bd364931', 'f439cfcbd075b6bf'),
('98ffaea06e268eda', 'ba5bad1a9b1676eb'),
('3c042e60e1ca15b6', '14d7b3a418f5510a'),
('367ca722d496bf33', '052a7ce6679df288'),
('f7321cb15555ce55', '231060e55e4e987d'),
('361d0c54a66fe7e2', '2dc6219f8ee96f1d'),
('f6e04ac9005b6d58', '64ec9871f8d58228'),
))
np.random.seed(1234)
im_dim = 96
num = 10
for which, n in enumerate([4, 8]):
spec = SampleSpec(n, n, im_dim=im_dim, min_cell=15, max_cell=18)
vis_input, vis_labels, stats = spec.blocking_generate_with_stats(
num)
for i in range(num):
im = vis_input[i, 0, :, :]
(h0, h1) = image_hash(im)
h0 = h0[0:16]
h1 = h1[0:16]
# print(h0, h1)
self.assertEqual(h0, correct[which][i][0])
self.assertEqual(h1, correct[which][i][1])
# print()
self.assertTrue(True)
def run():
data = get_data_loader()
typenet = TypeNet(data.img_shp,
data.output_size,
args.num_top_features,
activations=args.activations)
if args.ngpu > 1: # parallelize across multiple GPU's
typenet = nn.DataParallel(typenet)
if args.cuda: # push to cuda
typenet = typenet.cuda()
# main training loop
optimizer = optim.Adam(typenet.parameters(), lr=args.lr)
num_trained_on = 0
while num_trained_on < args.max_train_samples:
train(num_trained_on, typenet, optimizer, data)
num_trained_on += args.batch_size
test(typenet, data)
# close all the generators, needed to deal with multi-process generators
SampleSpec.close_generators()
import torch
from torch.utils.data import Dataset
import numpy as np
from allpairs.grid_generator import SampleSpec
DEFAULT_PAIRS = 4
DEFAULT_CLASSES = 4
sample_spec = SampleSpec(num_pairs=DEFAULT_PAIRS,
num_classes=DEFAULT_CLASSES,
im_dim=76,
min_cell=15,
max_cell=18)
def set_sample_spec(num_pairs, num_classes, reset_every=None, im_dim=76):
global sample_spec
assert sample_spec.generators is None, 'attempting to redefine spec after it has been used'
sample_spec = SampleSpec(num_pairs=num_pairs,
num_classes=num_classes,
im_dim=im_dim,
min_cell=15,
max_cell=18,
reset_every=reset_every)
class ToTensor(object):
"""simple override to add context to ToTensor"""
def __init__(self, numpy_base_type=np.float32):
self.numpy_base_type = numpy_base_type
def __call__(self, index, img, context):