def generate_layers(path, filename, out_path):
for i, layer in enumerate(layers):
print i
sys.stdout.flush()
ts = time.time()
im = np.asarray(Image.open(os.path.join(path, filename)).convert('RGB').resize((256, 256), resample=Image.ANTIALIAS)) \
.transpose(2, 0, 1).reshape((1, 3, 256, 256))[:, :, 16:-16][:, :, :, 16:-16]
im_name = filename[:filename.find('.')]
dirname = os.path.join(out_path,im_name)
if not os.path.exists(dirname):
os.mkdir(dirname)
if layer is not None:
STATS = STAT_LIST[:layer]
else:
STATS = STAT_LIST
stim_name = os.path.join(dirname, im_name+'_layer'+str(i)+'.npy')
r, targets = thing4.main(pf, sf, stim_name,
[(im, STATS)],
use_bounds=True,
data_layer='data',
start_layer='conv1_1',
start_normal=False,
crop=(16, -16, 16, -16),
save_dir = os.path.join(dirname, im_name+'_layer'+str(i)+'_history'),
save_freq=100,
seed=0)
save_stim(stim_name)
print time.time()-ts
sys.stdout.flush()
def generate_layers(path, filename, out_path):
seed = np.random.randint(2**32)
print 'Seed used:', seed
for i, layer in enumerate(layers):
print i
sys.stdout.flush()
ts = time.time()
im = get_image(os.path.join(path, filename))
im_name = filename[:filename.find('.')]
dirname = os.path.join(out_path, im_name)
if not os.path.exists(dirname):
os.mkdir(dirname)
orig = im.squeeze().transpose((1, 2, 0)).astype('uint8')
img = Image.fromarray(orig, 'RGB')
img.save(os.path.join(dirname, im_name + '.o.jpg'))
if layer is not None:
STATS = STAT_LIST[:layer]
else:
STATS = STAT_LIST
stim_name = os.path.join(dirname, im_name + '_layer' + str(i) + '.npy')
r, targets = thing4.main(pf,
sf,
stim_name, [(im, STATS)],
use_bounds=True,
data_layer='data',
start_layer='conv1_1',
start_normal=False,
crop=(16, -16, 16, -16),
save_dir=os.path.join(
dirname,
im_name + '_layer' + str(i) + '_history'),
save_freq=100,
seed=seed)
save_stim(stim_name)
print time.time() - ts
sys.stdout.flush()
('pool4', 100, 'corr_rs')]
pf = 'model.prototxt'
sf = 'model_parameters.caffemodel'
layers = [1, 3, 6, 11, None]
for i, layer in enumerate(layers):
print i
sys.stdout.flush()
ts = time.time()
im = 'campbell256.o.jpg'
im2 = np.asarray(Image.open('textures/texture_jpgs/'+im).convert('RGB').resize((256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape((1, 3, 256, 256))[:, :, 16:-16][:, :, :, 16:-16]
im_name = im[:im.find('.')]+'_p2_layer'+str(i)
dirname = 'textures/generated/'+im_name
if layer is not None:
STATS = STAT_LIST[:layer]
else:
STATS = STAT_LIST
r, targets = thing4.main(pf, sf, os.path.join(dirname, 'genstim.npy'),
[(im2, STATS)],
use_bounds=True,
data_layer='data',
start_layer='conv1_1',
start_normal=False,
crop=(16, -16, 16, -16),
save_dir = os.path.join(dirname, 'history'),
save_freq=100,
seed=0)
print time.time()-ts
sys.stdout.flush()
mean = np.load('/home/ubuntu/new/caffe/audio_batches_mean.npy')
x = cPickle.load(open('/home/ubuntu/new/caffe/cgrams.cpy'))
im = Image.fromarray((np.tile(x['all_cgrams'][:, :, -14],
(3, 1, 1)) * 255).astype(np.uint8).transpose(
(1, 2, 0))).resize((225, 225))
im = np.asarray(im).transpose((2, 0, 1)).reshape((1, 3, 225, 225))
print(im.shape)
r, targets = thing4.main(
pf,
sf,
'/home/ubuntu/test_audio_real2/test_real.npy',
[(
im,
[ #('conv1', 1, 'corr_t'),
#('pool1', 1, 'corr_t'),
#('conv2', 1, 'corr_t'),
#('pool2', 1, 'corr_t'),
#('conv3', 1, 'corr_t'),
#('conv4', 1, 'corr_t'),
#('conv5', 1, 'corr_t'),
('pool5', 1, 'corr_t')
])],
'data',
seed=0,
use_bounds=True,
mean=mean,
save_dir='/home/ubuntu/test_audio_real2/things/',
save_freq=100,
start_normal='zeros')
import numpy as np
import thing4
reload(thing4)
import dldata.stimulus_sets.hvm as hvm
from collections import OrderedDict
pf = '/home/ubuntu/new/caffe/examples/cifar10/caffenet_rand.prototxt'
sf = '/home/ubuntu/new/caffe/examples/cifar10/bvlc_reference_caffenet.caffemodel'
#im0 = 255*imgs[0].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
#im1 = 255*imgs[-1].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
var = np.array([-1] + [0] * 4095).astype(np.float32)
r = thing4.main(pf,
sf,
'/home/ubuntu/test_O3_bounds_s.npy',
targets0=[(None, [('fc6', 1, ('max', var))])],
crop=(14, -15, 14, -15),
use_bounds=True,
seed=1)
import thing4
import dldata.stimulus_sets.hvm as hvm
from collections import OrderedDict
pf = '/home/ubuntu/new/caffe/examples/cifar10/cifar10_quick_train_test_rand.prototxt'
sf = '/home/ubuntu/new/caffe/examples/cifar10/cifar10_quick_iter_5000.caffemodel.h5'
import numpy as np
v = np.load('/home/ubuntu/new/caffe/data_target.npy')
im = v[2]
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_xc.npy', [(im, [('conv1', 200, 'corr'), ('conv2', 200, 'corr'), ('c#onv3', 200, 'corr')])], 'data', mean=np.zeros_like(v[0]))
r, targets = thing4.main(pf,
sf,
'/home/ubuntu/test_yc1.npy',
[(im, [('pool3', 100, 'corr')])],
'data',
mean=np.zeros_like(v[0]))
im = Image.fromarray((np.tile(x['all_cgrams'][:, :, 8],
(3, 1, 1)) * 255).astype(np.uint8).transpose(
(1, 2, 0))).resize((225, 225),
resample=Image.ANTIALIAS)
im = np.asarray(im).transpose((2, 0, 1)).reshape((1, 3, 225, 225))
print(im.shape)
r, targets = thing4.main(
pf,
sf,
'/home/ubuntu/test_audio_stim4a_diag/test_real.npy',
[(
im,
[
('conv1', 1, 'corr_diag'),
('pool1', 1, 'corr_diag'),
('conv2', 1, 'corr_diag'),
('pool2', 1, 'corr_diag'),
('conv3', 1000, 'corr_diag'),
('conv4', 100000, 'corr_diag'),
('conv5', 100000, 'corr_diag'),
('pool5', 100000, 'corr_diag'),
#('fc6', 100000, 'ss')
])],
'data',
seed=2,
use_bounds=True,
#start_normal=False, start_im = im,
mean=mean,
save_dir='/home/ubuntu/test_audio_stim4a_diag/things/',
save_freq=100)
r, targets = thing4.main(
pf,
sf,
os.path.join(dirname, 'genstim.npy'),
[(
im2,
[
('conv1', 100, 'corr'),
('norm1', 100, 'corr'),
('pool1', 100, 'corr'),
(
'conv2',
10000,
'corr',
),
('norm2', 100, 'corr'),
('pool2', 100, 'corr'),
('conv3', 10000, 'corr'),
('conv4', 100000, 'corr'),
('conv5', 1000000, 'corr'),
('pool5', 1000000, 'corr'),
#('fc6', 10000, 'ss'),
#('fc7', 10000, 'ss'),
#('fc8', 10000, 'ss')
])],
use_bounds=True,
data_layer='data',
start_normal=False,
crop=(14, -15, 14, -15),
save_dir=os.path.join(dirname, 'history'),
save_freq=100,
seed=0)
im = np.asarray(
Image.open('/home/ubuntu/imgres.jpg').resize(
(256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape(
(1, 3, 256, 256))[:, :, 14:-15][:, :, :, 14:-15]
im1 = np.asarray(
Image.open('/home/ubuntu/imgres-1.jpg').resize(
(256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape(
(1, 3, 256, 256))[:, :, 14:-15][:, :, :, 14:-15]
var = np.zeros((1000, )).astype(np.float32)
#var = np.zeros((4096,)).astype(np.float32)
var[0] = 1
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_something4.npy', [([im], [('pool1', 100, 'corr')]), (None, [('fc8', 1, ('max', var))])], 'data', use_bounds=True, crop=(14, -15, 14, -15))
r, targets = thing4.main(pf,
sf,
'/home/ubuntu/genstim10/genstim.npy',
[(None, [('data', .01, 'smooth'),
('fc8', (10000, var), 'softmax')]),
([im], [('pool1', 20, 'corr')])],
data_layer='data',
use_bounds=True,
crop=(14, -15, 14, -15),
start_normal=True,
seed=1,
start_im=im,
save_dir='/home/ubuntu/genstim10/things',
save_freq=100)
import thing4
import dldata.stimulus_sets.hvm as hvm
from collections import OrderedDict
preproc = OrderedDict([(u'normalize', False), (u'dtype', u'float32'),
(u'resize_to', [256, 256, 3]), (u'mode', u'RGB'),
(u'crop', None), (u'mask', None)])
dataset = hvm.HvMWithDiscfade()
imgs = dataset.get_images(preproc=preproc)
pf = '/home/ubuntu/new/caffe/examples/cifar10/caffenet_rand.prototxt'
sf = '/home/ubuntu/new/caffe/examples/cifar10/bvlc_reference_caffenet.caffemodel'
#im0 = 255*imgs[0].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
#im1 = 255*imgs[-1].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
meta = dataset.meta
inds = (meta['category'] == 'Cars').nonzero()[0]
ims = [
255 * imgs[i][14:-15][:, 14:-15].transpose((2, 0, 1)).reshape(
(1, 3, 227, 227)) for i in inds[::20][:10]
]
r, targets = thing4.main(pf,
sf,
'/home/ubuntu/test_fc6_cn_cars3.npy',
[(ims[0], [('conv1', .5, 'ss')]),
(ims[1], [('conv5', 10, 'corr')])],
'data',
crop=(14, -15, 14, -15))
im_name = im[:im.find('.')] + '_p2'
dirname = 'textures/generated/' + im_name
r, targets = thing4.main(pf,
sf,
os.path.join(dirname, 'genstim.npy'),
[(im2, [('conv1_1', 1000, 'corr_rs'),
('conv1_2', 100, 'corr_rs'),
('pool1', 100, 'corr_rs'),
('conv2_1', 100, 'corr_rs'),
('conv2_2', 100, 'corr_rs'),
('pool2', 100, 'corr_rs'),
('conv3_1', 100, 'corr_rs'),
('conv3_2', 100, 'corr_rs'),
('conv3_3', 100, 'corr_rs'),
('conv3_4', 100, 'corr_rs'),
('pool3', 10, 'corr_rs'),
('conv4_1', 100, 'corr_rs'),
('conv4_2', 100, 'corr_rs'),
('conv4_3', 100, 'corr_rs'),
('conv4_4', 1000, 'corr_rs'),
('pool4', 100, 'corr_rs')])],
use_bounds=True,
data_layer='data',
start_layer='conv1_1',
start_normal=False,
crop=(16, -16, 16, -16),
save_dir=os.path.join(dirname, 'history'),
save_freq=100,
seed=0)
print time.time() - ts
sys.stdout.flush()
meta = dataset.meta
inds = (meta['category'] == 'Cars').nonzero()[0]
ims = [
255 * imgs[i][14:-15][:, 14:-15].transpose((2, 0, 1)).reshape(
(1, 3, 227, 227)) for i in inds
]
import numpy as np
from PIL import Image
im = np.asarray(
Image.open('/home/ubuntu/imgres.jpg').resize(
(256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape(
(1, 3, 256, 256))[:, :, 14:-15][:, :, :, 14:-15]
var = np.zeros((1000, )).astype(np.float32)
var[55] = -1
r, targets = thing4.main(pf,
sf,
'/home/ubuntu/genstim2.npy',
[(None, [('data', .1, 'smooth'),
('fc8', (100, var), 'max2')])],
'data',
use_bounds=True,
crop=(14, -15, 14, -15),
start_normal=True,
seed=2)
#r, targets = thing4.main(pf, sf, '/home/ubuntu/genstim2.npy', [(None, [('fc8', (100, var), 'max')])], 'data', use_bounds=True, crop=(14, -15, 14, -15), start_normal=True, seed=2)
#r, targets = thing4.main(pf, sf, '/home/ubuntu/genstim2.npy', [(None, [('data', 1, 'smooth')])], 'data', use_bounds=False, crop=(14, -15, 14, -15), start_normal=True, seed=0)
im1 = np.asarray(
Image.open('/home/ubuntu/test.png').convert('RGB').resize(
(256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape(
(1, 3, 256, 256))[:, :, 14:-15][:, :, :, 14:-15]
im2 = np.asarray(
Image.open('/home/ubuntu/imgres-2.png').convert('RGB').resize(
(256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape(
(1, 3, 256, 256))[:, :, 14:-15][:, :, :, 14:-15]
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_y4a.npy', [([ims[i] for i in range(640)], [('fc7', 100, 'ss'), ('fc6', 100, 'ss'), ('fc8', 100, 'ss')])], 'data', crop=(14, -15, 14, -15))
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_rocks_cn_pool.npy', [(im1, [('conv1', 100, 'corr'), ('pool1', 100, 'corr'), ('conv2', 100, 'corr', ), ('pool2', 100, 'corr'), ('conv3', 100, 'corr'), ('conv4', 100, 'corr'), ('conv5', 100, 'corr'), ('pool5', 100, 'corr')])], 'data', crop=(14, -15, 14, -15))
r, targets = thing4.main(pf,
sf,
'/home/ubuntu/test_fruits_cn_all.npy',
[(im2, [('conv1', 1000, 'corr'),
('pool1', 1000, 'corr'),
('conv2', 1000, 'corr'),
('pool2', 1000, 'corr'),
('conv5', 1000, 'corr'),
('pool5', 1000, 'corr'), ('fc6', 1000, 'ss'),
('fc7', 1000, 'ss')])],
use_bounds=True,
data_layer='data',
crop=(14, -15, 14, -15))
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_y3d.npy', [(ims[0], [('conv1', 100, 'corr'), ('conv2', 100, 'corr', ), ('conv3', 100, 'corr'), ('conv4', 100, 'corr'), ('conv5', 100, 'corr')])], 'data', crop=(14, -15, 14, -15), use_bounds=True, start_normal=False)
inds = (meta['category'] == 'Cars').nonzero()[0]
ims = [255*imgs[i][14:-15][:, 14:-15].transpose((2, 0, 1)).reshape((1, 3, 227, 227)) for i in inds]
import numpy as np
from PIL import Image
im = np.asarray(Image.open('/home/ubuntu/imgres.jpg').resize((256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape((1, 3, 256, 256))[:, :, 14:-15][:, :, :, 14:-15]
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_y4a.npy', [([ims[i] for i in range(640)], [('fc7', 100, 'ss'), ('fc6', 100, 'ss'), ('fc8', 100, 'ss')])], 'data', crop=(14, -15, 14, -15))
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_y3b.npy', [(ims[-1], [('conv1', 100, 'corr'), ('conv2', 100, 'corr', ), ('conv5', 100, 'corr'), ('pool5', 100, 'ss')])], 'data', crop=(14, -15, 14, -15))
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_X.npy', [([ims[i] for i in range(1)], [('conv5', 100, 'corr'), ('pool5', 100, 'ss'), ('fc7', 100, 'ss'), ('fc6', 100, 'ss'), ('fc8', 100, 'ss')])], 'data', crop=(14, -15, 14, -15))
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_X.npy', [([ims[i] for i in range(640)], [('fc8', 10000, 'ss')])], 'data', crop=(14, -15, 14, -15), use_bounds=True)
#inds = (meta['category'] == 'Faces').nonzero()[0]
#ims = [255*imgs[i][14:-15][:, 14:-15].transpose((2, 0, 1)).reshape((1, 3, 227, 227)) for i in inds]
im = np.asarray(Image.open('/home/ubuntu/imgres.jpg').resize((256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape((1, 3, 256, 256))[:, :, 14:-15][:, :, :, 14:-15]
im1 = np.asarray(Image.open('/home/ubuntu/imgres-1.jpg').resize((256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape((1, 3, 256, 256))[:, :, 14:-15][:, :,:, 14:-15]
var = np.zeros((1000,)).astype(np.float32)
var[23] = -1
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_something4.npy', [([im], [('pool1', 100, 'corr')]), (None, [('fc8', 1, ('max', var))])], 'data', use_bounds=True, crop=(14, -15, 14, -15))
r, targets = thing4.main(pf, sf, '/home/ubuntu/genstim7/genstim.npy',
[(None, [('data', .001, 'smooth')]),
([im], [('fc8', 1, 'ss')])],
data_layer = 'data', use_bounds=True, crop=(14, -15, 14, -15), start_normal=True, seed=0,
save_dir='/home/ubuntu/genstim7/things', save_freq=100)
var = .001 * np.ones((1000, )).astype(np.float32)
#var = np.zeros((4096,)).astype(np.float32)
var[801] = -1
var2 = np.zeros((1000, )).astype(np.float32)
#var = np.zeros((4096,)).astype(np.float32)
var2[2] = 1
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_something4.npy', [([im], [('pool1', 100, 'corr')]), (None, [('fc8', 1, ('max', var))])], 'data', use_bounds=True, crop=(14, -15, 14, -15))
dirname = '/home/ubuntu/genstim11d'
r, targets = thing4.main(
pf,
sf,
os.path.join(dirname, 'genstim.npy'),
[(
None,
[('data', 0.5, 'smoothsep2'), ('data', .01, 'smoothsep'),
('fc8', (25000, var), 'max')
#('fc8', (25000, var2), 'logsoftmax')
])],
data_layer='data',
use_bounds=True,
crop=(14, -15, 14, -15),
start_normal=False,
seed=2,
#start_im = im,
save_dir=os.path.join(dirname, 'things'),
save_freq=100)
pf = '/home/ubuntu/new/caffe/examples/cifar10/caffenet_rand.prototxt'
sf = '/home/ubuntu/new/caffe/examples/cifar10/bvlc_reference_caffenet.caffemodel'
#im0 = 255*imgs[0].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
#im1 = 255*imgs[-1].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
meta = dataset.meta
inds = (meta['category'] == 'Cars').nonzero()[0]
ims = [
255 * imgs[i][14:-15][:, 14:-15].transpose((2, 0, 1)).reshape(
(1, 3, 227, 227)) for i in inds
]
import numpy as np
from PIL import Image
im = np.asarray(
Image.open('/home/ubuntu/imgres.jpg').resize(
(256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape(
(1, 3, 256, 256))[:, :, 14:-15][:, :, :, 14:-15]
var = np.zeros((1000, )).astype(np.float32)
var[23] = -1
r, targets = thing4.main(pf,
sf,
'/home/ubuntu/genstim1.npy',
[(None, [('fc8', (100, var), 'max')])],
'data',
use_bounds=False,
crop=(14, -15, 14, -15),
start_normal=False,
start_im=im)
import thing4
import dldata.stimulus_sets.hvm as hvm
from collections import OrderedDict
preproc = OrderedDict([(u'normalize', False), (u'dtype', u'float32'), (u'resize_to', [256, 256, 3]), (u'mode', u'RGB'), (u'crop', None), (u'mask', None)])
pf = '/home/ubuntu/new/caffe/examples/cifar10/audio2.prototxt'
sf = '/home/ubuntu/new/caffe/examples/cifar10/audio.caffemodel.h5'
import numpy as np
from PIL import Image
im = np.asarray(Image.open('/home/ubuntu/imgres.jpg').resize((256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape((1, 3, 256, 256))
im1 = np.asarray(Image.open('/home/ubuntu/imgres-2.png').convert('RGB').resize((225, 225), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape((1, 3, 225, 225))
mean = np.load('/home/ubuntu/new/caffe/audio_batches_mean.npy')
r, targets = thing4.main(pf, sf, '/home/ubuntu/test_audio_corr/test_bork.npy', [(im1, [('conv1', 1, 'corr_rs'), ('pool1', 1, 'corr_rs'), ('conv2', 1, 'corr_rs'), ('pool2', 1, 'corr_rs'), ('conv3', 1, 'corr_rs'), ('conv4', 1, 'corr_rs'), ('conv5', 1, 'corr_rs'), ('pool5', 1, 'corr_rs')])], 'data', seed=2, use_bounds=True, mean=mean, save_dir = '/home/ubuntu/test_audio_corr/things/', save_freq=100)
import thing4
import dldata.stimulus_sets.hvm as hvm
from collections import OrderedDict
preproc = OrderedDict([(u'normalize', False), (u'dtype', u'float32'),
(u'resize_to', [256, 256, 3]), (u'mode', u'RGB'),
(u'crop', None), (u'mask', None)])
dataset = hvm.HvMWithDiscfade()
imgs = dataset.get_images(preproc=preproc)
pf = '/home/ubuntu/new/caffe/examples/cifar10/roschinet_larger_rand_test.prototxt'
sf = '/home/ubuntu/new/caffe/examples/cifar10/roschinet_larger.caffemodel.h5'
im0 = imgs[0].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
im1 = imgs[-1].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
meta = dataset.meta
inds = meta['category'] == 'Cars'
ims = [
imgs[i].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
for i in inds[::20][:10]
]
r, targets = thing4.main(pf, sf, '/home/ubuntu/test_ig.npy',
[(ims[i], [('conv4', 100, 'ss')]) for i in range(10)],
'data')
(1, 3, 256, 256))[:, :, 14:-15][:, :, :, 14:-15]
var = .001 * np.ones((1000, )).astype(np.float32)
#var = np.zeros((4096,)).astype(np.float32)
var[350] = -1
var2 = np.zeros((1000, )).astype(np.float32)
#var = np.zeros((4096,)).astype(np.float32)
var2[2] = 1
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_something4.npy', [([im], [('pool1', 100, 'corr')]), (None, [('fc8', 1, ('max', var))])], 'data', use_bounds=True, crop=(14, -15, 14, -15))
r, targets = thing4.main(
pf,
sf,
'/home/ubuntu/genstim15/genstim.npy',
[(
None,
[('data', 0.5, 'smoothsep2'), ('data', .001, 'smoothsep'),
('fc8', (25000, var), 'max')
#('fc8', (25000, var2), 'logsoftmax')
])],
data_layer='data',
use_bounds=True,
crop=(14, -15, 14, -15),
start_normal=False,
seed=10,
#start_im = im,
save_dir='/home/ubuntu/genstim15/things',
save_freq=100)
def get_stim(fname, stat_n, stat_n2, seed, out_fname, save_freq, maxfun):
impath = os.path.join(
'/Users/babylab/Desktop/sandbox/textures/texture_jpgs/', fname)
im_arr = np.asarray(
Image.open(impath).convert('RGB').resize(
(224, 224), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape(
(1, 3, 224, 224))
if stat_n is not None:
stat_list = STAT_LIST[:stat_n]
else:
stat_list = STAT_LIST[:]
if stat_n2 is not None:
stat_list += STAT_LIST_2[:stat_n2]
dirname = os.path.join('/Users/babylab/Desktop/sandbox/textures/generated',
out_fname)
final_path = os.path.join(dirname, 'genstim.npy')
if os.path.exists(final_path):
print('%s already finished, exiting' % dirname)
return
else:
hdir = os.path.join(dirname, 'history')
if os.path.isdir(hdir):
hist = filter(lambda x: x.startswith('im_'), os.listdir(hdir))
else:
hist = []
if len(hist) > 0:
histints = [int(x.split('_')[-1].split('.')[0]) for x in hist]
mhist = max(histints)
start_path = os.path.join(hdir, 'im_%d.npy' % mhist)
print('Starting with %s' % start_path)
x0 = np.load(start_path)
r, targets = thing4.main(pf,
sf,
final_path, [(im_arr, stat_list)],
use_bounds=True,
data_layer='data',
start_layer='conv1_1',
start_normal=False,
start_im=x0,
count_start=mhist,
crop=(16, -16, 16, -16),
save_dir=hdir,
save_freq=save_freq,
seed=seed,
opt_kwargs={'maxfun': maxfun})
else:
print('Starting %s from scratch' % dirname)
r, targets = thing4.main(pf,
sf,
final_path, [(im_arr, stat_list)],
use_bounds=True,
data_layer='data',
start_layer='conv1_1',
start_normal=False,
crop=(16, -16, 16, -16),
save_dir=hdir,
save_freq=save_freq,
seed=seed,
opt_kwargs={'maxfun': maxfun})
return r, targets
x = cPickle.load(open('/home/ubuntu/new/caffe/cgrams.cpy'))
im = Image.fromarray((np.tile(x['all_cgrams'][:, :, -14],
(3, 1, 1)) * 255).astype(np.uint8).transpose(
(1, 2, 0))).resize((225, 225),
resample=Image.ANTIALIAS)
im = np.asarray(im).transpose((2, 0, 1)).reshape((1, 3, 225, 225))
print(im.shape)
r, targets = thing4.main(
pf,
sf,
'/home/ubuntu/test_audio_real_stim1a_max/test_real.npy',
[(
im,
[ #('conv1', .1, 'corr_rs'),
#('pool1', .1, 'corr_rs'),
#('conv2', 1, 'corr_rs'),
('pool2', 1, 'ss'),
('conv3', 1, 'ss'),
('conv4', 1, 'ss'),
('conv5', 1, 'ss'),
('pool5', 1, 'ss'),
('fc6', 1, 'ss')
])],
'data',
seed=0,
use_bounds=True,
mean=mean,
save_dir='/home/ubuntu/test_audio_real_stim1a_max/things/',
save_freq=100)
from collections import OrderedDict
preproc = OrderedDict([(u'normalize', False), (u'dtype', u'float32'),
(u'resize_to', [256, 256, 3]), (u'mode', u'RGB'),
(u'crop', None), (u'mask', None)])
pf = '/home/ubuntu/new/caffe/examples/cifar10/audio2.prototxt'
sf = '/home/ubuntu/new/caffe/examples/cifar10/audio.caffemodel.h5'
import numpy as np
from PIL import Image
im = np.asarray(
Image.open('/home/ubuntu/imgres.jpg').resize(
(256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape(
(1, 3, 256, 256))
im1 = np.asarray(
Image.open('/home/ubuntu/imgres-2.png').convert('RGB').resize(
(225, 225), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape(
(1, 3, 225, 225))
mean = np.load('/home/ubuntu/new/caffe/audio_batches_mean.npy')
r, targets = thing4.main(pf,
sf,
'/home/ubuntu/test_audio2.npy',
[(im1, [('pool1', 1, 'ss')])],
'data',
seed=2,
use_bounds=True,
mean=mean)
import thing4
import dldata.stimulus_sets.hvm as hvm
from collections import OrderedDict
preproc = OrderedDict([(u'normalize', False), (u'dtype', u'float32'), (u'resize_to', [256, 256, 3]), (u'mode', u'RGB'), (u'crop', None), (u'mask', None)])
dataset = hvm.HvMWithDiscfade()
imgs = dataset.get_images(preproc=preproc)
pf = '/home/ubuntu/new/caffe/examples/cifar10/roschinet_larger_rand_test.prototxt'
sf = '/home/ubuntu/new/caffe/examples/cifar10/roschinet_larger.caffemodel.h5'
im0 = 255*imgs[0].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
im1 = 255*imgs[-1].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
ims = [255*imgs[i].transpose((2, 0, 1)).reshape((1, 3, 256, 256)) for i in range(10)]
r, targets = thing4.main(pf, sf, '/home/ubuntu/test_x2_1.npy', [(ims[i], [('conv3', 10, 'ss')]) for i in range(10)] + \
[(ims[i], [('pool3', 10, 'ss')]) for i in range(10)], 'data')
import thing4
import dldata.stimulus_sets.hvm as hvm
from collections import OrderedDict
preproc = OrderedDict([(u'normalize', False), (u'dtype', u'float32'),
(u'resize_to', [256, 256, 3]), (u'mode', u'RGB'),
(u'crop', None), (u'mask', None)])
dataset = hvm.HvMWithDiscfade()
imgs = dataset.get_images(preproc=preproc)
pf = '/home/ubuntu/new/caffe/examples/cifar10/roschinet_larger_rand_test.prototxt'
sf = '/home/ubuntu/new/caffe/examples/cifar10/roschinet_larger.caffemodel.h5'
im0 = imgs[0].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
im1 = imgs[-1].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
meta = dataset.meta
inds = (meta['category'] == 'Cars').nonzero()[0]
ims = [
imgs[i].transpose((2, 0, 1)).reshape((1, 3, 256, 256))
for i in inds[::20][:10]
]
r, targets = thing4.main(pf, sf, '/home/ubuntu/test_ih1.npy',
[([ims[i] for i in range(10)], [('fc6', 100, 'ss')])],
'data')
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_something4.npy', [([im], [('pool1', 100, 'corr')]), (None, [('fc8', 1, ('max', var))])], 'data', use_bounds=True, crop=(14, -15, 14, -15))
dirname = '/home/ubuntu/genstim14a'
r, targets = thing4.main(
pf,
sf,
os.path.join(dirname, 'genstim.npy'),
#[(None, [('data', .5, 'smoothsep2'),
# ('data', .005, 'smoothsep')])] + \
[([im1], [
('fc8', 10000000, 'ss'),
('fc7', 100000, 'ss'),
('fc6', 1000000, 'ss'),
('pool5', 1000, 'corr_diag'),
('conv5', 100, 'corr_diag'),
('conv4', 100, 'corr_diag'),
('conv3', 100, 'corr_diag'),
('pool2', 100, 'corr_diag'),
('pool1', 100, 'corr_diag'),
('conv2', 100, 'corr_diag'),
('conv1', 100, 'corr_diag'),
]) for i in range(1)],
data_layer='data',
use_bounds=True,
crop=(14, -15, 14, -15),
start_normal=False,
seed=1,
#start_im = im,
save_dir=os.path.join(dirname, 'things'),
save_freq=100)
ims = [
255 * imgs[i][14:-15][:, 14:-15].transpose((2, 0, 1)).reshape(
(1, 3, 227, 227)) for i in inds
]
import numpy as np
from PIL import Image
im = np.asarray(
Image.open('/home/ubuntu/imgres.jpg').resize(
(256, 256), resample=Image.ANTIALIAS)).transpose(2, 0, 1).reshape(
(1, 3, 256, 256))[:, :, 14:-15][:, :, :, 14:-15]
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_y4a.npy', [([ims[i] for i in range(640)], [('fc7', 100, 'ss'), ('fc6', 100, 'ss'), ('fc8', 100, 'ss')])], 'data', crop=(14, -15, 14, -15))
#r, targets = thing4.main(pf, sf, '/home/ubuntu/test_y3c.npy', [(im, [('conv1', 100, 'corr'), ('conv2', 100, 'corr', ), ('conv3', 100, 'corr'), ('conv4', 100, 'corr'), ('conv5', 100, 'corr')])], 'data', crop=(14, -15, 14, -15), use_bounds=True start_normal=False)
r, targets = thing4.main(pf,
sf,
'/home/ubuntu/test_y3d.npy',
[(ims[0], [('conv1', 100, 'corr'),
(
'conv2',
100,
'corr',
), ('conv3', 100, 'corr'),
('conv4', 100, 'corr'),
('conv5', 100, 'corr')])],
'data',
crop=(14, -15, 14, -15),
use_bounds=True,
start_normal=False)
